ZTWHHH commited on
Commit
e0e47d0
·
verified ·
1 Parent(s): 68de8b5

Add files using upload-large-folder tool

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. .gitattributes +1 -0
  2. evalkit_cambrian/lib/python3.10/site-packages/transformers/models/chinese_clip/convert_chinese_clip_original_pytorch_to_hf.py +134 -0
  3. evalkit_cambrian/lib/python3.10/site-packages/transformers/models/chinese_clip/feature_extraction_chinese_clip.py +33 -0
  4. evalkit_cambrian/lib/python3.10/site-packages/transformers/models/efficientformer/__init__.py +109 -0
  5. evalkit_cambrian/lib/python3.10/site-packages/transformers/models/efficientformer/__pycache__/__init__.cpython-310.pyc +0 -0
  6. evalkit_cambrian/lib/python3.10/site-packages/transformers/models/efficientformer/__pycache__/convert_efficientformer_original_pytorch_checkpoint_to_pytorch.cpython-310.pyc +0 -0
  7. evalkit_cambrian/lib/python3.10/site-packages/transformers/models/efficientformer/__pycache__/image_processing_efficientformer.cpython-310.pyc +0 -0
  8. evalkit_cambrian/lib/python3.10/site-packages/transformers/models/efficientformer/__pycache__/modeling_tf_efficientformer.cpython-310.pyc +0 -0
  9. evalkit_cambrian/lib/python3.10/site-packages/transformers/models/univnet/__init__.py +65 -0
  10. evalkit_cambrian/lib/python3.10/site-packages/transformers/models/univnet/__pycache__/feature_extraction_univnet.cpython-310.pyc +0 -0
  11. evalkit_cambrian/lib/python3.10/site-packages/transformers/models/univnet/convert_univnet.py +162 -0
  12. evalkit_cambrian/lib/python3.10/site-packages/transformers/models/univnet/feature_extraction_univnet.py +456 -0
  13. evalkit_cambrian/lib/python3.10/site-packages/transformers/models/univnet/modeling_univnet.py +636 -0
  14. evalkit_cambrian/lib/python3.10/site-packages/transformers/models/yoso/__pycache__/__init__.cpython-310.pyc +0 -0
  15. evalkit_cambrian/lib/python3.10/site-packages/transformers/models/yoso/__pycache__/configuration_yoso.cpython-310.pyc +0 -0
  16. evalkit_cambrian/lib/python3.10/site-packages/transformers/models/yoso/__pycache__/modeling_yoso.cpython-310.pyc +0 -0
  17. evalkit_cambrian/lib/python3.10/site-packages/transformers/models/yoso/convert_yoso_pytorch_to_pytorch.py +108 -0
  18. evalkit_cambrian/lib/python3.10/site-packages/transformers/models/yoso/modeling_yoso.py +1304 -0
  19. infer_4_47_1/lib/python3.10/site-packages/fontTools/designspaceLib/__init__.py +0 -0
  20. infer_4_47_1/lib/python3.10/site-packages/fontTools/feaLib/__main__.py +78 -0
  21. infer_4_47_1/lib/python3.10/site-packages/fontTools/feaLib/__pycache__/__init__.cpython-310.pyc +0 -0
  22. infer_4_47_1/lib/python3.10/site-packages/fontTools/feaLib/__pycache__/__main__.cpython-310.pyc +0 -0
  23. infer_4_47_1/lib/python3.10/site-packages/fontTools/feaLib/__pycache__/ast.cpython-310.pyc +0 -0
  24. infer_4_47_1/lib/python3.10/site-packages/fontTools/feaLib/__pycache__/error.cpython-310.pyc +0 -0
  25. infer_4_47_1/lib/python3.10/site-packages/fontTools/feaLib/__pycache__/lexer.cpython-310.pyc +0 -0
  26. infer_4_47_1/lib/python3.10/site-packages/fontTools/feaLib/__pycache__/location.cpython-310.pyc +0 -0
  27. infer_4_47_1/lib/python3.10/site-packages/fontTools/feaLib/__pycache__/lookupDebugInfo.cpython-310.pyc +0 -0
  28. infer_4_47_1/lib/python3.10/site-packages/fontTools/feaLib/__pycache__/parser.cpython-310.pyc +0 -0
  29. infer_4_47_1/lib/python3.10/site-packages/fontTools/feaLib/__pycache__/variableScalar.cpython-310.pyc +0 -0
  30. infer_4_47_1/lib/python3.10/site-packages/fontTools/feaLib/error.py +22 -0
  31. infer_4_47_1/lib/python3.10/site-packages/fontTools/feaLib/lookupDebugInfo.py +12 -0
  32. infer_4_47_1/lib/python3.10/site-packages/fontTools/feaLib/parser.py +0 -0
  33. infer_4_47_1/lib/python3.10/site-packages/fontTools/feaLib/variableScalar.py +113 -0
  34. infer_4_47_1/lib/python3.10/site-packages/fontTools/merge/__init__.py +248 -0
  35. infer_4_47_1/lib/python3.10/site-packages/fontTools/merge/__main__.py +6 -0
  36. infer_4_47_1/lib/python3.10/site-packages/fontTools/merge/__pycache__/__init__.cpython-310.pyc +0 -0
  37. infer_4_47_1/lib/python3.10/site-packages/fontTools/merge/__pycache__/__main__.cpython-310.pyc +0 -0
  38. infer_4_47_1/lib/python3.10/site-packages/fontTools/merge/__pycache__/base.cpython-310.pyc +0 -0
  39. infer_4_47_1/lib/python3.10/site-packages/fontTools/merge/__pycache__/cmap.cpython-310.pyc +0 -0
  40. infer_4_47_1/lib/python3.10/site-packages/fontTools/merge/__pycache__/layout.cpython-310.pyc +0 -0
  41. infer_4_47_1/lib/python3.10/site-packages/fontTools/merge/__pycache__/options.cpython-310.pyc +0 -0
  42. infer_4_47_1/lib/python3.10/site-packages/fontTools/merge/__pycache__/tables.cpython-310.pyc +0 -0
  43. infer_4_47_1/lib/python3.10/site-packages/fontTools/merge/__pycache__/unicode.cpython-310.pyc +0 -0
  44. infer_4_47_1/lib/python3.10/site-packages/fontTools/merge/__pycache__/util.cpython-310.pyc +0 -0
  45. infer_4_47_1/lib/python3.10/site-packages/fontTools/merge/base.py +81 -0
  46. infer_4_47_1/lib/python3.10/site-packages/fontTools/merge/cmap.py +141 -0
  47. infer_4_47_1/lib/python3.10/site-packages/fontTools/merge/layout.py +526 -0
  48. infer_4_47_1/lib/python3.10/site-packages/fontTools/merge/options.py +85 -0
  49. infer_4_47_1/lib/python3.10/site-packages/fontTools/merge/tables.py +341 -0
  50. infer_4_47_1/lib/python3.10/site-packages/fontTools/merge/unicode.py +78 -0
.gitattributes CHANGED
@@ -1576,3 +1576,4 @@ evalkit_tf446/lib/python3.10/site-packages/nvidia/cufft/lib/libcufft.so.10 filte
1576
  evalkit_tf449/lib/python3.10/site-packages/nvidia/cublas/lib/libcublasLt.so.12 filter=lfs diff=lfs merge=lfs -text
1577
  infer_4_47_1/lib/python3.10/site-packages/gradio/_frontend_code/lite/dist/assets/gradio_client-1.5.3-py3-none-any.whl filter=lfs diff=lfs merge=lfs -text
1578
  evalkit_cambrian/lib/python3.10/site-packages/transformers/models/perceiver/__pycache__/modeling_perceiver.cpython-310.pyc filter=lfs diff=lfs merge=lfs -text
 
 
1576
  evalkit_tf449/lib/python3.10/site-packages/nvidia/cublas/lib/libcublasLt.so.12 filter=lfs diff=lfs merge=lfs -text
1577
  infer_4_47_1/lib/python3.10/site-packages/gradio/_frontend_code/lite/dist/assets/gradio_client-1.5.3-py3-none-any.whl filter=lfs diff=lfs merge=lfs -text
1578
  evalkit_cambrian/lib/python3.10/site-packages/transformers/models/perceiver/__pycache__/modeling_perceiver.cpython-310.pyc filter=lfs diff=lfs merge=lfs -text
1579
+ infer_4_47_1/lib/python3.10/site-packages/fontTools/pens/momentsPen.cpython-310-x86_64-linux-gnu.so filter=lfs diff=lfs merge=lfs -text
evalkit_cambrian/lib/python3.10/site-packages/transformers/models/chinese_clip/convert_chinese_clip_original_pytorch_to_hf.py ADDED
@@ -0,0 +1,134 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright 2022 The OFA-Sys Team Authors and The HuggingFace Team. All rights reserved.
3
+ #
4
+ # Licensed under the Apache License, Version 2.0 (the "License");
5
+ # you may not use this file except in compliance with the License.
6
+ # You may obtain a copy of the License at
7
+ #
8
+ # http://www.apache.org/licenses/LICENSE-2.0
9
+ #
10
+ # Unless required by applicable law or agreed to in writing, software
11
+ # distributed under the License is distributed on an "AS IS" BASIS,
12
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
+ # See the License for the specific language governing permissions and
14
+ # limitations under the License.
15
+
16
+ import argparse
17
+
18
+ import torch
19
+
20
+ from transformers import ChineseCLIPConfig, ChineseCLIPModel
21
+
22
+
23
+ def copy_attn_layer(hf_attn_layer, pt_weights, prefix):
24
+ q_proj, k_proj, v_proj = pt_weights[f"{prefix}.in_proj_weight"].chunk(3, dim=0)
25
+ q_proj_bias, k_proj_bias, v_proj_bias = pt_weights[f"{prefix}.in_proj_bias"].chunk(3, dim=0)
26
+
27
+ out_proj_weights = pt_weights[f"{prefix}.out_proj.weight"]
28
+ out_proj_bias = pt_weights[f"{prefix}.out_proj.bias"]
29
+
30
+ hf_attn_layer.q_proj.weight.data = q_proj
31
+ hf_attn_layer.q_proj.bias.data = q_proj_bias
32
+
33
+ hf_attn_layer.k_proj.weight.data = k_proj
34
+ hf_attn_layer.k_proj.bias.data = k_proj_bias
35
+
36
+ hf_attn_layer.v_proj.weight.data = v_proj
37
+ hf_attn_layer.v_proj.bias.data = v_proj_bias
38
+
39
+ hf_attn_layer.out_proj.weight.data = out_proj_weights
40
+ hf_attn_layer.out_proj.bias.data = out_proj_bias
41
+
42
+
43
+ def copy_mlp(hf_mlp, pt_weights, prefix):
44
+ copy_linear(hf_mlp.fc1, pt_weights, f"{prefix}.c_fc")
45
+ copy_linear(hf_mlp.fc2, pt_weights, f"{prefix}.c_proj")
46
+
47
+
48
+ def copy_linear(hf_linear, pt_weights, prefix):
49
+ hf_linear.weight.data = pt_weights[f"{prefix}.weight"].data
50
+ hf_linear.bias.data = pt_weights[f"{prefix}.bias"].data
51
+
52
+
53
+ def copy_layer(hf_layer, pt_weights, prefix):
54
+ # copy layer norms
55
+ copy_linear(hf_layer.layer_norm1, pt_weights, f"{prefix}.ln_1")
56
+ copy_linear(hf_layer.layer_norm2, pt_weights, f"{prefix}.ln_2")
57
+
58
+ # copy MLP
59
+ copy_mlp(hf_layer.mlp, pt_weights, f"{prefix}.mlp")
60
+
61
+ # copy attn
62
+ copy_attn_layer(hf_layer.self_attn, pt_weights, f"{prefix}.attn")
63
+
64
+
65
+ def copy_layers(hf_layers, pt_weights, prefix):
66
+ for layer_id, hf_layer in enumerate(hf_layers):
67
+ copy_layer(hf_layer, pt_weights, f"{prefix}.{layer_id}")
68
+
69
+
70
+ def copy_text_model_and_projection(hf_model, pt_weights):
71
+ # copy projection
72
+ hf_model.text_projection.weight.data = pt_weights["text_projection"].data.T
73
+
74
+ # copy text encoder
75
+ for name, param in hf_model.text_model.named_parameters():
76
+ param.data = pt_weights[f"bert.{name}"].data
77
+
78
+
79
+ def copy_vision_model_and_projection(hf_model, pt_weights):
80
+ # copy projection
81
+ hf_model.visual_projection.weight.data = pt_weights["visual.proj"].data.T
82
+
83
+ # copy layer norms
84
+ copy_linear(hf_model.vision_model.pre_layrnorm, pt_weights, "visual.ln_pre")
85
+ copy_linear(hf_model.vision_model.post_layernorm, pt_weights, "visual.ln_post")
86
+
87
+ # copy embeddings
88
+ hf_model.vision_model.embeddings.patch_embedding.weight.data = pt_weights["visual.conv1.weight"].data
89
+ hf_model.vision_model.embeddings.class_embedding.data = pt_weights["visual.class_embedding"].data
90
+ hf_model.vision_model.embeddings.position_embedding.weight.data = pt_weights["visual.positional_embedding"].data
91
+
92
+ # copy encoder
93
+ copy_layers(hf_model.vision_model.encoder.layers, pt_weights, "visual.transformer.resblocks")
94
+
95
+
96
+ @torch.no_grad()
97
+ def convert_chinese_clip_checkpoint(checkpoint_path, pytorch_dump_folder_path, config_path=None):
98
+ """
99
+ Copy/paste/tweak model's weights to transformers design.
100
+ """
101
+
102
+ assert config_path is not None, "Please specify the ChineseCLIP model config of the corresponding model size."
103
+ config = ChineseCLIPConfig.from_pretrained(config_path)
104
+
105
+ hf_model = ChineseCLIPModel(config).eval()
106
+
107
+ pt_weights = torch.load(checkpoint_path, map_location="cpu")["state_dict"]
108
+ pt_weights = {(name[7:] if name.startswith("module.") else name): value for name, value in pt_weights.items()}
109
+
110
+ copy_text_model_and_projection(hf_model, pt_weights)
111
+ copy_vision_model_and_projection(hf_model, pt_weights)
112
+ hf_model.logit_scale.data = pt_weights["logit_scale"].data
113
+
114
+ hf_model.save_pretrained(pytorch_dump_folder_path)
115
+
116
+
117
+ if __name__ == "__main__":
118
+ parser = argparse.ArgumentParser()
119
+ parser.add_argument(
120
+ "--pytorch_dump_folder_path",
121
+ default=None,
122
+ type=str,
123
+ help="Path to the output folder storing converted hf PyTorch model.",
124
+ )
125
+ parser.add_argument(
126
+ "--checkpoint_path", default=None, type=str, help="Path to original github format ChineseCLIP checkpoint."
127
+ )
128
+ parser.add_argument(
129
+ "--config_path", default=None, required=True, type=str, help="Path to hf config.json of model to convert."
130
+ )
131
+ args = parser.parse_args()
132
+
133
+ convert_chinese_clip_checkpoint(args.checkpoint_path, args.pytorch_dump_folder_path, args.config_path)
134
+ print("The conversion is finished!")
evalkit_cambrian/lib/python3.10/site-packages/transformers/models/chinese_clip/feature_extraction_chinese_clip.py ADDED
@@ -0,0 +1,33 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright 2021 The OFA-Sys Team Authors and The HuggingFace Team. All rights reserved.
3
+ #
4
+ # Licensed under the Apache License, Version 2.0 (the "License");
5
+ # you may not use this file except in compliance with the License.
6
+ # You may obtain a copy of the License at
7
+ #
8
+ # http://www.apache.org/licenses/LICENSE-2.0
9
+ #
10
+ # Unless required by applicable law or agreed to in writing, software
11
+ # distributed under the License is distributed on an "AS IS" BASIS,
12
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
+ # See the License for the specific language governing permissions and
14
+ # limitations under the License.
15
+ """Feature extractor class for Chinese-CLIP."""
16
+
17
+ import warnings
18
+
19
+ from ...utils import logging
20
+ from .image_processing_chinese_clip import ChineseCLIPImageProcessor
21
+
22
+
23
+ logger = logging.get_logger(__name__)
24
+
25
+
26
+ class ChineseCLIPFeatureExtractor(ChineseCLIPImageProcessor):
27
+ def __init__(self, *args, **kwargs) -> None:
28
+ warnings.warn(
29
+ "The class ChineseCLIPFeatureExtractor is deprecated and will be removed in version 5 of Transformers."
30
+ " Please use ChineseCLIPImageProcessor instead.",
31
+ FutureWarning,
32
+ )
33
+ super().__init__(*args, **kwargs)
evalkit_cambrian/lib/python3.10/site-packages/transformers/models/efficientformer/__init__.py ADDED
@@ -0,0 +1,109 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2022 The HuggingFace Team. All rights reserved.
2
+ #
3
+ # Licensed under the Apache License, Version 2.0 (the "License");
4
+ # you may not use this file except in compliance with the License.
5
+ # You may obtain a copy of the License at
6
+ #
7
+ # http://www.apache.org/licenses/LICENSE-2.0
8
+ #
9
+ # Unless required by applicable law or agreed to in writing, software
10
+ # distributed under the License is distributed on an "AS IS" BASIS,
11
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12
+ # See the License for the specific language governing permissions and
13
+ # limitations under the License.
14
+ from typing import TYPE_CHECKING
15
+
16
+ from ...utils import (
17
+ OptionalDependencyNotAvailable,
18
+ _LazyModule,
19
+ is_tf_available,
20
+ is_torch_available,
21
+ is_vision_available,
22
+ )
23
+
24
+
25
+ _import_structure = {
26
+ "configuration_efficientformer": [
27
+ "EFFICIENTFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP",
28
+ "EfficientFormerConfig",
29
+ ]
30
+ }
31
+
32
+ try:
33
+ if not is_vision_available():
34
+ raise OptionalDependencyNotAvailable()
35
+ except OptionalDependencyNotAvailable:
36
+ pass
37
+ else:
38
+ _import_structure["image_processing_efficientformer"] = ["EfficientFormerImageProcessor"]
39
+
40
+ try:
41
+ if not is_torch_available():
42
+ raise OptionalDependencyNotAvailable()
43
+ except OptionalDependencyNotAvailable:
44
+ pass
45
+ else:
46
+ _import_structure["modeling_efficientformer"] = [
47
+ "EFFICIENTFORMER_PRETRAINED_MODEL_ARCHIVE_LIST",
48
+ "EfficientFormerForImageClassification",
49
+ "EfficientFormerForImageClassificationWithTeacher",
50
+ "EfficientFormerModel",
51
+ "EfficientFormerPreTrainedModel",
52
+ ]
53
+
54
+ try:
55
+ if not is_tf_available():
56
+ raise OptionalDependencyNotAvailable()
57
+ except OptionalDependencyNotAvailable:
58
+ pass
59
+ else:
60
+ _import_structure["modeling_tf_efficientformer"] = [
61
+ "TF_EFFICIENTFORMER_PRETRAINED_MODEL_ARCHIVE_LIST",
62
+ "TFEfficientFormerForImageClassification",
63
+ "TFEfficientFormerForImageClassificationWithTeacher",
64
+ "TFEfficientFormerModel",
65
+ "TFEfficientFormerPreTrainedModel",
66
+ ]
67
+
68
+ if TYPE_CHECKING:
69
+ from .configuration_efficientformer import EFFICIENTFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP, EfficientFormerConfig
70
+
71
+ try:
72
+ if not is_vision_available():
73
+ raise OptionalDependencyNotAvailable()
74
+ except OptionalDependencyNotAvailable:
75
+ pass
76
+ else:
77
+ from .image_processing_efficientformer import EfficientFormerImageProcessor
78
+
79
+ try:
80
+ if not is_torch_available():
81
+ raise OptionalDependencyNotAvailable()
82
+ except OptionalDependencyNotAvailable:
83
+ pass
84
+ else:
85
+ from .modeling_efficientformer import (
86
+ EFFICIENTFORMER_PRETRAINED_MODEL_ARCHIVE_LIST,
87
+ EfficientFormerForImageClassification,
88
+ EfficientFormerForImageClassificationWithTeacher,
89
+ EfficientFormerModel,
90
+ EfficientFormerPreTrainedModel,
91
+ )
92
+ try:
93
+ if not is_tf_available():
94
+ raise OptionalDependencyNotAvailable()
95
+ except OptionalDependencyNotAvailable:
96
+ pass
97
+ else:
98
+ from .modeling_tf_efficientformer import (
99
+ TF_EFFICIENTFORMER_PRETRAINED_MODEL_ARCHIVE_LIST,
100
+ TFEfficientFormerForImageClassification,
101
+ TFEfficientFormerForImageClassificationWithTeacher,
102
+ TFEfficientFormerModel,
103
+ TFEfficientFormerPreTrainedModel,
104
+ )
105
+
106
+ else:
107
+ import sys
108
+
109
+ sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)
evalkit_cambrian/lib/python3.10/site-packages/transformers/models/efficientformer/__pycache__/__init__.cpython-310.pyc ADDED
Binary file (1.73 kB). View file
 
evalkit_cambrian/lib/python3.10/site-packages/transformers/models/efficientformer/__pycache__/convert_efficientformer_original_pytorch_checkpoint_to_pytorch.cpython-310.pyc ADDED
Binary file (6.14 kB). View file
 
evalkit_cambrian/lib/python3.10/site-packages/transformers/models/efficientformer/__pycache__/image_processing_efficientformer.cpython-310.pyc ADDED
Binary file (12.8 kB). View file
 
evalkit_cambrian/lib/python3.10/site-packages/transformers/models/efficientformer/__pycache__/modeling_tf_efficientformer.cpython-310.pyc ADDED
Binary file (37.2 kB). View file
 
evalkit_cambrian/lib/python3.10/site-packages/transformers/models/univnet/__init__.py ADDED
@@ -0,0 +1,65 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2023 The HuggingFace Team. All rights reserved.
2
+ #
3
+ # Licensed under the Apache License, Version 2.0 (the "License");
4
+ # you may not use this file except in compliance with the License.
5
+ # You may obtain a copy of the License at
6
+ #
7
+ # http://www.apache.org/licenses/LICENSE-2.0
8
+ #
9
+ # Unless required by applicable law or agreed to in writing, software
10
+ # distributed under the License is distributed on an "AS IS" BASIS,
11
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12
+ # See the License for the specific language governing permissions and
13
+ # limitations under the License.
14
+
15
+ from typing import TYPE_CHECKING
16
+
17
+ from ...utils import (
18
+ OptionalDependencyNotAvailable,
19
+ _LazyModule,
20
+ is_torch_available,
21
+ )
22
+
23
+
24
+ _import_structure = {
25
+ "configuration_univnet": [
26
+ "UNIVNET_PRETRAINED_CONFIG_ARCHIVE_MAP",
27
+ "UnivNetConfig",
28
+ ],
29
+ "feature_extraction_univnet": ["UnivNetFeatureExtractor"],
30
+ }
31
+
32
+ try:
33
+ if not is_torch_available():
34
+ raise OptionalDependencyNotAvailable()
35
+ except OptionalDependencyNotAvailable:
36
+ pass
37
+ else:
38
+ _import_structure["modeling_univnet"] = [
39
+ "UNIVNET_PRETRAINED_MODEL_ARCHIVE_LIST",
40
+ "UnivNetModel",
41
+ ]
42
+
43
+
44
+ if TYPE_CHECKING:
45
+ from .configuration_univnet import (
46
+ UNIVNET_PRETRAINED_CONFIG_ARCHIVE_MAP,
47
+ UnivNetConfig,
48
+ )
49
+ from .feature_extraction_univnet import UnivNetFeatureExtractor
50
+
51
+ try:
52
+ if not is_torch_available():
53
+ raise OptionalDependencyNotAvailable()
54
+ except OptionalDependencyNotAvailable:
55
+ pass
56
+ else:
57
+ from .modeling_univnet import (
58
+ UNIVNET_PRETRAINED_MODEL_ARCHIVE_LIST,
59
+ UnivNetModel,
60
+ )
61
+
62
+ else:
63
+ import sys
64
+
65
+ sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)
evalkit_cambrian/lib/python3.10/site-packages/transformers/models/univnet/__pycache__/feature_extraction_univnet.cpython-310.pyc ADDED
Binary file (19.5 kB). View file
 
evalkit_cambrian/lib/python3.10/site-packages/transformers/models/univnet/convert_univnet.py ADDED
@@ -0,0 +1,162 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2023 The HuggingFace Team. All rights reserved.
2
+ #
3
+ # Licensed under the Apache License, Version 2.0 (the "License");
4
+ # you may not use this file except in compliance with the License.
5
+ # You may obtain a copy of the License at
6
+ #
7
+ # http://www.apache.org/licenses/LICENSE-2.0
8
+ #
9
+ # Unless required by applicable law or agreed to in writing, software
10
+ # distributed under the License is distributed on an "AS IS" BASIS,
11
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12
+ # See the License for the specific language governing permissions and
13
+ # limitations under the License.
14
+
15
+ import argparse
16
+
17
+ import torch
18
+
19
+ from transformers import UnivNetConfig, UnivNetModel, logging
20
+
21
+
22
+ logging.set_verbosity_info()
23
+ logger = logging.get_logger("transformers.models.univnet")
24
+
25
+
26
+ def get_kernel_predictor_key_mapping(config: UnivNetConfig, old_prefix: str = "", new_prefix: str = ""):
27
+ mapping = {}
28
+ # Initial conv layer
29
+ mapping[f"{old_prefix}.input_conv.0.weight_g"] = f"{new_prefix}.input_conv.weight_g"
30
+ mapping[f"{old_prefix}.input_conv.0.weight_v"] = f"{new_prefix}.input_conv.weight_v"
31
+ mapping[f"{old_prefix}.input_conv.0.bias"] = f"{new_prefix}.input_conv.bias"
32
+
33
+ # Kernel predictor resnet blocks
34
+ for i in range(config.kernel_predictor_num_blocks):
35
+ mapping[f"{old_prefix}.residual_convs.{i}.1.weight_g"] = f"{new_prefix}.resblocks.{i}.conv1.weight_g"
36
+ mapping[f"{old_prefix}.residual_convs.{i}.1.weight_v"] = f"{new_prefix}.resblocks.{i}.conv1.weight_v"
37
+ mapping[f"{old_prefix}.residual_convs.{i}.1.bias"] = f"{new_prefix}.resblocks.{i}.conv1.bias"
38
+
39
+ mapping[f"{old_prefix}.residual_convs.{i}.3.weight_g"] = f"{new_prefix}.resblocks.{i}.conv2.weight_g"
40
+ mapping[f"{old_prefix}.residual_convs.{i}.3.weight_v"] = f"{new_prefix}.resblocks.{i}.conv2.weight_v"
41
+ mapping[f"{old_prefix}.residual_convs.{i}.3.bias"] = f"{new_prefix}.resblocks.{i}.conv2.bias"
42
+
43
+ # Kernel output conv
44
+ mapping[f"{old_prefix}.kernel_conv.weight_g"] = f"{new_prefix}.kernel_conv.weight_g"
45
+ mapping[f"{old_prefix}.kernel_conv.weight_v"] = f"{new_prefix}.kernel_conv.weight_v"
46
+ mapping[f"{old_prefix}.kernel_conv.bias"] = f"{new_prefix}.kernel_conv.bias"
47
+
48
+ # Bias output conv
49
+ mapping[f"{old_prefix}.bias_conv.weight_g"] = f"{new_prefix}.bias_conv.weight_g"
50
+ mapping[f"{old_prefix}.bias_conv.weight_v"] = f"{new_prefix}.bias_conv.weight_v"
51
+ mapping[f"{old_prefix}.bias_conv.bias"] = f"{new_prefix}.bias_conv.bias"
52
+
53
+ return mapping
54
+
55
+
56
+ def get_key_mapping(config: UnivNetConfig):
57
+ mapping = {}
58
+
59
+ # NOTE: inital conv layer keys are the same
60
+
61
+ # LVC Residual blocks
62
+ for i in range(len(config.resblock_stride_sizes)):
63
+ # LVCBlock initial convt layer
64
+ mapping[f"res_stack.{i}.convt_pre.1.weight_g"] = f"resblocks.{i}.convt_pre.weight_g"
65
+ mapping[f"res_stack.{i}.convt_pre.1.weight_v"] = f"resblocks.{i}.convt_pre.weight_v"
66
+ mapping[f"res_stack.{i}.convt_pre.1.bias"] = f"resblocks.{i}.convt_pre.bias"
67
+
68
+ # Kernel predictor
69
+ kernel_predictor_mapping = get_kernel_predictor_key_mapping(
70
+ config, old_prefix=f"res_stack.{i}.kernel_predictor", new_prefix=f"resblocks.{i}.kernel_predictor"
71
+ )
72
+ mapping.update(kernel_predictor_mapping)
73
+
74
+ # LVC Residual blocks
75
+ for j in range(len(config.resblock_dilation_sizes[i])):
76
+ mapping[f"res_stack.{i}.conv_blocks.{j}.1.weight_g"] = f"resblocks.{i}.resblocks.{j}.conv.weight_g"
77
+ mapping[f"res_stack.{i}.conv_blocks.{j}.1.weight_v"] = f"resblocks.{i}.resblocks.{j}.conv.weight_v"
78
+ mapping[f"res_stack.{i}.conv_blocks.{j}.1.bias"] = f"resblocks.{i}.resblocks.{j}.conv.bias"
79
+
80
+ # Output conv layer
81
+ mapping["conv_post.1.weight_g"] = "conv_post.weight_g"
82
+ mapping["conv_post.1.weight_v"] = "conv_post.weight_v"
83
+ mapping["conv_post.1.bias"] = "conv_post.bias"
84
+
85
+ return mapping
86
+
87
+
88
+ def rename_state_dict(state_dict, keys_to_modify, keys_to_remove):
89
+ model_state_dict = {}
90
+ for key, value in state_dict.items():
91
+ if key in keys_to_remove:
92
+ continue
93
+
94
+ if key in keys_to_modify:
95
+ new_key = keys_to_modify[key]
96
+ model_state_dict[new_key] = value
97
+ else:
98
+ model_state_dict[key] = value
99
+ return model_state_dict
100
+
101
+
102
+ def convert_univnet_checkpoint(
103
+ checkpoint_path,
104
+ pytorch_dump_folder_path,
105
+ config_path=None,
106
+ repo_id=None,
107
+ safe_serialization=False,
108
+ ):
109
+ model_state_dict_base = torch.load(checkpoint_path, map_location="cpu")
110
+ # Get the generator's state dict
111
+ state_dict = model_state_dict_base["model_g"]
112
+
113
+ if config_path is not None:
114
+ config = UnivNetConfig.from_pretrained(config_path)
115
+ else:
116
+ config = UnivNetConfig()
117
+
118
+ keys_to_modify = get_key_mapping(config)
119
+ keys_to_remove = set()
120
+ hf_state_dict = rename_state_dict(state_dict, keys_to_modify, keys_to_remove)
121
+
122
+ model = UnivNetModel(config)
123
+ # Apply weight norm since the original checkpoint has weight norm applied
124
+ model.apply_weight_norm()
125
+ model.load_state_dict(hf_state_dict)
126
+ # Remove weight norm in preparation for inference
127
+ model.remove_weight_norm()
128
+
129
+ model.save_pretrained(pytorch_dump_folder_path, safe_serialization=safe_serialization)
130
+
131
+ if repo_id:
132
+ print("Pushing to the hub...")
133
+ model.push_to_hub(repo_id)
134
+
135
+
136
+ def main():
137
+ parser = argparse.ArgumentParser()
138
+ parser.add_argument("--checkpoint_path", required=True, default=None, type=str, help="Path to original checkpoint")
139
+ parser.add_argument("--config_path", default=None, type=str, help="Path to hf config.json of model to convert")
140
+ parser.add_argument(
141
+ "--pytorch_dump_folder_path", required=True, default=None, type=str, help="Path to the output PyTorch model."
142
+ )
143
+ parser.add_argument(
144
+ "--push_to_hub", default=None, type=str, help="Where to upload the converted model on the 🤗 hub."
145
+ )
146
+ parser.add_argument(
147
+ "--safe_serialization", action="store_true", help="Whether to save the model using `safetensors`."
148
+ )
149
+
150
+ args = parser.parse_args()
151
+
152
+ convert_univnet_checkpoint(
153
+ args.checkpoint_path,
154
+ args.pytorch_dump_folder_path,
155
+ args.config_path,
156
+ args.push_to_hub,
157
+ args.safe_serialization,
158
+ )
159
+
160
+
161
+ if __name__ == "__main__":
162
+ main()
evalkit_cambrian/lib/python3.10/site-packages/transformers/models/univnet/feature_extraction_univnet.py ADDED
@@ -0,0 +1,456 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2023 The HuggingFace Team. All rights reserved.
2
+ #
3
+ # Licensed under the Apache License, Version 2.0 (the "License");
4
+ # you may not use this file except in compliance with the License.
5
+ # You may obtain a copy of the License at
6
+ #
7
+ # http://www.apache.org/licenses/LICENSE-2.0
8
+ #
9
+ # Unless required by applicable law or agreed to in writing, software
10
+ # distributed under the License is distributed on an "AS IS" BASIS,
11
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12
+ # See the License for the specific language governing permissions and
13
+ # limitations under the License.
14
+ """Feature extractor class for UnivNetModel."""
15
+
16
+ from typing import Any, Dict, List, Optional, Union
17
+
18
+ import numpy as np
19
+
20
+ from ...audio_utils import mel_filter_bank, optimal_fft_length, spectrogram, window_function
21
+ from ...feature_extraction_sequence_utils import SequenceFeatureExtractor
22
+ from ...feature_extraction_utils import BatchFeature
23
+ from ...utils import PaddingStrategy, TensorType, logging
24
+
25
+
26
+ logger = logging.get_logger(__name__)
27
+
28
+
29
+ class UnivNetFeatureExtractor(SequenceFeatureExtractor):
30
+ r"""
31
+ Constructs a UnivNet feature extractor.
32
+
33
+ This class extracts log-mel-filter bank features from raw speech using the short time Fourier Transform (STFT). The
34
+ STFT implementation follows that of TacoTron 2 and Hifi-GAN.
35
+
36
+ This feature extractor inherits from [`~feature_extraction_sequence_utils.SequenceFeatureExtractor`] which contains
37
+ most of the main methods. Users should refer to this superclass for more information regarding those methods.
38
+
39
+ Args:
40
+ feature_size (`int`, *optional*, defaults to 1):
41
+ The feature dimension of the extracted features.
42
+ sampling_rate (`int`, *optional*, defaults to 24000):
43
+ The sampling rate at which the audio files should be digitalized expressed in hertz (Hz).
44
+ padding_value (`float`, *optional*, defaults to 0.0):
45
+ The value to pad with when applying the padding strategy defined by the `padding` argument to
46
+ [`UnivNetFeatureExtractor.__call__`]. Should correspond to audio silence. The `pad_end` argument to
47
+ `__call__` will also use this padding value.
48
+ do_normalize (`bool`, *optional*, defaults to `False`):
49
+ Whether to perform Tacotron 2 normalization on the input. Normalizing can help to significantly improve the
50
+ performance for some models.
51
+ num_mel_bins (`int`, *optional*, defaults to 100):
52
+ The number of mel-frequency bins in the extracted spectrogram features. This should match
53
+ `UnivNetModel.config.num_mel_bins`.
54
+ hop_length (`int`, *optional*, defaults to 256):
55
+ The direct number of samples between sliding windows. Otherwise referred to as "shift" in many papers. Note
56
+ that this is different from other audio feature extractors such as [`SpeechT5FeatureExtractor`] which take
57
+ the `hop_length` in ms.
58
+ win_length (`int`, *optional*, defaults to 1024):
59
+ The direct number of samples for each sliding window. Note that this is different from other audio feature
60
+ extractors such as [`SpeechT5FeatureExtractor`] which take the `win_length` in ms.
61
+ win_function (`str`, *optional*, defaults to `"hann_window"`):
62
+ Name for the window function used for windowing, must be accessible via `torch.{win_function}`
63
+ filter_length (`int`, *optional*, defaults to 1024):
64
+ The number of FFT components to use. If `None`, this is determined using
65
+ `transformers.audio_utils.optimal_fft_length`.
66
+ max_length_s (`int`, *optional*, defaults to 10):
67
+ The maximum input lenght of the model in seconds. This is used to pad the audio.
68
+ fmin (`float`, *optional*, defaults to 0.0):
69
+ Minimum mel frequency in Hz.
70
+ fmax (`float`, *optional*):
71
+ Maximum mel frequency in Hz. If not set, defaults to `sampling_rate / 2`.
72
+ mel_floor (`float`, *optional*, defaults to 1e-09):
73
+ Minimum value of mel frequency banks. Note that the way [`UnivNetFeatureExtractor`] uses `mel_floor` is
74
+ different than in [`transformers.audio_utils.spectrogram`].
75
+ center (`bool`, *optional*, defaults to `False`):
76
+ Whether to pad the waveform so that frame `t` is centered around time `t * hop_length`. If `False`, frame
77
+ `t` will start at time `t * hop_length`.
78
+ compression_factor (`float`, *optional*, defaults to 1.0):
79
+ The multiplicative compression factor for dynamic range compression during spectral normalization.
80
+ compression_clip_val (`float`, *optional*, defaults to 1e-05):
81
+ The clip value applied to the waveform before applying dynamic range compression during spectral
82
+ normalization.
83
+ normalize_min (`float`, *optional*, defaults to -11.512925148010254):
84
+ The min value used for Tacotron 2-style linear normalization. The default is the original value from the
85
+ Tacotron 2 implementation.
86
+ normalize_max (`float`, *optional*, defaults to 2.3143386840820312):
87
+ The max value used for Tacotron 2-style linear normalization. The default is the original value from the
88
+ Tacotron 2 implementation.
89
+ model_in_channels (`int`, *optional*, defaults to 64):
90
+ The number of input channels to the [`UnivNetModel`] model. This should match
91
+ `UnivNetModel.config.model_in_channels`.
92
+ pad_end_length (`int`, *optional*, defaults to 10):
93
+ If padding the end of each waveform, the number of spectrogram frames worth of samples to append. The
94
+ number of appended samples will be `pad_end_length * hop_length`.
95
+ return_attention_mask (`bool`, *optional*, defaults to `True`):
96
+ Whether or not [`~UnivNetFeatureExtractor.__call__`] should return `attention_mask`.
97
+ """
98
+
99
+ model_input_names = ["input_features", "noise_sequence", "padding_mask"]
100
+
101
+ def __init__(
102
+ self,
103
+ feature_size: int = 1,
104
+ sampling_rate: int = 24000,
105
+ padding_value: float = 0.0,
106
+ do_normalize: bool = False,
107
+ num_mel_bins: int = 100,
108
+ hop_length: int = 256,
109
+ win_length: int = 1024,
110
+ win_function: str = "hann_window",
111
+ filter_length: Optional[int] = 1024,
112
+ max_length_s: int = 10,
113
+ fmin: float = 0.0,
114
+ fmax: Optional[float] = None,
115
+ mel_floor: float = 1e-9,
116
+ center: bool = False,
117
+ compression_factor: float = 1.0,
118
+ compression_clip_val: float = 1e-5,
119
+ normalize_min: float = -11.512925148010254,
120
+ normalize_max: float = 2.3143386840820312,
121
+ model_in_channels: int = 64,
122
+ pad_end_length: int = 10,
123
+ return_attention_mask=True,
124
+ **kwargs,
125
+ ):
126
+ super().__init__(
127
+ feature_size=feature_size,
128
+ sampling_rate=sampling_rate,
129
+ padding_value=padding_value,
130
+ return_attention_mask=return_attention_mask,
131
+ **kwargs,
132
+ )
133
+
134
+ self.do_normalize = do_normalize
135
+
136
+ self.num_mel_bins = num_mel_bins
137
+ self.hop_length = hop_length
138
+ self.win_length = win_length
139
+ self.win_function = win_function
140
+ self.filter_length = filter_length
141
+ self.fmin = fmin
142
+ if fmax is None:
143
+ # Follows the librosa.filters.mel implementation
144
+ fmax = float(sampling_rate) / 2
145
+ self.fmax = fmax
146
+ self.mel_floor = mel_floor
147
+
148
+ self.max_length_s = max_length_s
149
+ self.num_max_samples = max_length_s * sampling_rate
150
+
151
+ if self.filter_length is None:
152
+ self.n_fft = optimal_fft_length(self.win_length)
153
+ else:
154
+ self.n_fft = self.filter_length
155
+ self.n_freqs = (self.n_fft // 2) + 1
156
+
157
+ self.window = window_function(window_length=self.win_length, name=self.win_function, periodic=True)
158
+
159
+ self.mel_filters = mel_filter_bank(
160
+ num_frequency_bins=self.n_freqs,
161
+ num_mel_filters=self.num_mel_bins,
162
+ min_frequency=self.fmin,
163
+ max_frequency=self.fmax,
164
+ sampling_rate=self.sampling_rate,
165
+ norm="slaney",
166
+ mel_scale="slaney",
167
+ )
168
+
169
+ self.center = center
170
+ self.compression_factor = compression_factor
171
+ self.compression_clip_val = compression_clip_val
172
+ self.normalize_min = normalize_min
173
+ self.normalize_max = normalize_max
174
+ self.model_in_channels = model_in_channels
175
+ self.pad_end_length = pad_end_length
176
+
177
+ def normalize(self, spectrogram):
178
+ return 2 * ((spectrogram - self.normalize_min) / (self.normalize_max - self.normalize_min)) - 1
179
+
180
+ def denormalize(self, spectrogram):
181
+ return self.normalize_min + (self.normalize_max - self.normalize_min) * ((spectrogram + 1) / 2)
182
+
183
+ def mel_spectrogram(self, waveform: np.ndarray) -> np.ndarray:
184
+ """
185
+ Calculates log MEL spectrograms from a batch of waveforms. Note that the input waveform(s) will be padded by
186
+ `int(self.n_fft - self.hop_length) / 2` on both sides using the `reflect` padding mode.
187
+
188
+ Args:
189
+ waveform (`np.ndarray` of shape `(length,)`):
190
+ The input waveform. This must be a single real-valued, mono waveform.
191
+
192
+ Returns:
193
+ `numpy.ndarray`: Array containing a log-mel spectrogram of shape `(num_frames, num_mel_bins)`.
194
+ """
195
+ # Do custom padding based on the official MelGAN and Hifi-GAN implementations
196
+ # See https://github.com/maum-ai/univnet/blob/9bb2b54838bb6d7ce767131cc7b8b61198bc7558/utils/stft.py#L84-L86
197
+ waveform = np.pad(
198
+ waveform,
199
+ (int((self.n_fft - self.hop_length) / 2), int((self.n_fft - self.hop_length) / 2)),
200
+ mode="reflect",
201
+ )
202
+
203
+ # Get the complex spectrogram.
204
+ # Note: waveform must be unbatched currently due to the implementation of spectrogram(...).
205
+ complex_spectrogram = spectrogram(
206
+ waveform,
207
+ window=self.window,
208
+ frame_length=self.n_fft,
209
+ hop_length=self.hop_length,
210
+ fft_length=self.n_fft,
211
+ power=None,
212
+ center=self.center,
213
+ mel_filters=None,
214
+ mel_floor=None,
215
+ )
216
+
217
+ # Apply the MEL filter bank and MEL floor manually since UnivNet uses a slightly different implementation
218
+ amplitude_spectrogram = np.sqrt(
219
+ np.real(complex_spectrogram) ** 2 + np.imag(complex_spectrogram) ** 2 + self.mel_floor
220
+ )
221
+ mel_spectrogram = np.matmul(self.mel_filters.T, amplitude_spectrogram)
222
+
223
+ # Perform spectral normalization to get the log mel spectrogram.
224
+ log_mel_spectrogram = np.log(
225
+ np.clip(mel_spectrogram, a_min=self.compression_clip_val, a_max=None) * self.compression_factor
226
+ )
227
+
228
+ # Return spectrogram with num_mel_bins last
229
+ return log_mel_spectrogram.T
230
+
231
+ def generate_noise(
232
+ self,
233
+ noise_length: int,
234
+ generator: Optional[np.random.Generator] = None,
235
+ ) -> np.ndarray:
236
+ """
237
+ Generates a random noise sequence of standard Gaussian noise for use in the `noise_sequence` argument of
238
+ [`UnivNetModel.forward`].
239
+
240
+ Args:
241
+ spectrogram_length (`int`):
242
+ The length (dim 0) of the generated noise.
243
+ model_in_channels (`int`, *optional*, defaults to `None`):
244
+ The number of features (dim 1) of the generated noise. This should correspond to the
245
+ `model_in_channels` of the [`UnivNetGan`] model. If not set, this will default to
246
+ `self.config.model_in_channels`.
247
+ generator (`numpy.random.Generator`, *optional*, defaults to `None`)
248
+ An optional `numpy.random.Generator` random number generator to control noise generation. If not set, a
249
+ new generator with fresh entropy will be created.
250
+
251
+ Returns:
252
+ `numpy.ndarray`: Array containing random standard Gaussian noise of shape `(noise_length,
253
+ model_in_channels)`.
254
+ """
255
+ if generator is None:
256
+ generator = np.random.default_rng()
257
+
258
+ noise_shape = (noise_length, self.model_in_channels)
259
+ noise = generator.standard_normal(noise_shape, dtype=np.float32)
260
+
261
+ return noise
262
+
263
+ def batch_decode(self, waveforms, waveform_lengths=None) -> List[np.ndarray]:
264
+ r"""
265
+ Removes padding from generated audio after running [`UnivNetModel.forward`]. This returns a ragged list of 1D
266
+ audio waveform arrays and not a single tensor/array because in general the waveforms will have different
267
+ lengths after removing padding.
268
+
269
+ Args:
270
+ waveforms (`torch.FloatTensor` of shape `(batch_size, sequence_length)`):
271
+ The batched output waveforms from the [`UnivNetModel`].
272
+ waveform_lengths (`torch.FloatTensor` of shape `(batch_size,)`, *optional*):
273
+ The batched lengths of each waveform before padding.
274
+
275
+ Returns:
276
+ `List[np.ndarray]`: A ragged list of 1D waveform arrays with padding removed.
277
+ """
278
+ # Collapse the batched waveform tensor to a list of 1D audio waveforms
279
+ waveforms = [waveform.detach().clone().cpu().numpy() for waveform in waveforms]
280
+
281
+ if waveform_lengths is not None:
282
+ waveforms = [waveform[: waveform_lengths[i]] for i, waveform in enumerate(waveforms)]
283
+
284
+ return waveforms
285
+
286
+ def __call__(
287
+ self,
288
+ raw_speech: Union[np.ndarray, List[float], List[np.ndarray], List[List[float]]],
289
+ sampling_rate: Optional[int] = None,
290
+ padding: Union[bool, str, PaddingStrategy] = True,
291
+ max_length: Optional[int] = None,
292
+ truncation: bool = True,
293
+ pad_to_multiple_of: Optional[int] = None,
294
+ return_noise: bool = True,
295
+ generator: Optional[np.random.Generator] = None,
296
+ pad_end: bool = False,
297
+ pad_length: Optional[int] = None,
298
+ do_normalize: Optional[str] = None,
299
+ return_attention_mask: Optional[bool] = None,
300
+ return_tensors: Optional[Union[str, TensorType]] = None,
301
+ ) -> BatchFeature:
302
+ """
303
+ Main method to featurize and prepare for the model one or several sequence(s).
304
+
305
+ Args:
306
+ raw_speech (`np.ndarray`, `List[float]`, `List[np.ndarray]`, `List[List[float]]`):
307
+ The sequence or batch of sequences to be padded. Each sequence can be a numpy array, a list of float
308
+ values, a list of numpy arrays or a list of list of float values. Must be mono channel audio, not
309
+ stereo, i.e. single float per timestep.
310
+ sampling_rate (`int`, *optional*):
311
+ The sampling rate at which the `raw_speech` input was sampled. It is strongly recommended to pass
312
+ `sampling_rate` at the forward call to prevent silent errors and allow automatic speech recognition
313
+ pipeline.
314
+ padding (`bool`, `str` or [`~utils.PaddingStrategy`], *optional*, defaults to `True`):
315
+ Select a strategy to pad the input `raw_speech` waveforms (according to the model's padding side and
316
+ padding index) among:
317
+
318
+ - `True` or `'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
319
+ sequence if provided).
320
+ - `'max_length'`: Pad to a maximum length specified with the argument `max_length` or to the maximum
321
+ acceptable input length for the model if that argument is not provided.
322
+ - `False` or `'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of different
323
+ lengths).
324
+
325
+ If `pad_end = True`, that padding will occur before the `padding` strategy is applied.
326
+ max_length (`int`, *optional*):
327
+ Maximum length of the returned list and optionally padding length (see above).
328
+ truncation (`bool`, *optional*, defaults to `True`):
329
+ Activates truncation to cut input sequences longer than `max_length` to `max_length`.
330
+ pad_to_multiple_of (`int`, *optional*):
331
+ If set will pad the sequence to a multiple of the provided value.
332
+
333
+ This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability
334
+ `>= 7.5` (Volta), or on TPUs which benefit from having sequence lengths be a multiple of 128.
335
+ return_noise (`bool`, *optional*, defaults to `True`):
336
+ Whether to generate and return a noise waveform for use in [`UnivNetModel.forward`].
337
+ generator (`numpy.random.Generator`, *optional*, defaults to `None`):
338
+ An optional `numpy.random.Generator` random number generator to use when generating noise.
339
+ pad_end (`bool`, *optional*, defaults to `False`):
340
+ Whether to pad the end of each waveform with silence. This can help reduce artifacts at the end of the
341
+ generated audio sample; see https://github.com/seungwonpark/melgan/issues/8 for more details. This
342
+ padding will be done before the padding strategy specified in `padding` is performed.
343
+ pad_length (`int`, *optional*, defaults to `None`):
344
+ If padding the end of each waveform, the length of the padding in spectrogram frames. If not set, this
345
+ will default to `self.config.pad_end_length`.
346
+ do_normalize (`bool`, *optional*):
347
+ Whether to perform Tacotron 2 normalization on the input. Normalizing can help to significantly improve
348
+ the performance for some models. If not set, this will default to `self.config.do_normalize`.
349
+ return_attention_mask (`bool`, *optional*):
350
+ Whether to return the attention mask. If left to the default, will return the attention mask according
351
+ to the specific feature_extractor's default.
352
+
353
+ [What are attention masks?](../glossary#attention-mask)
354
+
355
+ return_tensors (`str` or [`~utils.TensorType`], *optional*):
356
+ If set, will return tensors instead of list of python integers. Acceptable values are:
357
+
358
+ - `'tf'`: Return TensorFlow `tf.constant` objects.
359
+ - `'pt'`: Return PyTorch `torch.np.array` objects.
360
+ - `'np'`: Return Numpy `np.ndarray` objects.
361
+ """
362
+ do_normalize = do_normalize if do_normalize is not None else self.do_normalize
363
+
364
+ if sampling_rate is not None:
365
+ if sampling_rate != self.sampling_rate:
366
+ raise ValueError(
367
+ f"The model corresponding to this feature extractor: {self.__class__.__name__} was trained using a"
368
+ f" sampling rate of {self.sampling_rate}. Please make sure that the provided `raw_speech` input"
369
+ f" was sampled with {self.sampling_rate} and not {sampling_rate}."
370
+ )
371
+ else:
372
+ logger.warning(
373
+ "It is strongly recommended to pass the `sampling_rate` argument to this function. "
374
+ "Failing to do so can result in silent errors that might be hard to debug."
375
+ )
376
+
377
+ is_batched_numpy = isinstance(raw_speech, np.ndarray) and len(raw_speech.shape) > 1
378
+ if is_batched_numpy and len(raw_speech.shape) > 2:
379
+ raise ValueError(f"Only mono-channel audio is supported for input to {self}")
380
+ is_batched = is_batched_numpy or (
381
+ isinstance(raw_speech, (list, tuple)) and (isinstance(raw_speech[0], (np.ndarray, tuple, list)))
382
+ )
383
+
384
+ if is_batched:
385
+ raw_speech = [np.asarray(speech, dtype=np.float32) for speech in raw_speech]
386
+ elif not is_batched and not isinstance(raw_speech, np.ndarray):
387
+ raw_speech = np.asarray(raw_speech, dtype=np.float32)
388
+ elif isinstance(raw_speech, np.ndarray) and raw_speech.dtype is np.dtype(np.float64):
389
+ raw_speech = raw_speech.astype(np.float32)
390
+
391
+ # always return batch
392
+ if not is_batched:
393
+ raw_speech = [np.asarray(raw_speech, dtype=np.float32)]
394
+
395
+ # Pad end to reduce artifacts
396
+ if pad_end:
397
+ pad_length = pad_length if pad_length is not None else self.pad_end_length
398
+ raw_speech = [
399
+ np.pad(waveform, (0, pad_length * self.hop_length), constant_values=self.padding_value)
400
+ for waveform in raw_speech
401
+ ]
402
+
403
+ batched_speech = BatchFeature({"input_features": raw_speech})
404
+
405
+ padded_inputs = self.pad(
406
+ batched_speech,
407
+ padding=padding,
408
+ max_length=max_length if max_length is not None else self.num_max_samples,
409
+ truncation=truncation,
410
+ pad_to_multiple_of=pad_to_multiple_of,
411
+ return_attention_mask=return_attention_mask,
412
+ )
413
+
414
+ # make sure list is in array format
415
+ # input_features = padded_inputs.get("input_features").transpose(2, 0, 1)
416
+ input_features = padded_inputs.get("input_features")
417
+
418
+ mel_spectrograms = [self.mel_spectrogram(waveform) for waveform in input_features]
419
+
420
+ if isinstance(input_features[0], List):
421
+ batched_speech["input_features"] = [np.asarray(mel, dtype=np.float32) for mel in mel_spectrograms]
422
+ else:
423
+ batched_speech["input_features"] = [mel.astype(np.float32) for mel in mel_spectrograms]
424
+
425
+ # convert attention_mask to correct format
426
+ attention_mask = padded_inputs.get("attention_mask")
427
+ if attention_mask is not None:
428
+ batched_speech["padding_mask"] = [np.asarray(array, dtype=np.int32) for array in attention_mask]
429
+
430
+ if return_noise:
431
+ noise = [
432
+ self.generate_noise(spectrogram.shape[0], generator)
433
+ for spectrogram in batched_speech["input_features"]
434
+ ]
435
+ batched_speech["noise_sequence"] = noise
436
+
437
+ if do_normalize:
438
+ batched_speech["input_features"] = [
439
+ self.normalize(spectrogram) for spectrogram in batched_speech["input_features"]
440
+ ]
441
+
442
+ if return_tensors is not None:
443
+ batched_speech = batched_speech.convert_to_tensors(return_tensors)
444
+
445
+ return batched_speech
446
+
447
+ def to_dict(self) -> Dict[str, Any]:
448
+ output = super().to_dict()
449
+
450
+ # Don't serialize these as they are derived from the other properties.
451
+ names = ["window", "mel_filters", "n_fft", "n_freqs", "num_max_samples"]
452
+ for name in names:
453
+ if name in output:
454
+ del output[name]
455
+
456
+ return output
evalkit_cambrian/lib/python3.10/site-packages/transformers/models/univnet/modeling_univnet.py ADDED
@@ -0,0 +1,636 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2023 The HuggingFace Team. All rights reserved.
2
+ #
3
+ # Licensed under the Apache License, Version 2.0 (the "License");
4
+ # you may not use this file except in compliance with the License.
5
+ # You may obtain a copy of the License at
6
+ #
7
+ # http://www.apache.org/licenses/LICENSE-2.0
8
+ #
9
+ # Unless required by applicable law or agreed to in writing, software
10
+ # distributed under the License is distributed on an "AS IS" BASIS,
11
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12
+ # See the License for the specific language governing permissions and
13
+ # limitations under the License.
14
+ """ PyTorch UnivNetModel model."""
15
+
16
+ from dataclasses import dataclass
17
+ from typing import Optional, Tuple, Union
18
+
19
+ import torch
20
+ import torch.utils.checkpoint
21
+ from torch import nn
22
+
23
+ from ...modeling_utils import ModelOutput, PreTrainedModel
24
+ from ...utils import add_start_docstrings, add_start_docstrings_to_model_forward, logging, replace_return_docstrings
25
+ from .configuration_univnet import UnivNetConfig
26
+
27
+
28
+ logger = logging.get_logger(__name__)
29
+
30
+ # General docstring
31
+ _CONFIG_FOR_DOC = "UnivNetConfig"
32
+
33
+ _CHECKPOINT_FOR_DOC = "dg845/univnet-dev"
34
+
35
+ UNIVNET_PRETRAINED_MODEL_ARCHIVE_LIST = [
36
+ "dg845/univnet-dev",
37
+ # See all UnivNet models at https://huggingface.co/models?filter=univnet
38
+ ]
39
+
40
+
41
+ @dataclass
42
+ class UnivNetModelOutput(ModelOutput):
43
+ """
44
+ Output class for the [`UnivNetModel`], which includes the generated audio waveforms and the original unpadded
45
+ lengths of those waveforms (so that the padding can be removed by [`UnivNetModel.batch_decode`]).
46
+
47
+ Args:
48
+ waveforms (`torch.FloatTensor` of shape `(batch_size, sequence_length)`):
49
+ Batched 1D (mono-channel) output audio waveforms.
50
+ waveform_lengths (`torch.FloatTensor` of shape `(batch_size,)`):
51
+ The batched length in samples of each unpadded waveform in `waveforms`.
52
+ """
53
+
54
+ waveforms: torch.FloatTensor = None
55
+ waveform_lengths: torch.FloatTensor = None
56
+
57
+
58
+ class UnivNetKernelPredictorResidualBlock(nn.Module):
59
+ """
60
+ Implementation of the residual block for the kernel predictor network inside each location variable convolution
61
+ block (LVCBlock).
62
+
63
+ Parameters:
64
+ config: (`UnivNetConfig`):
65
+ Config for the `UnivNetModel` model.
66
+ """
67
+
68
+ def __init__(
69
+ self,
70
+ config: UnivNetConfig,
71
+ ):
72
+ super().__init__()
73
+ self.channels = config.model_in_channels
74
+ self.kernel_size = config.kernel_predictor_conv_size
75
+ self.dropout_prob = config.kernel_predictor_dropout
76
+ self.leaky_relu_slope = config.leaky_relu_slope
77
+
78
+ padding = (self.kernel_size - 1) // 2
79
+
80
+ self.dropout = nn.Dropout(self.dropout_prob)
81
+ self.conv1 = nn.Conv1d(self.channels, self.channels, self.kernel_size, padding=padding, bias=True)
82
+ self.conv2 = nn.Conv1d(self.channels, self.channels, self.kernel_size, padding=padding, bias=True)
83
+
84
+ def forward(self, hidden_states: torch.FloatTensor):
85
+ # hidden_states should have shape (batch_size, channels, seq_length)
86
+ residual = hidden_states
87
+ hidden_states = self.dropout(hidden_states)
88
+ hidden_states = self.conv1(hidden_states)
89
+ hidden_states = nn.functional.leaky_relu(hidden_states, self.leaky_relu_slope)
90
+ hidden_states = self.conv2(hidden_states)
91
+ hidden_states = nn.functional.leaky_relu(hidden_states, self.leaky_relu_slope)
92
+ return hidden_states + residual
93
+
94
+ def apply_weight_norm(self):
95
+ nn.utils.weight_norm(self.conv1)
96
+ nn.utils.weight_norm(self.conv2)
97
+
98
+ def remove_weight_norm(self):
99
+ nn.utils.remove_weight_norm(self.conv1)
100
+ nn.utils.remove_weight_norm(self.conv2)
101
+
102
+
103
+ class UnivNetKernelPredictor(nn.Module):
104
+ """
105
+ Implementation of the kernel predictor network which supplies the kernel and bias for the location variable
106
+ convolutional layers (LVCs) in each UnivNet LVCBlock.
107
+
108
+ Based on the KernelPredictor implementation in
109
+ [maum-ai/univnet](https://github.com/maum-ai/univnet/blob/9bb2b54838bb6d7ce767131cc7b8b61198bc7558/model/lvcnet.py#L7).
110
+
111
+ Parameters:
112
+ config: (`UnivNetConfig`):
113
+ Config for the `UnivNetModel` model.
114
+ conv_kernel_size (`int`, *optional*, defaults to 3):
115
+ The kernel size for the location variable convolutional layer kernels (convolutional weight tensor).
116
+ conv_layers (`int`, *optional*, defaults to 4):
117
+ The number of location variable convolutional layers to output kernels and biases for.
118
+ """
119
+
120
+ def __init__(
121
+ self,
122
+ config: UnivNetConfig,
123
+ conv_kernel_size: int = 3,
124
+ conv_layers: int = 4,
125
+ ):
126
+ super().__init__()
127
+
128
+ self.conv_in_channels = config.model_hidden_channels
129
+ self.conv_out_channels = 2 * config.model_hidden_channels
130
+ self.conv_kernel_size = conv_kernel_size
131
+ self.conv_layers = conv_layers
132
+
133
+ self.kernel_channels = (
134
+ self.conv_in_channels * self.conv_out_channels * self.conv_kernel_size * self.conv_layers
135
+ )
136
+ self.bias_channels = self.conv_out_channels * self.conv_layers
137
+
138
+ self.resnet_in_channels = config.num_mel_bins
139
+ self.resnet_hidden_channels = config.kernel_predictor_hidden_channels
140
+ self.resnet_kernel_size = config.kernel_predictor_conv_size
141
+ self.num_blocks = config.kernel_predictor_num_blocks
142
+
143
+ self.leaky_relu_slope = config.leaky_relu_slope
144
+
145
+ padding = (self.resnet_kernel_size - 1) // 2
146
+
147
+ self.input_conv = nn.Conv1d(self.resnet_in_channels, self.resnet_hidden_channels, 5, padding=2, bias=True)
148
+
149
+ self.resblocks = nn.ModuleList([UnivNetKernelPredictorResidualBlock(config) for _ in range(self.num_blocks)])
150
+
151
+ self.kernel_conv = nn.Conv1d(
152
+ self.resnet_hidden_channels, self.kernel_channels, self.resnet_kernel_size, padding=padding, bias=True
153
+ )
154
+ self.bias_conv = nn.Conv1d(
155
+ self.resnet_hidden_channels, self.bias_channels, self.resnet_kernel_size, padding=padding, bias=True
156
+ )
157
+
158
+ def forward(self, spectrogram: torch.FloatTensor):
159
+ """
160
+ Maps a conditioning log-mel spectrogram to a tensor of convolutional kernels and biases, for use in location
161
+ variable convolutional layers. Note that the input spectrogram should have shape (batch_size, input_channels,
162
+ seq_length).
163
+
164
+ Args:
165
+ spectrogram (`torch.FloatTensor` of shape `(batch_size, input_channels, seq_length)`):
166
+ Tensor containing the log-mel spectrograms.
167
+
168
+ Returns:
169
+ Tuple[`torch.FloatTensor, `torch.FloatTensor`]: tuple of tensors where the first element is the tensor of
170
+ location variable convolution kernels of shape `(batch_size, self.conv_layers, self.conv_in_channels,
171
+ self.conv_out_channels, self.conv_kernel_size, seq_length)` and the second element is the tensor of
172
+ location variable convolution biases of shape `(batch_size, self.conv_layers. self.conv_out_channels,
173
+ seq_length)`.
174
+ """
175
+ batch_size, _, seq_length = spectrogram.shape
176
+
177
+ hidden_states = self.input_conv(spectrogram)
178
+ hidden_states = nn.functional.leaky_relu(hidden_states, self.leaky_relu_slope)
179
+
180
+ for resblock in self.resblocks:
181
+ hidden_states = resblock(hidden_states)
182
+
183
+ kernel_hidden_states = self.kernel_conv(hidden_states)
184
+ bias_hidden_states = self.bias_conv(hidden_states)
185
+
186
+ # Reshape kernels and biases to appropriate shape
187
+ kernels = kernel_hidden_states.view(
188
+ batch_size,
189
+ self.conv_layers,
190
+ self.conv_in_channels,
191
+ self.conv_out_channels,
192
+ self.conv_kernel_size,
193
+ seq_length,
194
+ ).contiguous()
195
+ biases = bias_hidden_states.view(
196
+ batch_size,
197
+ self.conv_layers,
198
+ self.conv_out_channels,
199
+ seq_length,
200
+ ).contiguous()
201
+
202
+ return kernels, biases
203
+
204
+ def apply_weight_norm(self):
205
+ nn.utils.weight_norm(self.input_conv)
206
+ for layer in self.resblocks:
207
+ layer.apply_weight_norm()
208
+ nn.utils.weight_norm(self.kernel_conv)
209
+ nn.utils.weight_norm(self.bias_conv)
210
+
211
+ def remove_weight_norm(self):
212
+ nn.utils.remove_weight_norm(self.input_conv)
213
+ for layer in self.resblocks:
214
+ layer.remove_weight_norm()
215
+ nn.utils.remove_weight_norm(self.kernel_conv)
216
+ nn.utils.remove_weight_norm(self.bias_conv)
217
+
218
+
219
+ class UnivNetLvcResidualBlock(nn.Module):
220
+ """
221
+ Implementation of the location variable convolution (LVC) residual block for the UnivNet residual network.
222
+
223
+ Parameters:
224
+ config: (`UnivNetConfig`):
225
+ Config for the `UnivNetModel` model.
226
+ kernel_size (`int`):
227
+ The kernel size for the dilated 1D convolutional layer.
228
+ dilation (`int`):
229
+ The dilation for the dilated 1D convolutional layer.
230
+ """
231
+
232
+ def __init__(
233
+ self,
234
+ config: UnivNetConfig,
235
+ kernel_size: int,
236
+ dilation: int,
237
+ ):
238
+ super().__init__()
239
+ self.hidden_channels = config.model_hidden_channels
240
+ self.kernel_size = kernel_size
241
+ self.dilation = dilation
242
+ self.leaky_relu_slope = config.leaky_relu_slope
243
+
244
+ padding = self.dilation * (self.kernel_size - 1) // 2
245
+
246
+ self.conv = nn.Conv1d(
247
+ self.hidden_channels,
248
+ self.hidden_channels,
249
+ self.kernel_size,
250
+ padding=padding,
251
+ dilation=self.dilation,
252
+ )
253
+
254
+ def forward(self, hidden_states, kernel, bias, hop_size=256):
255
+ residual = hidden_states
256
+ hidden_states = nn.functional.leaky_relu(hidden_states, self.leaky_relu_slope)
257
+ hidden_states = self.conv(hidden_states)
258
+ hidden_states = nn.functional.leaky_relu(hidden_states, self.leaky_relu_slope)
259
+ hidden_states = self.location_variable_convolution(hidden_states, kernel, bias, hop_size=hop_size)
260
+ # Gated activation unit
261
+ hidden_states = torch.sigmoid(hidden_states[:, : self.hidden_channels, :]) * torch.tanh(
262
+ hidden_states[:, self.hidden_channels :, :]
263
+ )
264
+ # Skip connection
265
+ hidden_states = residual + hidden_states
266
+
267
+ return hidden_states
268
+
269
+ # Based on https://github.com/maum-ai/univnet/blob/9bb2b54838bb6d7ce767131cc7b8b61198bc7558/model/lvcnet.py#L171
270
+ def location_variable_convolution(
271
+ self,
272
+ hidden_states: torch.FloatTensor,
273
+ kernel: torch.FloatTensor,
274
+ bias: torch.FloatTensor,
275
+ dilation: int = 1,
276
+ hop_size: int = 256,
277
+ ):
278
+ """
279
+ Performs location-variable convolution operation on the input sequence (hidden_states) using the local
280
+ convolution kernel. This was introduced in [LVCNet: Efficient Condition-Dependent Modeling Network for Waveform
281
+ Generation](https://arxiv.org/abs/2102.10815) by Zhen Zheng, Jianzong Wang, Ning Cheng, and Jing Xiao.
282
+
283
+ Time: 414 μs ± 309 ns per loop (mean ± std. dev. of 7 runs, 1000 loops each), test on NVIDIA V100.
284
+
285
+ Args:
286
+ hidden_states (`torch.FloatTensor` of shape `(batch_size, in_channels, in_length)`):
287
+ The input sequence of shape (batch, in_channels, in_length).
288
+ kernel (`torch.FloatTensor` of shape `(batch_size, in_channels, out_channels, kernel_size, kernel_length)`):
289
+ The local convolution kernel of shape (batch, in_channels, out_channels, kernel_size, kernel_length).
290
+ bias (`torch.FloatTensor` of shape `(batch_size, out_channels, kernel_length)`):
291
+ The bias for the local convolution of shape (batch, out_channels, kernel_length).
292
+ dilation (`int`, *optional*, defaults to 1):
293
+ The dilation of convolution.
294
+ hop_size (`int`, *optional*, defaults to 256):
295
+ The hop_size of the conditioning sequence.
296
+ Returns:
297
+ `torch.FloatTensor`: the output sequence after performing local convolution with shape (batch_size,
298
+ out_channels, in_length).
299
+ """
300
+ batch, _, in_length = hidden_states.shape
301
+ batch, _, out_channels, kernel_size, kernel_length = kernel.shape
302
+ if in_length != (kernel_length * hop_size):
303
+ raise ValueError(
304
+ f"Dim 2 of `hidden_states` should be {kernel_length * hop_size}) but got {in_length}. Please check"
305
+ " `hidden_states` or `kernel` and `hop_size` to make sure they are correct."
306
+ )
307
+
308
+ padding = dilation * int((kernel_size - 1) / 2)
309
+
310
+ # (batch, in_channels, in_length + 2*padding)
311
+ hidden_states = nn.functional.pad(hidden_states, (padding, padding), "constant", 0)
312
+ # (batch, in_channels, kernel_length, hop_size + 2*padding)
313
+ hidden_states = hidden_states.unfold(2, hop_size + 2 * padding, hop_size)
314
+
315
+ if hop_size < dilation:
316
+ hidden_states = nn.functional.pad(hidden_states, (0, dilation), "constant", 0)
317
+ # (batch, in_channels, kernel_length, (hop_size + 2*padding)/dilation, dilation)
318
+ hidden_states = hidden_states.unfold(3, dilation, dilation)
319
+ hidden_states = hidden_states[:, :, :, :, :hop_size]
320
+ # (batch, in_channels, kernel_length, dilation, (hop_size + 2*padding)/dilation)
321
+ hidden_states = hidden_states.transpose(3, 4)
322
+ # (batch, in_channels, kernel_length, dilation, _, kernel_size)
323
+ hidden_states = hidden_states.unfold(4, kernel_size, 1)
324
+
325
+ # Apply local convolution kernel to hidden_states.
326
+ output_hidden_states = torch.einsum("bildsk,biokl->bolsd", hidden_states, kernel)
327
+
328
+ output_hidden_states = output_hidden_states.to(memory_format=torch.channels_last_3d)
329
+ bias = bias.unsqueeze(-1).unsqueeze(-1).to(memory_format=torch.channels_last_3d)
330
+ output_hidden_states = output_hidden_states + bias
331
+ output_hidden_states = output_hidden_states.contiguous().view(batch, out_channels, -1)
332
+
333
+ return output_hidden_states
334
+
335
+ def apply_weight_norm(self):
336
+ nn.utils.weight_norm(self.conv)
337
+
338
+ def remove_weight_norm(self):
339
+ nn.utils.remove_weight_norm(self.conv)
340
+
341
+
342
+ class UnivNetLvcBlock(nn.Module):
343
+ """
344
+ Implementation of the location variable convolution (LVC) residual block of the UnivNet residual block. Includes a
345
+ `UnivNetKernelPredictor` inside to predict the kernels and biases of the LVC layers.
346
+
347
+ Based on LVCBlock in
348
+ [maum-ai/univnet](https://github.com/maum-ai/univnet/blob/9bb2b54838bb6d7ce767131cc7b8b61198bc7558/model/lvcnet.py#L98)
349
+
350
+ Parameters:
351
+ config (`UnivNetConfig`):
352
+ Config for the `UnivNetModel` model.
353
+ layer_id (`int`):
354
+ An integer corresponding to the index of the current LVC resnet block layer. This should be between 0 and
355
+ `len(config.resblock_stride_sizes) - 1)` inclusive.
356
+ lvc_hop_size (`int`, *optional*, defaults to 256):
357
+ The hop size for the location variable convolutional layers.
358
+ """
359
+
360
+ def __init__(
361
+ self,
362
+ config: UnivNetConfig,
363
+ layer_id: int,
364
+ lvc_hop_size: int = 256,
365
+ ):
366
+ super().__init__()
367
+ self.hidden_channels = config.model_hidden_channels
368
+ self.kernel_size = config.resblock_kernel_sizes[layer_id]
369
+ self.stride = config.resblock_stride_sizes[layer_id]
370
+ self.dilations = config.resblock_dilation_sizes[layer_id]
371
+ self.cond_hop_length = lvc_hop_size
372
+ self.leaky_relu_slope = config.leaky_relu_slope
373
+ self.num_blocks = len(self.dilations)
374
+
375
+ self.convt_pre = nn.ConvTranspose1d(
376
+ self.hidden_channels,
377
+ self.hidden_channels,
378
+ 2 * self.stride,
379
+ stride=self.stride,
380
+ padding=self.stride // 2 + self.stride % 2,
381
+ output_padding=self.stride % 2,
382
+ )
383
+
384
+ self.kernel_predictor = UnivNetKernelPredictor(config, self.kernel_size, self.num_blocks)
385
+
386
+ self.resblocks = nn.ModuleList(
387
+ [UnivNetLvcResidualBlock(config, self.kernel_size, self.dilations[i]) for i in range(self.num_blocks)]
388
+ )
389
+
390
+ def forward(self, hidden_states: torch.FloatTensor, spectrogram: torch.FloatTensor):
391
+ # hidden_states: (batch_size, hidden_channels, seq_length)
392
+ # spectrogram: (batch_size, cond_channels, cond_length)
393
+ hidden_states = nn.functional.leaky_relu(hidden_states, self.leaky_relu_slope)
394
+ hidden_states = self.convt_pre(hidden_states)
395
+
396
+ kernels, biases = self.kernel_predictor(spectrogram)
397
+
398
+ for i, resblock in enumerate(self.resblocks):
399
+ kernel = kernels[:, i, :, :, :, :]
400
+ bias = biases[:, i, :, :]
401
+ hidden_states = resblock(hidden_states, kernel, bias, hop_size=self.cond_hop_length)
402
+
403
+ return hidden_states
404
+
405
+ def apply_weight_norm(self):
406
+ nn.utils.weight_norm(self.convt_pre)
407
+ self.kernel_predictor.apply_weight_norm()
408
+ for layer in self.resblocks:
409
+ layer.apply_weight_norm()
410
+
411
+ def remove_weight_norm(self):
412
+ nn.utils.remove_weight_norm(self.convt_pre)
413
+ self.kernel_predictor.remove_weight_norm()
414
+ for layer in self.resblocks:
415
+ layer.remove_weight_norm()
416
+
417
+
418
+ UNIVNET_START_DOCSTRING = r"""
419
+ This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
420
+ library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
421
+ etc.)
422
+
423
+ This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
424
+ Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
425
+ and behavior.
426
+
427
+ Parameters:
428
+ config ([`UnivNetConfig`]):
429
+ Model configuration class with all the parameters of the model. Initializing with a config file does not
430
+ load the weights associated with the model, only the configuration. Check out the
431
+ [`~PreTrainedModel.from_pretrained`] method to load the model weights.
432
+ """
433
+
434
+
435
+ UNIVNET_INPUTS_DOCSTRING = r"""
436
+ Converts a noise waveform and a conditioning spectrogram to a speech waveform. Passing a batch of log-mel
437
+ spectrograms returns a batch of speech waveforms. Passing a single, un-batched log-mel spectrogram returns a
438
+ single, un-batched speech waveform.
439
+
440
+ Args:
441
+ input_features (`torch.FloatTensor`):
442
+ Tensor containing the log-mel spectrograms. Can be batched and of shape `(batch_size, sequence_length,
443
+ config.num_mel_channels)`, or un-batched and of shape `(sequence_length, config.num_mel_channels)`.
444
+ noise_sequence (`torch.FloatTensor`, *optional*):
445
+ Tensor containing a noise sequence of standard Gaussian noise. Can be batched and of shape `(batch_size,
446
+ sequence_length, config.model_in_channels)`, or un-batched and of shape (sequence_length,
447
+ config.model_in_channels)`. If not supplied, will be randomly generated.
448
+ padding_mask (`torch.BoolTensor`, *optional*):
449
+ Mask indicating which parts of each sequence are padded. Mask values are selected in `[0, 1]`:
450
+
451
+ - 1 for tokens that are **not masked**
452
+ - 0 for tokens that are **masked**
453
+
454
+ The mask can be batched and of shape `(batch_size, sequence_length)` or un-batched and of shape
455
+ `(sequence_length,)`.
456
+ generator (`torch.Generator`, *optional*):
457
+ A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation
458
+ deterministic.
459
+ return_dict:
460
+ Whether to return a [`~utils.ModelOutput`] subclass instead of a plain tuple.
461
+ """
462
+
463
+
464
+ @add_start_docstrings(
465
+ """UnivNet GAN vocoder.""",
466
+ UNIVNET_START_DOCSTRING,
467
+ )
468
+ class UnivNetModel(PreTrainedModel):
469
+ config_class = UnivNetConfig
470
+ main_input_name = "input_features"
471
+
472
+ def __init__(self, config: UnivNetConfig):
473
+ super().__init__(config)
474
+
475
+ self.num_kernels = len(config.resblock_kernel_sizes)
476
+ self.leaky_relu_slope = config.leaky_relu_slope
477
+
478
+ self.conv_pre = nn.Conv1d(
479
+ config.model_in_channels,
480
+ config.model_hidden_channels,
481
+ kernel_size=7,
482
+ stride=1,
483
+ padding=3,
484
+ padding_mode="reflect",
485
+ )
486
+
487
+ # Initialize location-variable convolution ResNet Blocks.
488
+ num_layers = len(config.resblock_stride_sizes)
489
+ hop_length = 1
490
+ hop_lengths = []
491
+ for stride in config.resblock_stride_sizes:
492
+ hop_length = hop_length * stride
493
+ hop_lengths.append(hop_length)
494
+
495
+ self.resblocks = nn.ModuleList(
496
+ [
497
+ UnivNetLvcBlock(
498
+ config,
499
+ layer_id=i,
500
+ lvc_hop_size=hop_lengths[i],
501
+ )
502
+ for i in range(num_layers)
503
+ ]
504
+ )
505
+
506
+ self.conv_post = nn.Conv1d(config.model_hidden_channels, 1, 7, padding=3, padding_mode="reflect")
507
+
508
+ # Initialize weights and apply final processing
509
+ self.post_init()
510
+
511
+ @add_start_docstrings_to_model_forward(UNIVNET_INPUTS_DOCSTRING)
512
+ @replace_return_docstrings(output_type=UnivNetModelOutput, config_class=_CONFIG_FOR_DOC)
513
+ def forward(
514
+ self,
515
+ input_features: torch.FloatTensor,
516
+ noise_sequence: Optional[torch.FloatTensor] = None,
517
+ padding_mask: Optional[torch.FloatTensor] = None,
518
+ generator: Optional[torch.Generator] = None,
519
+ return_dict: Optional[bool] = None,
520
+ ) -> Union[Tuple[torch.FloatTensor], UnivNetModelOutput]:
521
+ r"""
522
+ Returns:
523
+
524
+ Example:
525
+
526
+ ```python
527
+ >>> from transformers import UnivNetFeatureExtractor, UnivNetModel
528
+ >>> from datasets import load_dataset, Audio
529
+
530
+ >>> model = UnivNetModel.from_pretrained("dg845/univnet-dev")
531
+ >>> feature_extractor = UnivNetFeatureExtractor.from_pretrained("dg845/univnet-dev")
532
+
533
+ >>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
534
+ >>> # Resample the audio to the feature extractor's sampling rate.
535
+ >>> ds = ds.cast_column("audio", Audio(sampling_rate=feature_extractor.sampling_rate))
536
+ >>> inputs = feature_extractor(
537
+ ... ds[0]["audio"]["array"], sampling_rate=ds[0]["audio"]["sampling_rate"], return_tensors="pt"
538
+ ... )
539
+ >>> audio = model(**inputs).waveforms
540
+ >>> list(audio.shape)
541
+ [1, 140288]
542
+ ```
543
+ """
544
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
545
+
546
+ # Resolve batch sizes for noise_sequence and spectrogram
547
+ spectrogram_batched = input_features.dim() == 3
548
+ if not spectrogram_batched:
549
+ input_features = input_features.unsqueeze(0)
550
+ spectrogram_batch_size, spectrogram_length, _ = input_features.shape
551
+
552
+ if noise_sequence is not None:
553
+ noise_sequence_batched = noise_sequence.dim() == 3
554
+ if not noise_sequence_batched:
555
+ noise_sequence = noise_sequence.unsqueeze(0)
556
+ else:
557
+ # Randomly generate noise_sequence
558
+ noise_sequence_shape = (spectrogram_batch_size, spectrogram_length, self.config.model_in_channels)
559
+ noise_sequence = torch.randn(
560
+ noise_sequence_shape, generator=generator, dtype=input_features.dtype, device=input_features.device
561
+ )
562
+ noise_sequence_batch_size = noise_sequence.shape[0]
563
+
564
+ if spectrogram_batch_size > 1 and noise_sequence_batch_size == 1:
565
+ # Repeat noise_sequence spectrogram_batch_size times
566
+ noise_sequence = noise_sequence.repeat(spectrogram_batch_size, 1, 1)
567
+ elif noise_sequence_batch_size > 1 and spectrogram_batch_size == 1:
568
+ # Repeat spectrogram noise_sequence_batch_size times
569
+ input_features = input_features.repeat(noise_sequence_batch_size, 1, 1)
570
+
571
+ if noise_sequence_batch_size != spectrogram_batch_size:
572
+ raise ValueError(
573
+ f"The batch size of `noise_sequence` is {noise_sequence_batch_size} and the batch size of"
574
+ f" `input_features` is {spectrogram_batch_size}, but the two are expected to be equal."
575
+ )
576
+
577
+ if padding_mask is not None:
578
+ if padding_mask.dim() == 1:
579
+ padding_mask = padding_mask.unsqueeze(0)
580
+ padding_mask_batch_size = padding_mask.shape[0]
581
+ if padding_mask_batch_size != spectrogram_batch_size:
582
+ raise ValueError(
583
+ f"The batch size of `padding_mask` is {padding_mask_batch_size} and the batch size of"
584
+ f" `input_features` is {spectrogram_batch_size}, but the two are expected to be equal."
585
+ )
586
+
587
+ # Change shapes to have channels before sequence lengths
588
+ hidden_states = noise_sequence.transpose(2, 1)
589
+ input_features = input_features.transpose(2, 1)
590
+
591
+ hidden_states = self.conv_pre(hidden_states)
592
+
593
+ for resblock in self.resblocks:
594
+ hidden_states = resblock(hidden_states, input_features)
595
+
596
+ hidden_states = nn.functional.leaky_relu(hidden_states, self.leaky_relu_slope)
597
+ hidden_states = self.conv_post(hidden_states)
598
+ hidden_states = torch.tanh(hidden_states)
599
+
600
+ # Remove sequence length dimension since this collapses to 1
601
+ # NOTE: keep waveforms batched even if there's only one
602
+ waveform = hidden_states.squeeze(1)
603
+
604
+ # Get sequence lengths for UnivNetFeatureExtractor.batch_decode.
605
+ waveform_lengths = None
606
+ if padding_mask is not None:
607
+ # Padding is always contiguous and added on the right
608
+ waveform_lengths = torch.sum(padding_mask, dim=1)
609
+
610
+ if not return_dict:
611
+ outputs = (waveform, waveform_lengths)
612
+ return outputs
613
+
614
+ return UnivNetModelOutput(
615
+ waveforms=waveform,
616
+ waveform_lengths=waveform_lengths,
617
+ )
618
+
619
+ def _init_weights(self, module):
620
+ """Initialize the weights."""
621
+ if isinstance(module, (nn.Linear, nn.Conv1d, nn.ConvTranspose1d)):
622
+ module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
623
+ if module.bias is not None:
624
+ module.bias.data.zero_()
625
+
626
+ def apply_weight_norm(self):
627
+ nn.utils.weight_norm(self.conv_pre)
628
+ for layer in self.resblocks:
629
+ layer.apply_weight_norm()
630
+ nn.utils.weight_norm(self.conv_post)
631
+
632
+ def remove_weight_norm(self):
633
+ nn.utils.remove_weight_norm(self.conv_pre)
634
+ for layer in self.resblocks:
635
+ layer.remove_weight_norm()
636
+ nn.utils.remove_weight_norm(self.conv_post)
evalkit_cambrian/lib/python3.10/site-packages/transformers/models/yoso/__pycache__/__init__.cpython-310.pyc ADDED
Binary file (1.05 kB). View file
 
evalkit_cambrian/lib/python3.10/site-packages/transformers/models/yoso/__pycache__/configuration_yoso.cpython-310.pyc ADDED
Binary file (5.93 kB). View file
 
evalkit_cambrian/lib/python3.10/site-packages/transformers/models/yoso/__pycache__/modeling_yoso.cpython-310.pyc ADDED
Binary file (35.8 kB). View file
 
evalkit_cambrian/lib/python3.10/site-packages/transformers/models/yoso/convert_yoso_pytorch_to_pytorch.py ADDED
@@ -0,0 +1,108 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright 2022 The HuggingFace Inc. team.
3
+ #
4
+ # Licensed under the Apache License, Version 2.0 (the "License");
5
+ # you may not use this file except in compliance with the License.
6
+ # You may obtain a copy of the License at
7
+ #
8
+ # http://www.apache.org/licenses/LICENSE-2.0
9
+ #
10
+ # Unless required by applicable law or agreed to in writing, software
11
+ # distributed under the License is distributed on an "AS IS" BASIS,
12
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
+ # See the License for the specific language governing permissions and
14
+ # limitations under the License.
15
+ """Convert YOSO checkpoints from the original repository. URL: https://github.com/mlpen/YOSO"""
16
+
17
+ import argparse
18
+
19
+ import torch
20
+
21
+ from transformers import YosoConfig, YosoForMaskedLM
22
+
23
+
24
+ def rename_key(orig_key):
25
+ if "model" in orig_key:
26
+ orig_key = orig_key.replace("model.", "")
27
+ if "norm1" in orig_key:
28
+ orig_key = orig_key.replace("norm1", "attention.output.LayerNorm")
29
+ if "norm2" in orig_key:
30
+ orig_key = orig_key.replace("norm2", "output.LayerNorm")
31
+ if "norm" in orig_key:
32
+ orig_key = orig_key.replace("norm", "LayerNorm")
33
+ if "transformer" in orig_key:
34
+ layer_num = orig_key.split(".")[0].split("_")[-1]
35
+ orig_key = orig_key.replace(f"transformer_{layer_num}", f"encoder.layer.{layer_num}")
36
+ if "mha.attn" in orig_key:
37
+ orig_key = orig_key.replace("mha.attn", "attention.self")
38
+ if "mha" in orig_key:
39
+ orig_key = orig_key.replace("mha", "attention")
40
+ if "W_q" in orig_key:
41
+ orig_key = orig_key.replace("W_q", "self.query")
42
+ if "W_k" in orig_key:
43
+ orig_key = orig_key.replace("W_k", "self.key")
44
+ if "W_v" in orig_key:
45
+ orig_key = orig_key.replace("W_v", "self.value")
46
+ if "ff1" in orig_key:
47
+ orig_key = orig_key.replace("ff1", "intermediate.dense")
48
+ if "ff2" in orig_key:
49
+ orig_key = orig_key.replace("ff2", "output.dense")
50
+ if "ff" in orig_key:
51
+ orig_key = orig_key.replace("ff", "output.dense")
52
+ if "mlm_class" in orig_key:
53
+ orig_key = orig_key.replace("mlm.mlm_class", "cls.predictions.decoder")
54
+ if "mlm" in orig_key:
55
+ orig_key = orig_key.replace("mlm", "cls.predictions.transform")
56
+ if "cls" not in orig_key:
57
+ orig_key = "yoso." + orig_key
58
+
59
+ return orig_key
60
+
61
+
62
+ def convert_checkpoint_helper(max_position_embeddings, orig_state_dict):
63
+ for key in orig_state_dict.copy().keys():
64
+ val = orig_state_dict.pop(key)
65
+
66
+ if ("pooler" in key) or ("sen_class" in key):
67
+ continue
68
+ else:
69
+ orig_state_dict[rename_key(key)] = val
70
+
71
+ orig_state_dict["cls.predictions.bias"] = orig_state_dict["cls.predictions.decoder.bias"]
72
+ orig_state_dict["yoso.embeddings.position_ids"] = torch.arange(max_position_embeddings).expand((1, -1)) + 2
73
+
74
+ return orig_state_dict
75
+
76
+
77
+ def convert_yoso_checkpoint(checkpoint_path, yoso_config_file, pytorch_dump_path):
78
+ orig_state_dict = torch.load(checkpoint_path, map_location="cpu")["model_state_dict"]
79
+ config = YosoConfig.from_json_file(yoso_config_file)
80
+ model = YosoForMaskedLM(config)
81
+
82
+ new_state_dict = convert_checkpoint_helper(config.max_position_embeddings, orig_state_dict)
83
+
84
+ print(model.load_state_dict(new_state_dict))
85
+ model.eval()
86
+ model.save_pretrained(pytorch_dump_path)
87
+
88
+ print(f"Checkpoint successfuly converted. Model saved at {pytorch_dump_path}")
89
+
90
+
91
+ if __name__ == "__main__":
92
+ parser = argparse.ArgumentParser()
93
+ # Required parameters
94
+ parser.add_argument(
95
+ "--pytorch_model_path", default=None, type=str, required=True, help="Path to YOSO pytorch checkpoint."
96
+ )
97
+ parser.add_argument(
98
+ "--config_file",
99
+ default=None,
100
+ type=str,
101
+ required=True,
102
+ help="The json file for YOSO model config.",
103
+ )
104
+ parser.add_argument(
105
+ "--pytorch_dump_path", default=None, type=str, required=True, help="Path to the output PyTorch model."
106
+ )
107
+ args = parser.parse_args()
108
+ convert_yoso_checkpoint(args.pytorch_model_path, args.config_file, args.pytorch_dump_path)
evalkit_cambrian/lib/python3.10/site-packages/transformers/models/yoso/modeling_yoso.py ADDED
@@ -0,0 +1,1304 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright 2022 University of Wisconsin-Madison and The HuggingFace Inc. team. All rights reserved.
3
+ #
4
+ # Licensed under the Apache License, Version 2.0 (the "License");
5
+ # you may not use this file except in compliance with the License.
6
+ # You may obtain a copy of the License at
7
+ #
8
+ # http://www.apache.org/licenses/LICENSE-2.0
9
+ #
10
+ # Unless required by applicable law or agreed to in writing, software
11
+ # distributed under the License is distributed on an "AS IS" BASIS,
12
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
+ # See the License for the specific language governing permissions and
14
+ # limitations under the License.
15
+ """ PyTorch YOSO model."""
16
+
17
+
18
+ import math
19
+ from pathlib import Path
20
+ from typing import Optional, Tuple, Union
21
+
22
+ import torch
23
+ import torch.utils.checkpoint
24
+ from torch import nn
25
+ from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
26
+
27
+ from ...activations import ACT2FN
28
+ from ...modeling_outputs import (
29
+ BaseModelOutputWithCrossAttentions,
30
+ MaskedLMOutput,
31
+ MultipleChoiceModelOutput,
32
+ QuestionAnsweringModelOutput,
33
+ SequenceClassifierOutput,
34
+ TokenClassifierOutput,
35
+ )
36
+ from ...modeling_utils import PreTrainedModel
37
+ from ...pytorch_utils import apply_chunking_to_forward, find_pruneable_heads_and_indices, prune_linear_layer
38
+ from ...utils import add_code_sample_docstrings, add_start_docstrings, add_start_docstrings_to_model_forward, logging
39
+ from .configuration_yoso import YosoConfig
40
+
41
+
42
+ logger = logging.get_logger(__name__)
43
+
44
+ _CHECKPOINT_FOR_DOC = "uw-madison/yoso-4096"
45
+ _CONFIG_FOR_DOC = "YosoConfig"
46
+
47
+ YOSO_PRETRAINED_MODEL_ARCHIVE_LIST = [
48
+ "uw-madison/yoso-4096",
49
+ # See all YOSO models at https://huggingface.co/models?filter=yoso
50
+ ]
51
+
52
+
53
+ def load_cuda_kernels():
54
+ global lsh_cumulation
55
+ try:
56
+ from torch.utils.cpp_extension import load
57
+
58
+ def append_root(files):
59
+ src_folder = Path(__file__).resolve().parent.parent.parent / "kernels" / "yoso"
60
+ return [src_folder / file for file in files]
61
+
62
+ src_files = append_root(
63
+ ["fast_lsh_cumulation_torch.cpp", "fast_lsh_cumulation.cu", "fast_lsh_cumulation_cuda.cu"]
64
+ )
65
+
66
+ load("fast_lsh_cumulation", src_files, verbose=True)
67
+
68
+ import fast_lsh_cumulation as lsh_cumulation
69
+
70
+ return True
71
+ except Exception:
72
+ lsh_cumulation = None
73
+ return False
74
+
75
+
76
+ def to_contiguous(input_tensors):
77
+ if isinstance(input_tensors, list):
78
+ out = []
79
+ for tensor in input_tensors:
80
+ if not tensor.is_contiguous():
81
+ tensor = tensor.contiguous()
82
+ out.append(tensor)
83
+ return out
84
+ else:
85
+ if not input_tensors.is_contiguous():
86
+ input_tensors = input_tensors.contiguous()
87
+ return input_tensors
88
+
89
+
90
+ def normalize(input_tensors):
91
+ if isinstance(input_tensors, list):
92
+ out = []
93
+ for tensor in input_tensors:
94
+ out.append(nn.functional.normalize(tensor, p=2, dim=-1))
95
+ return out
96
+ else:
97
+ return nn.functional.normalize(input_tensors, p=2, dim=-1)
98
+
99
+
100
+ def hashing(query, key, num_hash, hash_len):
101
+ if len(query.size()) != 3:
102
+ raise ValueError("Query has incorrect size.")
103
+ if len(key.size()) != 3:
104
+ raise ValueError("Key has incorrect size.")
105
+
106
+ rmat = torch.randn(query.size(0), query.size(2), num_hash * hash_len, device=query.device)
107
+ raise_pow = 2 ** torch.arange(hash_len, device=query.device)
108
+
109
+ query_projection = torch.matmul(query, rmat).reshape(query.size(0), query.size(1), num_hash, hash_len)
110
+ key_projection = torch.matmul(key, rmat).reshape(key.size(0), key.size(1), num_hash, hash_len)
111
+ query_binary = (query_projection > 0).int()
112
+ key_binary = (key_projection > 0).int()
113
+ query_hash = torch.sum(query_binary * raise_pow, dim=-1)
114
+ query_hash = torch.sum(key_binary * raise_pow, dim=-1)
115
+
116
+ return query_hash.int(), query_hash.int()
117
+
118
+
119
+ class YosoCumulation(torch.autograd.Function):
120
+ @staticmethod
121
+ def forward(ctx, query_mask, key_mask, query, key, value, config):
122
+ hash_code_len = config["hash_code_len"]
123
+
124
+ expectation = (1 - torch.acos(torch.matmul(query, key.transpose(-1, -2))) / math.pi) ** hash_code_len
125
+ expectation = expectation * query_mask[:, :, None] * key_mask[:, None, :]
126
+ cumulation_value = torch.matmul(expectation, value)
127
+
128
+ ctx.save_for_backward(query_mask, key_mask, expectation, query, key, value)
129
+ ctx.config = config
130
+
131
+ return cumulation_value
132
+
133
+ @staticmethod
134
+ def backward(ctx, grad):
135
+ grad = to_contiguous(grad)
136
+
137
+ query_mask, key_mask, expectation, query, key, value = ctx.saved_tensors
138
+ config = ctx.config
139
+
140
+ hash_code_len = config["hash_code_len"]
141
+
142
+ weighted_exp = torch.matmul(grad, value.transpose(-1, -2)) * expectation
143
+ grad_query = torch.matmul(weighted_exp, (hash_code_len / 2) * key)
144
+ grad_key = torch.matmul(weighted_exp.transpose(-1, -2), (hash_code_len / 2) * query)
145
+ grad_value = torch.matmul(expectation.transpose(-1, -2), grad)
146
+
147
+ return None, None, grad_query, grad_key, grad_value, None
148
+
149
+
150
+ class YosoLSHCumulation(torch.autograd.Function):
151
+ @staticmethod
152
+ def forward(ctx, query_mask, key_mask, query, key, value, config):
153
+ if query_mask.size(0) != key_mask.size(0):
154
+ raise ValueError("Query mask and Key mask differ in sizes in dimension 0")
155
+ if query_mask.size(0) != query.size(0):
156
+ raise ValueError("Query mask and Query differ in sizes in dimension 0")
157
+ if query_mask.size(0) != key.size(0):
158
+ raise ValueError("Query mask and Key differ in sizes in dimension 0")
159
+ if query_mask.size(0) != value.size(0):
160
+ raise ValueError("Query mask and Value mask differ in sizes in dimension 0")
161
+ if key.size(1) != value.size(1):
162
+ raise ValueError("Key and Value differ in sizes in dimension 1")
163
+ if query.size(2) != key.size(2):
164
+ raise ValueError("Query and Key differ in sizes in dimension 2")
165
+
166
+ query_mask, key_mask, query, key, value = to_contiguous([query_mask, key_mask, query, key, value])
167
+
168
+ use_cuda = query_mask.is_cuda
169
+ num_hash = config["num_hash"]
170
+ hash_code_len = config["hash_code_len"]
171
+ hashtable_capacity = int(2**hash_code_len)
172
+
173
+ if config["use_fast_hash"]:
174
+ query_hash_code, key_hash_code = lsh_cumulation.fast_hash(
175
+ query_mask, query, key_mask, key, num_hash, hash_code_len, use_cuda, 1
176
+ )
177
+ else:
178
+ query_hash_code, key_hash_code = hashing(query, key, num_hash, hash_code_len)
179
+
180
+ cumulation_value = lsh_cumulation.lsh_cumulation(
181
+ query_mask, query_hash_code, key_mask, key_hash_code, value, hashtable_capacity, use_cuda, 1
182
+ )
183
+
184
+ ctx.save_for_backward(query_mask, key_mask, query_hash_code, key_hash_code, query, key, value)
185
+ ctx.config = config
186
+
187
+ return cumulation_value
188
+
189
+ @staticmethod
190
+ def backward(ctx, grad):
191
+ grad = to_contiguous(grad)
192
+
193
+ query_mask, key_mask, query_hash_code, key_hash_code, query, key, value = ctx.saved_tensors
194
+ config = ctx.config
195
+
196
+ use_cuda = grad.is_cuda
197
+ hash_code_len = config["hash_code_len"]
198
+ hashtable_capacity = int(2**hash_code_len)
199
+
200
+ if config["lsh_backward"]:
201
+ grad_value = lsh_cumulation.lsh_cumulation(
202
+ key_mask, key_hash_code, query_mask, query_hash_code, grad, hashtable_capacity, use_cuda, 1
203
+ )
204
+ grad_query = lsh_cumulation.lsh_weighted_cumulation(
205
+ query_mask,
206
+ query_hash_code,
207
+ grad,
208
+ key_mask,
209
+ key_hash_code,
210
+ value,
211
+ (hash_code_len / 2) * key,
212
+ hashtable_capacity,
213
+ use_cuda,
214
+ 4,
215
+ )
216
+ grad_key = lsh_cumulation.lsh_weighted_cumulation(
217
+ key_mask,
218
+ key_hash_code,
219
+ value,
220
+ query_mask,
221
+ query_hash_code,
222
+ grad,
223
+ (hash_code_len / 2) * query,
224
+ hashtable_capacity,
225
+ use_cuda,
226
+ 4,
227
+ )
228
+ else:
229
+ expectation = (1 - torch.acos(torch.matmul(query, key.transpose(-1, -2))) / math.pi) ** hash_code_len
230
+ expectation = expectation * query_mask[:, :, None] * key_mask[:, None, :]
231
+ weighted_exp = torch.matmul(grad, value.transpose(-1, -2)) * expectation
232
+ grad_query = torch.matmul(weighted_exp, (hash_code_len / 2) * key)
233
+ grad_key = torch.matmul(weighted_exp.transpose(-1, -2), (hash_code_len / 2) * query)
234
+ grad_value = torch.matmul(expectation.transpose(-1, -2), grad)
235
+
236
+ return None, None, grad_query, grad_key, grad_value, None
237
+
238
+
239
+ # Copied from transformers.models.nystromformer.modeling_nystromformer.NystromformerEmbeddings
240
+ class YosoEmbeddings(nn.Module):
241
+ """Construct the embeddings from word, position and token_type embeddings."""
242
+
243
+ def __init__(self, config):
244
+ super().__init__()
245
+ self.word_embeddings = nn.Embedding(config.vocab_size, config.hidden_size, padding_idx=config.pad_token_id)
246
+ self.position_embeddings = nn.Embedding(config.max_position_embeddings + 2, config.hidden_size)
247
+ self.token_type_embeddings = nn.Embedding(config.type_vocab_size, config.hidden_size)
248
+
249
+ # self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load
250
+ # any TensorFlow checkpoint file
251
+ self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
252
+ self.dropout = nn.Dropout(config.hidden_dropout_prob)
253
+
254
+ # position_ids (1, len position emb) is contiguous in memory and exported when serialized
255
+ self.register_buffer(
256
+ "position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)) + 2, persistent=False
257
+ )
258
+ self.position_embedding_type = getattr(config, "position_embedding_type", "absolute")
259
+ self.register_buffer(
260
+ "token_type_ids",
261
+ torch.zeros(self.position_ids.size(), dtype=torch.long, device=self.position_ids.device),
262
+ persistent=False,
263
+ )
264
+
265
+ def forward(self, input_ids=None, token_type_ids=None, position_ids=None, inputs_embeds=None):
266
+ if input_ids is not None:
267
+ input_shape = input_ids.size()
268
+ else:
269
+ input_shape = inputs_embeds.size()[:-1]
270
+
271
+ seq_length = input_shape[1]
272
+
273
+ if position_ids is None:
274
+ position_ids = self.position_ids[:, :seq_length]
275
+
276
+ # Setting the token_type_ids to the registered buffer in constructor where it is all zeros, which usually occurs
277
+ # when its auto-generated, registered buffer helps users when tracing the model without passing token_type_ids, solves
278
+ # issue #5664
279
+ if token_type_ids is None:
280
+ if hasattr(self, "token_type_ids"):
281
+ buffered_token_type_ids = self.token_type_ids[:, :seq_length]
282
+ buffered_token_type_ids_expanded = buffered_token_type_ids.expand(input_shape[0], seq_length)
283
+ token_type_ids = buffered_token_type_ids_expanded
284
+ else:
285
+ token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=self.position_ids.device)
286
+
287
+ if inputs_embeds is None:
288
+ inputs_embeds = self.word_embeddings(input_ids)
289
+ token_type_embeddings = self.token_type_embeddings(token_type_ids)
290
+
291
+ embeddings = inputs_embeds + token_type_embeddings
292
+ if self.position_embedding_type == "absolute":
293
+ position_embeddings = self.position_embeddings(position_ids)
294
+ embeddings += position_embeddings
295
+ embeddings = self.LayerNorm(embeddings)
296
+ embeddings = self.dropout(embeddings)
297
+ return embeddings
298
+
299
+
300
+ class YosoSelfAttention(nn.Module):
301
+ def __init__(self, config, position_embedding_type=None):
302
+ super().__init__()
303
+ if config.hidden_size % config.num_attention_heads != 0 and not hasattr(config, "embedding_size"):
304
+ raise ValueError(
305
+ f"The hidden size ({config.hidden_size}) is not a multiple of the number of attention "
306
+ f"heads ({config.num_attention_heads})"
307
+ )
308
+
309
+ self.num_attention_heads = config.num_attention_heads
310
+ self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
311
+ self.all_head_size = self.num_attention_heads * self.attention_head_size
312
+
313
+ self.query = nn.Linear(config.hidden_size, self.all_head_size)
314
+ self.key = nn.Linear(config.hidden_size, self.all_head_size)
315
+ self.value = nn.Linear(config.hidden_size, self.all_head_size)
316
+
317
+ self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
318
+ self.position_embedding_type = (
319
+ position_embedding_type if position_embedding_type is not None else config.position_embedding_type
320
+ )
321
+
322
+ self.use_expectation = config.use_expectation
323
+ self.hash_code_len = config.hash_code_len
324
+ self.use_conv = config.conv_window is not None
325
+ self.use_fast_hash = config.use_fast_hash
326
+ self.num_hash = config.num_hash
327
+ self.lsh_backward = config.lsh_backward
328
+
329
+ self.lsh_config = {
330
+ "hash_code_len": self.hash_code_len,
331
+ "use_fast_hash": self.use_fast_hash,
332
+ "num_hash": self.num_hash,
333
+ "lsh_backward": self.lsh_backward,
334
+ }
335
+
336
+ if config.conv_window is not None:
337
+ self.conv = nn.Conv2d(
338
+ in_channels=config.num_attention_heads,
339
+ out_channels=config.num_attention_heads,
340
+ kernel_size=(config.conv_window, 1),
341
+ padding=(config.conv_window // 2, 0),
342
+ bias=False,
343
+ groups=config.num_attention_heads,
344
+ )
345
+
346
+ def transpose_for_scores(self, layer):
347
+ new_layer_shape = layer.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
348
+ layer = layer.view(*new_layer_shape)
349
+ return layer.permute(0, 2, 1, 3)
350
+
351
+ def forward(self, hidden_states, attention_mask=None, output_attentions=False):
352
+ mixed_query_layer = self.query(hidden_states)
353
+
354
+ key_layer = self.transpose_for_scores(self.key(hidden_states))
355
+ value_layer = self.transpose_for_scores(self.value(hidden_states))
356
+ query_layer = self.transpose_for_scores(mixed_query_layer)
357
+
358
+ if self.use_conv:
359
+ conv_value_layer = self.conv(value_layer * attention_mask[:, None, :, None])
360
+
361
+ batch_size, num_heads, seq_len, head_dim = query_layer.size()
362
+
363
+ query_layer = query_layer.reshape(batch_size * num_heads, seq_len, head_dim)
364
+ key_layer = key_layer.reshape(batch_size * num_heads, seq_len, head_dim)
365
+ value_layer = value_layer.reshape(batch_size * num_heads, seq_len, head_dim)
366
+
367
+ # revert changes made by get_extended_attention_mask
368
+ attention_mask = 1.0 + attention_mask / 10000.0
369
+ attention_mask = (
370
+ attention_mask.squeeze().repeat(1, num_heads, 1).reshape(batch_size * num_heads, seq_len).int()
371
+ )
372
+
373
+ # The CUDA kernels are most efficient with inputs whose size is a multiple of a GPU's warp size (32). Inputs
374
+ # smaller than this are padded with zeros.
375
+ gpu_warp_size = 32
376
+
377
+ if (not self.use_expectation) and head_dim < gpu_warp_size:
378
+ pad_size = batch_size * num_heads, seq_len, gpu_warp_size - head_dim
379
+
380
+ query_layer = torch.cat(
381
+ [
382
+ query_layer,
383
+ torch.zeros(pad_size, device=query_layer.device),
384
+ ],
385
+ dim=-1,
386
+ )
387
+ key_layer = torch.cat(
388
+ [
389
+ key_layer,
390
+ torch.zeros(pad_size, device=key_layer.device),
391
+ ],
392
+ dim=-1,
393
+ )
394
+ value_layer = torch.cat(
395
+ [
396
+ value_layer,
397
+ torch.zeros(pad_size, device=value_layer.device),
398
+ ],
399
+ dim=-1,
400
+ )
401
+
402
+ if self.use_expectation or self.training:
403
+ query_layer, key_layer = normalize([query_layer, key_layer])
404
+
405
+ if self.use_expectation:
406
+ context_layer = YosoCumulation.apply(
407
+ attention_mask, attention_mask, query_layer, key_layer, value_layer, self.lsh_config
408
+ )
409
+ else:
410
+ context_layer = YosoLSHCumulation.apply(
411
+ attention_mask, attention_mask, query_layer, key_layer, value_layer, self.lsh_config
412
+ )
413
+
414
+ if (not self.use_expectation) and head_dim < gpu_warp_size:
415
+ context_layer = context_layer[:, :, :head_dim]
416
+
417
+ context_layer = normalize(context_layer)
418
+
419
+ context_layer = context_layer.reshape(batch_size, num_heads, seq_len, head_dim)
420
+
421
+ if self.use_conv:
422
+ context_layer += conv_value_layer
423
+
424
+ context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
425
+ new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
426
+ context_layer = context_layer.view(*new_context_layer_shape)
427
+
428
+ outputs = (context_layer, context_layer) if output_attentions else (context_layer,)
429
+
430
+ return outputs
431
+
432
+
433
+ # Copied from transformers.models.bert.modeling_bert.BertSelfOutput
434
+ class YosoSelfOutput(nn.Module):
435
+ def __init__(self, config):
436
+ super().__init__()
437
+ self.dense = nn.Linear(config.hidden_size, config.hidden_size)
438
+ self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
439
+ self.dropout = nn.Dropout(config.hidden_dropout_prob)
440
+
441
+ def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor:
442
+ hidden_states = self.dense(hidden_states)
443
+ hidden_states = self.dropout(hidden_states)
444
+ hidden_states = self.LayerNorm(hidden_states + input_tensor)
445
+ return hidden_states
446
+
447
+
448
+ class YosoAttention(nn.Module):
449
+ def __init__(self, config, position_embedding_type=None):
450
+ super().__init__()
451
+ self.self = YosoSelfAttention(config, position_embedding_type=position_embedding_type)
452
+ self.output = YosoSelfOutput(config)
453
+ self.pruned_heads = set()
454
+
455
+ def prune_heads(self, heads):
456
+ if len(heads) == 0:
457
+ return
458
+ heads, index = find_pruneable_heads_and_indices(
459
+ heads, self.self.num_attention_heads, self.self.attention_head_size, self.pruned_heads
460
+ )
461
+
462
+ # Prune linear layers
463
+ self.self.query = prune_linear_layer(self.self.query, index)
464
+ self.self.key = prune_linear_layer(self.self.key, index)
465
+ self.self.value = prune_linear_layer(self.self.value, index)
466
+ self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)
467
+
468
+ # Update hyper params and store pruned heads
469
+ self.self.num_attention_heads = self.self.num_attention_heads - len(heads)
470
+ self.self.all_head_size = self.self.attention_head_size * self.self.num_attention_heads
471
+ self.pruned_heads = self.pruned_heads.union(heads)
472
+
473
+ def forward(self, hidden_states, attention_mask=None, output_attentions=False):
474
+ self_outputs = self.self(hidden_states, attention_mask, output_attentions)
475
+ attention_output = self.output(self_outputs[0], hidden_states)
476
+ outputs = (attention_output,) + self_outputs[1:] # add attentions if we output them
477
+ return outputs
478
+
479
+
480
+ # Copied from transformers.models.bert.modeling_bert.BertIntermediate
481
+ class YosoIntermediate(nn.Module):
482
+ def __init__(self, config):
483
+ super().__init__()
484
+ self.dense = nn.Linear(config.hidden_size, config.intermediate_size)
485
+ if isinstance(config.hidden_act, str):
486
+ self.intermediate_act_fn = ACT2FN[config.hidden_act]
487
+ else:
488
+ self.intermediate_act_fn = config.hidden_act
489
+
490
+ def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
491
+ hidden_states = self.dense(hidden_states)
492
+ hidden_states = self.intermediate_act_fn(hidden_states)
493
+ return hidden_states
494
+
495
+
496
+ # Copied from transformers.models.bert.modeling_bert.BertOutput
497
+ class YosoOutput(nn.Module):
498
+ def __init__(self, config):
499
+ super().__init__()
500
+ self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
501
+ self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
502
+ self.dropout = nn.Dropout(config.hidden_dropout_prob)
503
+
504
+ def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor:
505
+ hidden_states = self.dense(hidden_states)
506
+ hidden_states = self.dropout(hidden_states)
507
+ hidden_states = self.LayerNorm(hidden_states + input_tensor)
508
+ return hidden_states
509
+
510
+
511
+ class YosoLayer(nn.Module):
512
+ def __init__(self, config):
513
+ super().__init__()
514
+ self.chunk_size_feed_forward = config.chunk_size_feed_forward
515
+ self.seq_len_dim = 1
516
+ self.attention = YosoAttention(config)
517
+ self.add_cross_attention = config.add_cross_attention
518
+ self.intermediate = YosoIntermediate(config)
519
+ self.output = YosoOutput(config)
520
+
521
+ def forward(self, hidden_states, attention_mask=None, output_attentions=False):
522
+ self_attention_outputs = self.attention(hidden_states, attention_mask, output_attentions=output_attentions)
523
+ attention_output = self_attention_outputs[0]
524
+
525
+ outputs = self_attention_outputs[1:] # add self attentions if we output attention weights
526
+
527
+ layer_output = apply_chunking_to_forward(
528
+ self.feed_forward_chunk, self.chunk_size_feed_forward, self.seq_len_dim, attention_output
529
+ )
530
+ outputs = (layer_output,) + outputs
531
+
532
+ return outputs
533
+
534
+ def feed_forward_chunk(self, attention_output):
535
+ intermediate_output = self.intermediate(attention_output)
536
+ layer_output = self.output(intermediate_output, attention_output)
537
+ return layer_output
538
+
539
+
540
+ class YosoEncoder(nn.Module):
541
+ def __init__(self, config):
542
+ super().__init__()
543
+ self.config = config
544
+ self.layer = nn.ModuleList([YosoLayer(config) for _ in range(config.num_hidden_layers)])
545
+ self.gradient_checkpointing = False
546
+
547
+ def forward(
548
+ self,
549
+ hidden_states,
550
+ attention_mask=None,
551
+ head_mask=None,
552
+ output_attentions=False,
553
+ output_hidden_states=False,
554
+ return_dict=True,
555
+ ):
556
+ all_hidden_states = () if output_hidden_states else None
557
+ all_self_attentions = () if output_attentions else None
558
+
559
+ for i, layer_module in enumerate(self.layer):
560
+ if output_hidden_states:
561
+ all_hidden_states = all_hidden_states + (hidden_states,)
562
+
563
+ if self.gradient_checkpointing and self.training:
564
+ layer_outputs = self._gradient_checkpointing_func(
565
+ layer_module.__call__,
566
+ hidden_states,
567
+ attention_mask,
568
+ output_attentions,
569
+ )
570
+ else:
571
+ layer_outputs = layer_module(hidden_states, attention_mask, output_attentions)
572
+
573
+ hidden_states = layer_outputs[0]
574
+ if output_attentions:
575
+ all_self_attentions = all_self_attentions + (layer_outputs[1],)
576
+
577
+ if output_hidden_states:
578
+ all_hidden_states = all_hidden_states + (hidden_states,)
579
+
580
+ if not return_dict:
581
+ return tuple(v for v in [hidden_states, all_hidden_states, all_self_attentions] if v is not None)
582
+ return BaseModelOutputWithCrossAttentions(
583
+ last_hidden_state=hidden_states,
584
+ hidden_states=all_hidden_states,
585
+ attentions=all_self_attentions,
586
+ )
587
+
588
+
589
+ # Copied from transformers.models.bert.modeling_bert.BertPredictionHeadTransform
590
+ class YosoPredictionHeadTransform(nn.Module):
591
+ def __init__(self, config):
592
+ super().__init__()
593
+ self.dense = nn.Linear(config.hidden_size, config.hidden_size)
594
+ if isinstance(config.hidden_act, str):
595
+ self.transform_act_fn = ACT2FN[config.hidden_act]
596
+ else:
597
+ self.transform_act_fn = config.hidden_act
598
+ self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
599
+
600
+ def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
601
+ hidden_states = self.dense(hidden_states)
602
+ hidden_states = self.transform_act_fn(hidden_states)
603
+ hidden_states = self.LayerNorm(hidden_states)
604
+ return hidden_states
605
+
606
+
607
+ # Copied from transformers.models.bert.modeling_bert.BertLMPredictionHead with Bert->Yoso
608
+ class YosoLMPredictionHead(nn.Module):
609
+ def __init__(self, config):
610
+ super().__init__()
611
+ self.transform = YosoPredictionHeadTransform(config)
612
+
613
+ # The output weights are the same as the input embeddings, but there is
614
+ # an output-only bias for each token.
615
+ self.decoder = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
616
+
617
+ self.bias = nn.Parameter(torch.zeros(config.vocab_size))
618
+
619
+ # Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`
620
+ self.decoder.bias = self.bias
621
+
622
+ def forward(self, hidden_states):
623
+ hidden_states = self.transform(hidden_states)
624
+ hidden_states = self.decoder(hidden_states)
625
+ return hidden_states
626
+
627
+
628
+ # Copied from transformers.models.bert.modeling_bert.BertOnlyMLMHead with Bert->Yoso
629
+ class YosoOnlyMLMHead(nn.Module):
630
+ def __init__(self, config):
631
+ super().__init__()
632
+ self.predictions = YosoLMPredictionHead(config)
633
+
634
+ def forward(self, sequence_output: torch.Tensor) -> torch.Tensor:
635
+ prediction_scores = self.predictions(sequence_output)
636
+ return prediction_scores
637
+
638
+
639
+ class YosoPreTrainedModel(PreTrainedModel):
640
+ """
641
+ An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
642
+ models.
643
+ """
644
+
645
+ config_class = YosoConfig
646
+ base_model_prefix = "yoso"
647
+ supports_gradient_checkpointing = True
648
+
649
+ def _init_weights(self, module):
650
+ """Initialize the weights"""
651
+ if isinstance(module, nn.Linear):
652
+ # Slightly different from the TF version which uses truncated_normal for initialization
653
+ # cf https://github.com/pytorch/pytorch/pull/5617
654
+ module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
655
+ if module.bias is not None:
656
+ module.bias.data.zero_()
657
+ elif isinstance(module, nn.Embedding):
658
+ module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
659
+ if module.padding_idx is not None:
660
+ module.weight.data[module.padding_idx].zero_()
661
+ elif isinstance(module, nn.LayerNorm):
662
+ module.bias.data.zero_()
663
+ module.weight.data.fill_(1.0)
664
+
665
+
666
+ YOSO_START_DOCSTRING = r"""
667
+ This model is a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) sub-class. Use
668
+ it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and
669
+ behavior.
670
+
671
+ Parameters:
672
+ config ([`YosoConfig`]): Model configuration class with all the parameters of the model.
673
+ Initializing with a config file does not load the weights associated with the model, only the
674
+ configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
675
+ """
676
+
677
+ YOSO_INPUTS_DOCSTRING = r"""
678
+ Args:
679
+ input_ids (`torch.LongTensor` of shape `({0})`):
680
+ Indices of input sequence tokens in the vocabulary.
681
+
682
+ Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
683
+ [`PreTrainedTokenizer.__call__`] for details.
684
+
685
+ [What are input IDs?](../glossary#input-ids)
686
+ attention_mask (`torch.FloatTensor` of shape `({0})`, *optional*):
687
+ Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
688
+
689
+ - 1 for tokens that are **not masked**,
690
+ - 0 for tokens that are **masked**.
691
+
692
+ [What are attention masks?](../glossary#attention-mask)
693
+ token_type_ids (`torch.LongTensor` of shape `({0})`, *optional*):
694
+ Segment token indices to indicate first and second portions of the inputs. Indices are selected in `[0,
695
+ 1]`:
696
+
697
+ - 0 corresponds to a *sentence A* token,
698
+ - 1 corresponds to a *sentence B* token.
699
+
700
+ [What are token type IDs?](../glossary#token-type-ids)
701
+ position_ids (`torch.LongTensor` of shape `({0})`, *optional*):
702
+ Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
703
+ config.max_position_embeddings - 1]`.
704
+
705
+ [What are position IDs?](../glossary#position-ids)
706
+ head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
707
+ Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:
708
+
709
+ - 1 indicates the head is **not masked**,
710
+ - 0 indicates the head is **masked**.
711
+
712
+ inputs_embeds (`torch.FloatTensor` of shape `({0}, hidden_size)`, *optional*):
713
+ Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
714
+ is useful if you want more control over how to convert *input_ids* indices into associated vectors than the
715
+ model's internal embedding lookup matrix.
716
+ output_attentions (`bool`, *optional*):
717
+ Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
718
+ tensors for more detail.
719
+ output_hidden_states (`bool`, *optional*):
720
+ Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
721
+ more detail.
722
+ return_dict (`bool`, *optional*):
723
+ Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
724
+ """
725
+
726
+
727
+ @add_start_docstrings(
728
+ "The bare YOSO Model transformer outputting raw hidden-states without any specific head on top.",
729
+ YOSO_START_DOCSTRING,
730
+ )
731
+ class YosoModel(YosoPreTrainedModel):
732
+ def __init__(self, config):
733
+ super().__init__(config)
734
+ self.config = config
735
+
736
+ self.embeddings = YosoEmbeddings(config)
737
+ self.encoder = YosoEncoder(config)
738
+
739
+ # Initialize weights and apply final processing
740
+ self.post_init()
741
+
742
+ def get_input_embeddings(self):
743
+ return self.embeddings.word_embeddings
744
+
745
+ def set_input_embeddings(self, value):
746
+ self.embeddings.word_embeddings = value
747
+
748
+ def _prune_heads(self, heads_to_prune):
749
+ """
750
+ Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
751
+ class PreTrainedModel
752
+ """
753
+ for layer, heads in heads_to_prune.items():
754
+ self.encoder.layer[layer].attention.prune_heads(heads)
755
+
756
+ @add_start_docstrings_to_model_forward(YOSO_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
757
+ @add_code_sample_docstrings(
758
+ checkpoint=_CHECKPOINT_FOR_DOC,
759
+ output_type=BaseModelOutputWithCrossAttentions,
760
+ config_class=_CONFIG_FOR_DOC,
761
+ )
762
+ def forward(
763
+ self,
764
+ input_ids: Optional[torch.Tensor] = None,
765
+ attention_mask: Optional[torch.Tensor] = None,
766
+ token_type_ids: Optional[torch.Tensor] = None,
767
+ position_ids: Optional[torch.Tensor] = None,
768
+ head_mask: Optional[torch.Tensor] = None,
769
+ inputs_embeds: Optional[torch.Tensor] = None,
770
+ output_attentions: Optional[bool] = None,
771
+ output_hidden_states: Optional[bool] = None,
772
+ return_dict: Optional[bool] = None,
773
+ ) -> Union[Tuple, BaseModelOutputWithCrossAttentions]:
774
+ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
775
+ output_hidden_states = (
776
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
777
+ )
778
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
779
+
780
+ if input_ids is not None and inputs_embeds is not None:
781
+ raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
782
+ elif input_ids is not None:
783
+ self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask)
784
+ input_shape = input_ids.size()
785
+ elif inputs_embeds is not None:
786
+ input_shape = inputs_embeds.size()[:-1]
787
+ else:
788
+ raise ValueError("You have to specify either input_ids or inputs_embeds")
789
+
790
+ batch_size, seq_length = input_shape
791
+ device = input_ids.device if input_ids is not None else inputs_embeds.device
792
+
793
+ if attention_mask is None:
794
+ attention_mask = torch.ones(((batch_size, seq_length)), device=device)
795
+
796
+ if token_type_ids is None:
797
+ if hasattr(self.embeddings, "token_type_ids"):
798
+ buffered_token_type_ids = self.embeddings.token_type_ids[:, :seq_length]
799
+ buffered_token_type_ids_expanded = buffered_token_type_ids.expand(batch_size, seq_length)
800
+ token_type_ids = buffered_token_type_ids_expanded
801
+ else:
802
+ token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
803
+
804
+ # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
805
+ # ourselves in which case we just need to make it broadcastable to all heads.
806
+ extended_attention_mask: torch.Tensor = self.get_extended_attention_mask(attention_mask, input_shape)
807
+
808
+ # Prepare head mask if needed
809
+ # 1.0 in head_mask indicate we keep the head
810
+ # attention_probs has shape bsz x n_heads x N x N
811
+ # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
812
+ # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
813
+ head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
814
+
815
+ embedding_output = self.embeddings(
816
+ input_ids=input_ids,
817
+ position_ids=position_ids,
818
+ token_type_ids=token_type_ids,
819
+ inputs_embeds=inputs_embeds,
820
+ )
821
+ encoder_outputs = self.encoder(
822
+ embedding_output,
823
+ attention_mask=extended_attention_mask,
824
+ head_mask=head_mask,
825
+ output_attentions=output_attentions,
826
+ output_hidden_states=output_hidden_states,
827
+ return_dict=return_dict,
828
+ )
829
+ sequence_output = encoder_outputs[0]
830
+
831
+ if not return_dict:
832
+ return (sequence_output,) + encoder_outputs[1:]
833
+
834
+ return BaseModelOutputWithCrossAttentions(
835
+ last_hidden_state=sequence_output,
836
+ hidden_states=encoder_outputs.hidden_states,
837
+ attentions=encoder_outputs.attentions,
838
+ cross_attentions=encoder_outputs.cross_attentions,
839
+ )
840
+
841
+
842
+ @add_start_docstrings("""YOSO Model with a `language modeling` head on top.""", YOSO_START_DOCSTRING)
843
+ class YosoForMaskedLM(YosoPreTrainedModel):
844
+ _tied_weights_keys = ["cls.predictions.decoder.weight", "cls.predictions.decoder.bias"]
845
+
846
+ def __init__(self, config):
847
+ super().__init__(config)
848
+
849
+ self.yoso = YosoModel(config)
850
+ self.cls = YosoOnlyMLMHead(config)
851
+
852
+ # Initialize weights and apply final processing
853
+ self.post_init()
854
+
855
+ def get_output_embeddings(self):
856
+ return self.cls.predictions.decoder
857
+
858
+ def set_output_embeddings(self, new_embeddings):
859
+ self.cls.predictions.decoder = new_embeddings
860
+
861
+ @add_start_docstrings_to_model_forward(YOSO_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
862
+ @add_code_sample_docstrings(
863
+ checkpoint=_CHECKPOINT_FOR_DOC,
864
+ output_type=MaskedLMOutput,
865
+ config_class=_CONFIG_FOR_DOC,
866
+ )
867
+ def forward(
868
+ self,
869
+ input_ids: Optional[torch.Tensor] = None,
870
+ attention_mask: Optional[torch.Tensor] = None,
871
+ token_type_ids: Optional[torch.Tensor] = None,
872
+ position_ids: Optional[torch.Tensor] = None,
873
+ head_mask: Optional[torch.Tensor] = None,
874
+ inputs_embeds: Optional[torch.Tensor] = None,
875
+ labels: Optional[torch.Tensor] = None,
876
+ output_attentions: Optional[bool] = None,
877
+ output_hidden_states: Optional[bool] = None,
878
+ return_dict: Optional[bool] = None,
879
+ ) -> Union[Tuple, MaskedLMOutput]:
880
+ r"""
881
+ labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
882
+ Labels for computing the masked language modeling loss. Indices should be in `[-100, 0, ...,
883
+ config.vocab_size]` (see `input_ids` docstring) Tokens with indices set to `-100` are ignored (masked), the
884
+ loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
885
+ """
886
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
887
+
888
+ outputs = self.yoso(
889
+ input_ids,
890
+ attention_mask=attention_mask,
891
+ token_type_ids=token_type_ids,
892
+ position_ids=position_ids,
893
+ head_mask=head_mask,
894
+ inputs_embeds=inputs_embeds,
895
+ output_attentions=output_attentions,
896
+ output_hidden_states=output_hidden_states,
897
+ return_dict=return_dict,
898
+ )
899
+
900
+ sequence_output = outputs[0]
901
+ prediction_scores = self.cls(sequence_output)
902
+
903
+ masked_lm_loss = None
904
+ if labels is not None:
905
+ loss_fct = CrossEntropyLoss() # -100 index = padding token
906
+ masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), labels.view(-1))
907
+
908
+ if not return_dict:
909
+ output = (prediction_scores,) + outputs[1:]
910
+ return ((masked_lm_loss,) + output) if masked_lm_loss is not None else output
911
+
912
+ return MaskedLMOutput(
913
+ loss=masked_lm_loss,
914
+ logits=prediction_scores,
915
+ hidden_states=outputs.hidden_states,
916
+ attentions=outputs.attentions,
917
+ )
918
+
919
+
920
+ class YosoClassificationHead(nn.Module):
921
+ """Head for sentence-level classification tasks."""
922
+
923
+ def __init__(self, config):
924
+ super().__init__()
925
+ self.dense = nn.Linear(config.hidden_size, config.hidden_size)
926
+ self.dropout = nn.Dropout(config.hidden_dropout_prob)
927
+ self.out_proj = nn.Linear(config.hidden_size, config.num_labels)
928
+
929
+ self.config = config
930
+
931
+ def forward(self, features, **kwargs):
932
+ x = features[:, 0, :] # take <s> token (equiv. to [CLS])
933
+ x = self.dropout(x)
934
+ x = self.dense(x)
935
+ x = ACT2FN[self.config.hidden_act](x)
936
+ x = self.dropout(x)
937
+ x = self.out_proj(x)
938
+ return x
939
+
940
+
941
+ @add_start_docstrings(
942
+ """YOSO Model transformer with a sequence classification/regression head on top (a linear layer on top of
943
+ the pooled output) e.g. for GLUE tasks.""",
944
+ YOSO_START_DOCSTRING,
945
+ )
946
+ class YosoForSequenceClassification(YosoPreTrainedModel):
947
+ def __init__(self, config):
948
+ super().__init__(config)
949
+ self.num_labels = config.num_labels
950
+ self.yoso = YosoModel(config)
951
+ self.classifier = YosoClassificationHead(config)
952
+
953
+ # Initialize weights and apply final processing
954
+ self.post_init()
955
+
956
+ @add_start_docstrings_to_model_forward(YOSO_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
957
+ @add_code_sample_docstrings(
958
+ checkpoint=_CHECKPOINT_FOR_DOC,
959
+ output_type=SequenceClassifierOutput,
960
+ config_class=_CONFIG_FOR_DOC,
961
+ )
962
+ def forward(
963
+ self,
964
+ input_ids: Optional[torch.Tensor] = None,
965
+ attention_mask: Optional[torch.Tensor] = None,
966
+ token_type_ids: Optional[torch.Tensor] = None,
967
+ position_ids: Optional[torch.Tensor] = None,
968
+ head_mask: Optional[torch.Tensor] = None,
969
+ inputs_embeds: Optional[torch.Tensor] = None,
970
+ labels: Optional[torch.Tensor] = None,
971
+ output_attentions: Optional[bool] = None,
972
+ output_hidden_states: Optional[bool] = None,
973
+ return_dict: Optional[bool] = None,
974
+ ) -> Union[Tuple, SequenceClassifierOutput]:
975
+ r"""
976
+ labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
977
+ Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
978
+ config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
979
+ `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
980
+ """
981
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
982
+
983
+ outputs = self.yoso(
984
+ input_ids,
985
+ attention_mask=attention_mask,
986
+ token_type_ids=token_type_ids,
987
+ position_ids=position_ids,
988
+ head_mask=head_mask,
989
+ inputs_embeds=inputs_embeds,
990
+ output_attentions=output_attentions,
991
+ output_hidden_states=output_hidden_states,
992
+ return_dict=return_dict,
993
+ )
994
+
995
+ sequence_output = outputs[0]
996
+ logits = self.classifier(sequence_output)
997
+
998
+ loss = None
999
+ if labels is not None:
1000
+ if self.config.problem_type is None:
1001
+ if self.num_labels == 1:
1002
+ self.config.problem_type = "regression"
1003
+ elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
1004
+ self.config.problem_type = "single_label_classification"
1005
+ else:
1006
+ self.config.problem_type = "multi_label_classification"
1007
+
1008
+ if self.config.problem_type == "regression":
1009
+ loss_fct = MSELoss()
1010
+ if self.num_labels == 1:
1011
+ loss = loss_fct(logits.squeeze(), labels.squeeze())
1012
+ else:
1013
+ loss = loss_fct(logits, labels)
1014
+ elif self.config.problem_type == "single_label_classification":
1015
+ loss_fct = CrossEntropyLoss()
1016
+ loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
1017
+ elif self.config.problem_type == "multi_label_classification":
1018
+ loss_fct = BCEWithLogitsLoss()
1019
+ loss = loss_fct(logits, labels)
1020
+ if not return_dict:
1021
+ output = (logits,) + outputs[1:]
1022
+ return ((loss,) + output) if loss is not None else output
1023
+
1024
+ return SequenceClassifierOutput(
1025
+ loss=loss,
1026
+ logits=logits,
1027
+ hidden_states=outputs.hidden_states,
1028
+ attentions=outputs.attentions,
1029
+ )
1030
+
1031
+
1032
+ @add_start_docstrings(
1033
+ """YOSO Model with a multiple choice classification head on top (a linear layer on top of
1034
+ the pooled output and a softmax) e.g. for RocStories/SWAG tasks.""",
1035
+ YOSO_START_DOCSTRING,
1036
+ )
1037
+ class YosoForMultipleChoice(YosoPreTrainedModel):
1038
+ def __init__(self, config):
1039
+ super().__init__(config)
1040
+
1041
+ self.yoso = YosoModel(config)
1042
+ self.pre_classifier = nn.Linear(config.hidden_size, config.hidden_size)
1043
+ self.classifier = nn.Linear(config.hidden_size, 1)
1044
+
1045
+ # Initialize weights and apply final processing
1046
+ self.post_init()
1047
+
1048
+ @add_start_docstrings_to_model_forward(YOSO_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length"))
1049
+ @add_code_sample_docstrings(
1050
+ checkpoint=_CHECKPOINT_FOR_DOC,
1051
+ output_type=MultipleChoiceModelOutput,
1052
+ config_class=_CONFIG_FOR_DOC,
1053
+ )
1054
+ def forward(
1055
+ self,
1056
+ input_ids: Optional[torch.Tensor] = None,
1057
+ attention_mask: Optional[torch.Tensor] = None,
1058
+ token_type_ids: Optional[torch.Tensor] = None,
1059
+ position_ids: Optional[torch.Tensor] = None,
1060
+ head_mask: Optional[torch.Tensor] = None,
1061
+ inputs_embeds: Optional[torch.Tensor] = None,
1062
+ labels: Optional[torch.Tensor] = None,
1063
+ output_attentions: Optional[bool] = None,
1064
+ output_hidden_states: Optional[bool] = None,
1065
+ return_dict: Optional[bool] = None,
1066
+ ) -> Union[Tuple, MultipleChoiceModelOutput]:
1067
+ r"""
1068
+ labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
1069
+ Labels for computing the multiple choice classification loss. Indices should be in `[0, ...,
1070
+ num_choices-1]` where `num_choices` is the size of the second dimension of the input tensors. (See
1071
+ `input_ids` above)
1072
+ """
1073
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
1074
+ num_choices = input_ids.shape[1] if input_ids is not None else inputs_embeds.shape[1]
1075
+
1076
+ input_ids = input_ids.view(-1, input_ids.size(-1)) if input_ids is not None else None
1077
+ attention_mask = attention_mask.view(-1, attention_mask.size(-1)) if attention_mask is not None else None
1078
+ token_type_ids = token_type_ids.view(-1, token_type_ids.size(-1)) if token_type_ids is not None else None
1079
+ position_ids = position_ids.view(-1, position_ids.size(-1)) if position_ids is not None else None
1080
+ inputs_embeds = (
1081
+ inputs_embeds.view(-1, inputs_embeds.size(-2), inputs_embeds.size(-1))
1082
+ if inputs_embeds is not None
1083
+ else None
1084
+ )
1085
+
1086
+ outputs = self.yoso(
1087
+ input_ids,
1088
+ attention_mask=attention_mask,
1089
+ token_type_ids=token_type_ids,
1090
+ position_ids=position_ids,
1091
+ head_mask=head_mask,
1092
+ inputs_embeds=inputs_embeds,
1093
+ output_attentions=output_attentions,
1094
+ output_hidden_states=output_hidden_states,
1095
+ return_dict=return_dict,
1096
+ )
1097
+
1098
+ hidden_state = outputs[0] # (bs * num_choices, seq_len, dim)
1099
+ pooled_output = hidden_state[:, 0] # (bs * num_choices, dim)
1100
+ pooled_output = self.pre_classifier(pooled_output) # (bs * num_choices, dim)
1101
+ pooled_output = nn.ReLU()(pooled_output) # (bs * num_choices, dim)
1102
+ logits = self.classifier(pooled_output)
1103
+
1104
+ reshaped_logits = logits.view(-1, num_choices)
1105
+
1106
+ loss = None
1107
+ if labels is not None:
1108
+ loss_fct = CrossEntropyLoss()
1109
+ loss = loss_fct(reshaped_logits, labels)
1110
+
1111
+ if not return_dict:
1112
+ output = (reshaped_logits,) + outputs[1:]
1113
+ return ((loss,) + output) if loss is not None else output
1114
+
1115
+ return MultipleChoiceModelOutput(
1116
+ loss=loss,
1117
+ logits=reshaped_logits,
1118
+ hidden_states=outputs.hidden_states,
1119
+ attentions=outputs.attentions,
1120
+ )
1121
+
1122
+
1123
+ @add_start_docstrings(
1124
+ """YOSO Model with a token classification head on top (a linear layer on top of
1125
+ the hidden-states output) e.g. for Named-Entity-Recognition (NER) tasks.""",
1126
+ YOSO_START_DOCSTRING,
1127
+ )
1128
+ class YosoForTokenClassification(YosoPreTrainedModel):
1129
+ def __init__(self, config):
1130
+ super().__init__(config)
1131
+ self.num_labels = config.num_labels
1132
+
1133
+ self.yoso = YosoModel(config)
1134
+ self.dropout = nn.Dropout(config.hidden_dropout_prob)
1135
+ self.classifier = nn.Linear(config.hidden_size, config.num_labels)
1136
+
1137
+ # Initialize weights and apply final processing
1138
+ self.post_init()
1139
+
1140
+ @add_start_docstrings_to_model_forward(YOSO_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
1141
+ @add_code_sample_docstrings(
1142
+ checkpoint=_CHECKPOINT_FOR_DOC,
1143
+ output_type=TokenClassifierOutput,
1144
+ config_class=_CONFIG_FOR_DOC,
1145
+ )
1146
+ def forward(
1147
+ self,
1148
+ input_ids: Optional[torch.Tensor] = None,
1149
+ attention_mask: Optional[torch.Tensor] = None,
1150
+ token_type_ids: Optional[torch.Tensor] = None,
1151
+ position_ids: Optional[torch.Tensor] = None,
1152
+ head_mask: Optional[torch.Tensor] = None,
1153
+ inputs_embeds: Optional[torch.Tensor] = None,
1154
+ labels: Optional[torch.Tensor] = None,
1155
+ output_attentions: Optional[bool] = None,
1156
+ output_hidden_states: Optional[bool] = None,
1157
+ return_dict: Optional[bool] = None,
1158
+ ) -> Union[Tuple, TokenClassifierOutput]:
1159
+ r"""
1160
+ labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
1161
+ Labels for computing the token classification loss. Indices should be in `[0, ..., config.num_labels - 1]`.
1162
+ """
1163
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
1164
+
1165
+ outputs = self.yoso(
1166
+ input_ids,
1167
+ attention_mask=attention_mask,
1168
+ token_type_ids=token_type_ids,
1169
+ position_ids=position_ids,
1170
+ head_mask=head_mask,
1171
+ inputs_embeds=inputs_embeds,
1172
+ output_attentions=output_attentions,
1173
+ output_hidden_states=output_hidden_states,
1174
+ return_dict=return_dict,
1175
+ )
1176
+
1177
+ sequence_output = outputs[0]
1178
+
1179
+ sequence_output = self.dropout(sequence_output)
1180
+ logits = self.classifier(sequence_output)
1181
+
1182
+ loss = None
1183
+ if labels is not None:
1184
+ loss_fct = CrossEntropyLoss()
1185
+ # Only keep active parts of the loss
1186
+ if attention_mask is not None:
1187
+ active_loss = attention_mask.view(-1) == 1
1188
+ active_logits = logits.view(-1, self.num_labels)
1189
+ active_labels = torch.where(
1190
+ active_loss, labels.view(-1), torch.tensor(loss_fct.ignore_index).type_as(labels)
1191
+ )
1192
+ loss = loss_fct(active_logits, active_labels)
1193
+ else:
1194
+ loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
1195
+
1196
+ if not return_dict:
1197
+ output = (logits,) + outputs[1:]
1198
+ return ((loss,) + output) if loss is not None else output
1199
+
1200
+ return TokenClassifierOutput(
1201
+ loss=loss,
1202
+ logits=logits,
1203
+ hidden_states=outputs.hidden_states,
1204
+ attentions=outputs.attentions,
1205
+ )
1206
+
1207
+
1208
+ @add_start_docstrings(
1209
+ """YOSO Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear
1210
+ layers on top of the hidden-states output to compute `span start logits` and `span end logits`).""",
1211
+ YOSO_START_DOCSTRING,
1212
+ )
1213
+ class YosoForQuestionAnswering(YosoPreTrainedModel):
1214
+ def __init__(self, config):
1215
+ super().__init__(config)
1216
+
1217
+ config.num_labels = 2
1218
+ self.num_labels = config.num_labels
1219
+
1220
+ self.yoso = YosoModel(config)
1221
+ self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)
1222
+
1223
+ # Initialize weights and apply final processing
1224
+ self.post_init()
1225
+
1226
+ @add_start_docstrings_to_model_forward(YOSO_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
1227
+ @add_code_sample_docstrings(
1228
+ checkpoint=_CHECKPOINT_FOR_DOC,
1229
+ output_type=QuestionAnsweringModelOutput,
1230
+ config_class=_CONFIG_FOR_DOC,
1231
+ )
1232
+ def forward(
1233
+ self,
1234
+ input_ids: Optional[torch.Tensor] = None,
1235
+ attention_mask: Optional[torch.Tensor] = None,
1236
+ token_type_ids: Optional[torch.Tensor] = None,
1237
+ position_ids: Optional[torch.Tensor] = None,
1238
+ head_mask: Optional[torch.Tensor] = None,
1239
+ inputs_embeds: Optional[torch.Tensor] = None,
1240
+ start_positions: Optional[torch.Tensor] = None,
1241
+ end_positions: Optional[torch.Tensor] = None,
1242
+ output_attentions: Optional[bool] = None,
1243
+ output_hidden_states: Optional[bool] = None,
1244
+ return_dict: Optional[bool] = None,
1245
+ ) -> Union[Tuple, QuestionAnsweringModelOutput]:
1246
+ r"""
1247
+ start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
1248
+ Labels for position (index) of the start of the labelled span for computing the token classification loss.
1249
+ Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
1250
+ are not taken into account for computing the loss.
1251
+ end_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
1252
+ Labels for position (index) of the end of the labelled span for computing the token classification loss.
1253
+ Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
1254
+ are not taken into account for computing the loss.
1255
+ """
1256
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
1257
+
1258
+ outputs = self.yoso(
1259
+ input_ids,
1260
+ attention_mask=attention_mask,
1261
+ token_type_ids=token_type_ids,
1262
+ position_ids=position_ids,
1263
+ head_mask=head_mask,
1264
+ inputs_embeds=inputs_embeds,
1265
+ output_attentions=output_attentions,
1266
+ output_hidden_states=output_hidden_states,
1267
+ return_dict=return_dict,
1268
+ )
1269
+
1270
+ sequence_output = outputs[0]
1271
+
1272
+ logits = self.qa_outputs(sequence_output)
1273
+ start_logits, end_logits = logits.split(1, dim=-1)
1274
+ start_logits = start_logits.squeeze(-1)
1275
+ end_logits = end_logits.squeeze(-1)
1276
+
1277
+ total_loss = None
1278
+ if start_positions is not None and end_positions is not None:
1279
+ # If we are on multi-GPU, split add a dimension
1280
+ if len(start_positions.size()) > 1:
1281
+ start_positions = start_positions.squeeze(-1)
1282
+ if len(end_positions.size()) > 1:
1283
+ end_positions = end_positions.squeeze(-1)
1284
+ # sometimes the start/end positions are outside our model inputs, we ignore these terms
1285
+ ignored_index = start_logits.size(1)
1286
+ start_positions = start_positions.clamp(0, ignored_index)
1287
+ end_positions = end_positions.clamp(0, ignored_index)
1288
+
1289
+ loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
1290
+ start_loss = loss_fct(start_logits, start_positions)
1291
+ end_loss = loss_fct(end_logits, end_positions)
1292
+ total_loss = (start_loss + end_loss) / 2
1293
+
1294
+ if not return_dict:
1295
+ output = (start_logits, end_logits) + outputs[1:]
1296
+ return ((total_loss,) + output) if total_loss is not None else output
1297
+
1298
+ return QuestionAnsweringModelOutput(
1299
+ loss=total_loss,
1300
+ start_logits=start_logits,
1301
+ end_logits=end_logits,
1302
+ hidden_states=outputs.hidden_states,
1303
+ attentions=outputs.attentions,
1304
+ )
infer_4_47_1/lib/python3.10/site-packages/fontTools/designspaceLib/__init__.py ADDED
The diff for this file is too large to render. See raw diff
 
infer_4_47_1/lib/python3.10/site-packages/fontTools/feaLib/__main__.py ADDED
@@ -0,0 +1,78 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from fontTools.ttLib import TTFont
2
+ from fontTools.feaLib.builder import addOpenTypeFeatures, Builder
3
+ from fontTools.feaLib.error import FeatureLibError
4
+ from fontTools import configLogger
5
+ from fontTools.misc.cliTools import makeOutputFileName
6
+ import sys
7
+ import argparse
8
+ import logging
9
+
10
+
11
+ log = logging.getLogger("fontTools.feaLib")
12
+
13
+
14
+ def main(args=None):
15
+ """Add features from a feature file (.fea) into an OTF font"""
16
+ parser = argparse.ArgumentParser(
17
+ description="Use fontTools to compile OpenType feature files (*.fea)."
18
+ )
19
+ parser.add_argument(
20
+ "input_fea", metavar="FEATURES", help="Path to the feature file"
21
+ )
22
+ parser.add_argument(
23
+ "input_font", metavar="INPUT_FONT", help="Path to the input font"
24
+ )
25
+ parser.add_argument(
26
+ "-o",
27
+ "--output",
28
+ dest="output_font",
29
+ metavar="OUTPUT_FONT",
30
+ help="Path to the output font.",
31
+ )
32
+ parser.add_argument(
33
+ "-t",
34
+ "--tables",
35
+ metavar="TABLE_TAG",
36
+ choices=Builder.supportedTables,
37
+ nargs="+",
38
+ help="Specify the table(s) to be built.",
39
+ )
40
+ parser.add_argument(
41
+ "-d",
42
+ "--debug",
43
+ action="store_true",
44
+ help="Add source-level debugging information to font.",
45
+ )
46
+ parser.add_argument(
47
+ "-v",
48
+ "--verbose",
49
+ help="Increase the logger verbosity. Multiple -v " "options are allowed.",
50
+ action="count",
51
+ default=0,
52
+ )
53
+ parser.add_argument(
54
+ "--traceback", help="show traceback for exceptions.", action="store_true"
55
+ )
56
+ options = parser.parse_args(args)
57
+
58
+ levels = ["WARNING", "INFO", "DEBUG"]
59
+ configLogger(level=levels[min(len(levels) - 1, options.verbose)])
60
+
61
+ output_font = options.output_font or makeOutputFileName(options.input_font)
62
+ log.info("Compiling features to '%s'" % (output_font))
63
+
64
+ font = TTFont(options.input_font)
65
+ try:
66
+ addOpenTypeFeatures(
67
+ font, options.input_fea, tables=options.tables, debug=options.debug
68
+ )
69
+ except FeatureLibError as e:
70
+ if options.traceback:
71
+ raise
72
+ log.error(e)
73
+ sys.exit(1)
74
+ font.save(output_font)
75
+
76
+
77
+ if __name__ == "__main__":
78
+ sys.exit(main())
infer_4_47_1/lib/python3.10/site-packages/fontTools/feaLib/__pycache__/__init__.cpython-310.pyc ADDED
Binary file (256 Bytes). View file
 
infer_4_47_1/lib/python3.10/site-packages/fontTools/feaLib/__pycache__/__main__.cpython-310.pyc ADDED
Binary file (2.17 kB). View file
 
infer_4_47_1/lib/python3.10/site-packages/fontTools/feaLib/__pycache__/ast.cpython-310.pyc ADDED
Binary file (75.9 kB). View file
 
infer_4_47_1/lib/python3.10/site-packages/fontTools/feaLib/__pycache__/error.cpython-310.pyc ADDED
Binary file (1.11 kB). View file
 
infer_4_47_1/lib/python3.10/site-packages/fontTools/feaLib/__pycache__/lexer.cpython-310.pyc ADDED
Binary file (8.33 kB). View file
 
infer_4_47_1/lib/python3.10/site-packages/fontTools/feaLib/__pycache__/location.cpython-310.pyc ADDED
Binary file (671 Bytes). View file
 
infer_4_47_1/lib/python3.10/site-packages/fontTools/feaLib/__pycache__/lookupDebugInfo.cpython-310.pyc ADDED
Binary file (670 Bytes). View file
 
infer_4_47_1/lib/python3.10/site-packages/fontTools/feaLib/__pycache__/parser.cpython-310.pyc ADDED
Binary file (54.7 kB). View file
 
infer_4_47_1/lib/python3.10/site-packages/fontTools/feaLib/__pycache__/variableScalar.cpython-310.pyc ADDED
Binary file (5.36 kB). View file
 
infer_4_47_1/lib/python3.10/site-packages/fontTools/feaLib/error.py ADDED
@@ -0,0 +1,22 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ class FeatureLibError(Exception):
2
+ def __init__(self, message, location):
3
+ Exception.__init__(self, message)
4
+ self.location = location
5
+
6
+ def __str__(self):
7
+ message = Exception.__str__(self)
8
+ if self.location:
9
+ return f"{self.location}: {message}"
10
+ else:
11
+ return message
12
+
13
+
14
+ class IncludedFeaNotFound(FeatureLibError):
15
+ def __str__(self):
16
+ assert self.location is not None
17
+
18
+ message = (
19
+ "The following feature file should be included but cannot be found: "
20
+ f"{Exception.__str__(self)}"
21
+ )
22
+ return f"{self.location}: {message}"
infer_4_47_1/lib/python3.10/site-packages/fontTools/feaLib/lookupDebugInfo.py ADDED
@@ -0,0 +1,12 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from typing import NamedTuple
2
+
3
+ LOOKUP_DEBUG_INFO_KEY = "com.github.fonttools.feaLib"
4
+ LOOKUP_DEBUG_ENV_VAR = "FONTTOOLS_LOOKUP_DEBUGGING"
5
+
6
+
7
+ class LookupDebugInfo(NamedTuple):
8
+ """Information about where a lookup came from, to be embedded in a font"""
9
+
10
+ location: str
11
+ name: str
12
+ feature: list
infer_4_47_1/lib/python3.10/site-packages/fontTools/feaLib/parser.py ADDED
The diff for this file is too large to render. See raw diff
 
infer_4_47_1/lib/python3.10/site-packages/fontTools/feaLib/variableScalar.py ADDED
@@ -0,0 +1,113 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from fontTools.varLib.models import VariationModel, normalizeValue, piecewiseLinearMap
2
+
3
+
4
+ def Location(loc):
5
+ return tuple(sorted(loc.items()))
6
+
7
+
8
+ class VariableScalar:
9
+ """A scalar with different values at different points in the designspace."""
10
+
11
+ def __init__(self, location_value={}):
12
+ self.values = {}
13
+ self.axes = {}
14
+ for location, value in location_value.items():
15
+ self.add_value(location, value)
16
+
17
+ def __repr__(self):
18
+ items = []
19
+ for location, value in self.values.items():
20
+ loc = ",".join(["%s=%i" % (ax, loc) for ax, loc in location])
21
+ items.append("%s:%i" % (loc, value))
22
+ return "(" + (" ".join(items)) + ")"
23
+
24
+ @property
25
+ def does_vary(self):
26
+ values = list(self.values.values())
27
+ return any(v != values[0] for v in values[1:])
28
+
29
+ @property
30
+ def axes_dict(self):
31
+ if not self.axes:
32
+ raise ValueError(
33
+ ".axes must be defined on variable scalar before interpolating"
34
+ )
35
+ return {ax.axisTag: ax for ax in self.axes}
36
+
37
+ def _normalized_location(self, location):
38
+ location = self.fix_location(location)
39
+ normalized_location = {}
40
+ for axtag in location.keys():
41
+ if axtag not in self.axes_dict:
42
+ raise ValueError("Unknown axis %s in %s" % (axtag, location))
43
+ axis = self.axes_dict[axtag]
44
+ normalized_location[axtag] = normalizeValue(
45
+ location[axtag], (axis.minValue, axis.defaultValue, axis.maxValue)
46
+ )
47
+
48
+ return Location(normalized_location)
49
+
50
+ def fix_location(self, location):
51
+ location = dict(location)
52
+ for tag, axis in self.axes_dict.items():
53
+ if tag not in location:
54
+ location[tag] = axis.defaultValue
55
+ return location
56
+
57
+ def add_value(self, location, value):
58
+ if self.axes:
59
+ location = self.fix_location(location)
60
+
61
+ self.values[Location(location)] = value
62
+
63
+ def fix_all_locations(self):
64
+ self.values = {
65
+ Location(self.fix_location(l)): v for l, v in self.values.items()
66
+ }
67
+
68
+ @property
69
+ def default(self):
70
+ self.fix_all_locations()
71
+ key = Location({ax.axisTag: ax.defaultValue for ax in self.axes})
72
+ if key not in self.values:
73
+ raise ValueError("Default value could not be found")
74
+ # I *guess* we could interpolate one, but I don't know how.
75
+ return self.values[key]
76
+
77
+ def value_at_location(self, location, model_cache=None, avar=None):
78
+ loc = Location(location)
79
+ if loc in self.values.keys():
80
+ return self.values[loc]
81
+ values = list(self.values.values())
82
+ loc = dict(self._normalized_location(loc))
83
+ return self.model(model_cache, avar).interpolateFromMasters(loc, values)
84
+
85
+ def model(self, model_cache=None, avar=None):
86
+ if model_cache is not None:
87
+ key = tuple(self.values.keys())
88
+ if key in model_cache:
89
+ return model_cache[key]
90
+ locations = [dict(self._normalized_location(k)) for k in self.values.keys()]
91
+ if avar is not None:
92
+ mapping = avar.segments
93
+ locations = [
94
+ {
95
+ k: piecewiseLinearMap(v, mapping[k]) if k in mapping else v
96
+ for k, v in location.items()
97
+ }
98
+ for location in locations
99
+ ]
100
+ m = VariationModel(locations)
101
+ if model_cache is not None:
102
+ model_cache[key] = m
103
+ return m
104
+
105
+ def get_deltas_and_supports(self, model_cache=None, avar=None):
106
+ values = list(self.values.values())
107
+ return self.model(model_cache, avar).getDeltasAndSupports(values)
108
+
109
+ def add_to_variation_store(self, store_builder, model_cache=None, avar=None):
110
+ deltas, supports = self.get_deltas_and_supports(model_cache, avar)
111
+ store_builder.setSupports(supports)
112
+ index = store_builder.storeDeltas(deltas)
113
+ return int(self.default), index
infer_4_47_1/lib/python3.10/site-packages/fontTools/merge/__init__.py ADDED
@@ -0,0 +1,248 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2013 Google, Inc. All Rights Reserved.
2
+ #
3
+ # Google Author(s): Behdad Esfahbod, Roozbeh Pournader
4
+
5
+ from fontTools import ttLib
6
+ import fontTools.merge.base
7
+ from fontTools.merge.cmap import (
8
+ computeMegaGlyphOrder,
9
+ computeMegaCmap,
10
+ renameCFFCharStrings,
11
+ )
12
+ from fontTools.merge.layout import layoutPreMerge, layoutPostMerge
13
+ from fontTools.merge.options import Options
14
+ import fontTools.merge.tables
15
+ from fontTools.misc.loggingTools import Timer
16
+ from functools import reduce
17
+ import sys
18
+ import logging
19
+
20
+
21
+ log = logging.getLogger("fontTools.merge")
22
+ timer = Timer(logger=logging.getLogger(__name__ + ".timer"), level=logging.INFO)
23
+
24
+
25
+ class Merger(object):
26
+ """Font merger.
27
+
28
+ This class merges multiple files into a single OpenType font, taking into
29
+ account complexities such as OpenType layout (``GSUB``/``GPOS``) tables and
30
+ cross-font metrics (for example ``hhea.ascent`` is set to the maximum value
31
+ across all the fonts).
32
+
33
+ If multiple glyphs map to the same Unicode value, and the glyphs are considered
34
+ sufficiently different (that is, they differ in any of paths, widths, or
35
+ height), then subsequent glyphs are renamed and a lookup in the ``locl``
36
+ feature will be created to disambiguate them. For example, if the arguments
37
+ are an Arabic font and a Latin font and both contain a set of parentheses,
38
+ the Latin glyphs will be renamed to ``parenleft.1`` and ``parenright.1``,
39
+ and a lookup will be inserted into the to ``locl`` feature (creating it if
40
+ necessary) under the ``latn`` script to substitute ``parenleft`` with
41
+ ``parenleft.1`` etc.
42
+
43
+ Restrictions:
44
+
45
+ - All fonts must have the same units per em.
46
+ - If duplicate glyph disambiguation takes place as described above then the
47
+ fonts must have a ``GSUB`` table.
48
+
49
+ Attributes:
50
+ options: Currently unused.
51
+ """
52
+
53
+ def __init__(self, options=None):
54
+ if not options:
55
+ options = Options()
56
+
57
+ self.options = options
58
+
59
+ def _openFonts(self, fontfiles):
60
+ fonts = [ttLib.TTFont(fontfile) for fontfile in fontfiles]
61
+ for font, fontfile in zip(fonts, fontfiles):
62
+ font._merger__fontfile = fontfile
63
+ font._merger__name = font["name"].getDebugName(4)
64
+ return fonts
65
+
66
+ def merge(self, fontfiles):
67
+ """Merges fonts together.
68
+
69
+ Args:
70
+ fontfiles: A list of file names to be merged
71
+
72
+ Returns:
73
+ A :class:`fontTools.ttLib.TTFont` object. Call the ``save`` method on
74
+ this to write it out to an OTF file.
75
+ """
76
+ #
77
+ # Settle on a mega glyph order.
78
+ #
79
+ fonts = self._openFonts(fontfiles)
80
+ glyphOrders = [list(font.getGlyphOrder()) for font in fonts]
81
+ computeMegaGlyphOrder(self, glyphOrders)
82
+
83
+ # Take first input file sfntVersion
84
+ sfntVersion = fonts[0].sfntVersion
85
+
86
+ # Reload fonts and set new glyph names on them.
87
+ fonts = self._openFonts(fontfiles)
88
+ for font, glyphOrder in zip(fonts, glyphOrders):
89
+ font.setGlyphOrder(glyphOrder)
90
+ if "CFF " in font:
91
+ renameCFFCharStrings(self, glyphOrder, font["CFF "])
92
+
93
+ cmaps = [font["cmap"] for font in fonts]
94
+ self.duplicateGlyphsPerFont = [{} for _ in fonts]
95
+ computeMegaCmap(self, cmaps)
96
+
97
+ mega = ttLib.TTFont(sfntVersion=sfntVersion)
98
+ mega.setGlyphOrder(self.glyphOrder)
99
+
100
+ for font in fonts:
101
+ self._preMerge(font)
102
+
103
+ self.fonts = fonts
104
+
105
+ allTags = reduce(set.union, (list(font.keys()) for font in fonts), set())
106
+ allTags.remove("GlyphOrder")
107
+
108
+ for tag in sorted(allTags):
109
+ if tag in self.options.drop_tables:
110
+ continue
111
+
112
+ with timer("merge '%s'" % tag):
113
+ tables = [font.get(tag, NotImplemented) for font in fonts]
114
+
115
+ log.info("Merging '%s'.", tag)
116
+ clazz = ttLib.getTableClass(tag)
117
+ table = clazz(tag).merge(self, tables)
118
+ # XXX Clean this up and use: table = mergeObjects(tables)
119
+
120
+ if table is not NotImplemented and table is not False:
121
+ mega[tag] = table
122
+ log.info("Merged '%s'.", tag)
123
+ else:
124
+ log.info("Dropped '%s'.", tag)
125
+
126
+ del self.duplicateGlyphsPerFont
127
+ del self.fonts
128
+
129
+ self._postMerge(mega)
130
+
131
+ return mega
132
+
133
+ def mergeObjects(self, returnTable, logic, tables):
134
+ # Right now we don't use self at all. Will use in the future
135
+ # for options and logging.
136
+
137
+ allKeys = set.union(
138
+ set(),
139
+ *(vars(table).keys() for table in tables if table is not NotImplemented),
140
+ )
141
+ for key in allKeys:
142
+ log.info(" %s", key)
143
+ try:
144
+ mergeLogic = logic[key]
145
+ except KeyError:
146
+ try:
147
+ mergeLogic = logic["*"]
148
+ except KeyError:
149
+ raise Exception(
150
+ "Don't know how to merge key %s of class %s"
151
+ % (key, returnTable.__class__.__name__)
152
+ )
153
+ if mergeLogic is NotImplemented:
154
+ continue
155
+ value = mergeLogic(getattr(table, key, NotImplemented) for table in tables)
156
+ if value is not NotImplemented:
157
+ setattr(returnTable, key, value)
158
+
159
+ return returnTable
160
+
161
+ def _preMerge(self, font):
162
+ layoutPreMerge(font)
163
+
164
+ def _postMerge(self, font):
165
+ layoutPostMerge(font)
166
+
167
+ if "OS/2" in font:
168
+ # https://github.com/fonttools/fonttools/issues/2538
169
+ # TODO: Add an option to disable this?
170
+ font["OS/2"].recalcAvgCharWidth(font)
171
+
172
+
173
+ __all__ = ["Options", "Merger", "main"]
174
+
175
+
176
+ @timer("make one with everything (TOTAL TIME)")
177
+ def main(args=None):
178
+ """Merge multiple fonts into one"""
179
+ from fontTools import configLogger
180
+
181
+ if args is None:
182
+ args = sys.argv[1:]
183
+
184
+ options = Options()
185
+ args = options.parse_opts(args)
186
+ fontfiles = []
187
+ if options.input_file:
188
+ with open(options.input_file) as inputfile:
189
+ fontfiles = [
190
+ line.strip()
191
+ for line in inputfile.readlines()
192
+ if not line.lstrip().startswith("#")
193
+ ]
194
+ for g in args:
195
+ fontfiles.append(g)
196
+
197
+ if len(fontfiles) < 1:
198
+ print(
199
+ "usage: pyftmerge [font1 ... fontN] [--input-file=filelist.txt] [--output-file=merged.ttf] [--import-file=tables.ttx]",
200
+ file=sys.stderr,
201
+ )
202
+ print(
203
+ " [--drop-tables=tags] [--verbose] [--timing]",
204
+ file=sys.stderr,
205
+ )
206
+ print("", file=sys.stderr)
207
+ print(" font1 ... fontN Files to merge.", file=sys.stderr)
208
+ print(
209
+ " --input-file=<filename> Read files to merge from a text file, each path new line. # Comment lines allowed.",
210
+ file=sys.stderr,
211
+ )
212
+ print(
213
+ " --output-file=<filename> Specify output file name (default: merged.ttf).",
214
+ file=sys.stderr,
215
+ )
216
+ print(
217
+ " --import-file=<filename> TTX file to import after merging. This can be used to set metadata.",
218
+ file=sys.stderr,
219
+ )
220
+ print(
221
+ " --drop-tables=<table tags> Comma separated list of table tags to skip, case sensitive.",
222
+ file=sys.stderr,
223
+ )
224
+ print(
225
+ " --verbose Output progress information.",
226
+ file=sys.stderr,
227
+ )
228
+ print(" --timing Output progress timing.", file=sys.stderr)
229
+ return 1
230
+
231
+ configLogger(level=logging.INFO if options.verbose else logging.WARNING)
232
+ if options.timing:
233
+ timer.logger.setLevel(logging.DEBUG)
234
+ else:
235
+ timer.logger.disabled = True
236
+
237
+ merger = Merger(options=options)
238
+ font = merger.merge(fontfiles)
239
+
240
+ if options.import_file:
241
+ font.importXML(options.import_file)
242
+
243
+ with timer("compile and save font"):
244
+ font.save(options.output_file)
245
+
246
+
247
+ if __name__ == "__main__":
248
+ sys.exit(main())
infer_4_47_1/lib/python3.10/site-packages/fontTools/merge/__main__.py ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ import sys
2
+ from fontTools.merge import main
3
+
4
+
5
+ if __name__ == "__main__":
6
+ sys.exit(main())
infer_4_47_1/lib/python3.10/site-packages/fontTools/merge/__pycache__/__init__.cpython-310.pyc ADDED
Binary file (7.87 kB). View file
 
infer_4_47_1/lib/python3.10/site-packages/fontTools/merge/__pycache__/__main__.cpython-310.pyc ADDED
Binary file (288 Bytes). View file
 
infer_4_47_1/lib/python3.10/site-packages/fontTools/merge/__pycache__/base.cpython-310.pyc ADDED
Binary file (2.52 kB). View file
 
infer_4_47_1/lib/python3.10/site-packages/fontTools/merge/__pycache__/cmap.cpython-310.pyc ADDED
Binary file (3.34 kB). View file
 
infer_4_47_1/lib/python3.10/site-packages/fontTools/merge/__pycache__/layout.cpython-310.pyc ADDED
Binary file (11.7 kB). View file
 
infer_4_47_1/lib/python3.10/site-packages/fontTools/merge/__pycache__/options.cpython-310.pyc ADDED
Binary file (2.16 kB). View file
 
infer_4_47_1/lib/python3.10/site-packages/fontTools/merge/__pycache__/tables.cpython-310.pyc ADDED
Binary file (7.54 kB). View file
 
infer_4_47_1/lib/python3.10/site-packages/fontTools/merge/__pycache__/unicode.cpython-310.pyc ADDED
Binary file (1.12 kB). View file
 
infer_4_47_1/lib/python3.10/site-packages/fontTools/merge/__pycache__/util.cpython-310.pyc ADDED
Binary file (5.88 kB). View file
 
infer_4_47_1/lib/python3.10/site-packages/fontTools/merge/base.py ADDED
@@ -0,0 +1,81 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2013 Google, Inc. All Rights Reserved.
2
+ #
3
+ # Google Author(s): Behdad Esfahbod, Roozbeh Pournader
4
+
5
+ from fontTools.ttLib.tables.DefaultTable import DefaultTable
6
+ import logging
7
+
8
+
9
+ log = logging.getLogger("fontTools.merge")
10
+
11
+
12
+ def add_method(*clazzes, **kwargs):
13
+ """Returns a decorator function that adds a new method to one or
14
+ more classes."""
15
+ allowDefault = kwargs.get("allowDefaultTable", False)
16
+
17
+ def wrapper(method):
18
+ done = []
19
+ for clazz in clazzes:
20
+ if clazz in done:
21
+ continue # Support multiple names of a clazz
22
+ done.append(clazz)
23
+ assert allowDefault or clazz != DefaultTable, "Oops, table class not found."
24
+ assert (
25
+ method.__name__ not in clazz.__dict__
26
+ ), "Oops, class '%s' has method '%s'." % (clazz.__name__, method.__name__)
27
+ setattr(clazz, method.__name__, method)
28
+ return None
29
+
30
+ return wrapper
31
+
32
+
33
+ def mergeObjects(lst):
34
+ lst = [item for item in lst if item is not NotImplemented]
35
+ if not lst:
36
+ return NotImplemented
37
+ lst = [item for item in lst if item is not None]
38
+ if not lst:
39
+ return None
40
+
41
+ clazz = lst[0].__class__
42
+ assert all(type(item) == clazz for item in lst), lst
43
+
44
+ logic = clazz.mergeMap
45
+ returnTable = clazz()
46
+ returnDict = {}
47
+
48
+ allKeys = set.union(set(), *(vars(table).keys() for table in lst))
49
+ for key in allKeys:
50
+ try:
51
+ mergeLogic = logic[key]
52
+ except KeyError:
53
+ try:
54
+ mergeLogic = logic["*"]
55
+ except KeyError:
56
+ raise Exception(
57
+ "Don't know how to merge key %s of class %s" % (key, clazz.__name__)
58
+ )
59
+ if mergeLogic is NotImplemented:
60
+ continue
61
+ value = mergeLogic(getattr(table, key, NotImplemented) for table in lst)
62
+ if value is not NotImplemented:
63
+ returnDict[key] = value
64
+
65
+ returnTable.__dict__ = returnDict
66
+
67
+ return returnTable
68
+
69
+
70
+ @add_method(DefaultTable, allowDefaultTable=True)
71
+ def merge(self, m, tables):
72
+ if not hasattr(self, "mergeMap"):
73
+ log.info("Don't know how to merge '%s'.", self.tableTag)
74
+ return NotImplemented
75
+
76
+ logic = self.mergeMap
77
+
78
+ if isinstance(logic, dict):
79
+ return m.mergeObjects(self, self.mergeMap, tables)
80
+ else:
81
+ return logic(tables)
infer_4_47_1/lib/python3.10/site-packages/fontTools/merge/cmap.py ADDED
@@ -0,0 +1,141 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2013 Google, Inc. All Rights Reserved.
2
+ #
3
+ # Google Author(s): Behdad Esfahbod, Roozbeh Pournader
4
+
5
+ from fontTools.merge.unicode import is_Default_Ignorable
6
+ from fontTools.pens.recordingPen import DecomposingRecordingPen
7
+ import logging
8
+
9
+
10
+ log = logging.getLogger("fontTools.merge")
11
+
12
+
13
+ def computeMegaGlyphOrder(merger, glyphOrders):
14
+ """Modifies passed-in glyphOrders to reflect new glyph names.
15
+ Stores merger.glyphOrder."""
16
+ megaOrder = {}
17
+ for glyphOrder in glyphOrders:
18
+ for i, glyphName in enumerate(glyphOrder):
19
+ if glyphName in megaOrder:
20
+ n = megaOrder[glyphName]
21
+ while (glyphName + "." + repr(n)) in megaOrder:
22
+ n += 1
23
+ megaOrder[glyphName] = n
24
+ glyphName += "." + repr(n)
25
+ glyphOrder[i] = glyphName
26
+ megaOrder[glyphName] = 1
27
+ merger.glyphOrder = megaOrder = list(megaOrder.keys())
28
+
29
+
30
+ def _glyphsAreSame(
31
+ glyphSet1,
32
+ glyphSet2,
33
+ glyph1,
34
+ glyph2,
35
+ advanceTolerance=0.05,
36
+ advanceToleranceEmpty=0.20,
37
+ ):
38
+ pen1 = DecomposingRecordingPen(glyphSet1)
39
+ pen2 = DecomposingRecordingPen(glyphSet2)
40
+ g1 = glyphSet1[glyph1]
41
+ g2 = glyphSet2[glyph2]
42
+ g1.draw(pen1)
43
+ g2.draw(pen2)
44
+ if pen1.value != pen2.value:
45
+ return False
46
+ # Allow more width tolerance for glyphs with no ink
47
+ tolerance = advanceTolerance if pen1.value else advanceToleranceEmpty
48
+ # TODO Warn if advances not the same but within tolerance.
49
+ if abs(g1.width - g2.width) > g1.width * tolerance:
50
+ return False
51
+ if hasattr(g1, "height") and g1.height is not None:
52
+ if abs(g1.height - g2.height) > g1.height * tolerance:
53
+ return False
54
+ return True
55
+
56
+
57
+ # Valid (format, platformID, platEncID) triplets for cmap subtables containing
58
+ # Unicode BMP-only and Unicode Full Repertoire semantics.
59
+ # Cf. OpenType spec for "Platform specific encodings":
60
+ # https://docs.microsoft.com/en-us/typography/opentype/spec/name
61
+ class _CmapUnicodePlatEncodings:
62
+ BMP = {(4, 3, 1), (4, 0, 3), (4, 0, 4), (4, 0, 6)}
63
+ FullRepertoire = {(12, 3, 10), (12, 0, 4), (12, 0, 6)}
64
+
65
+
66
+ def computeMegaCmap(merger, cmapTables):
67
+ """Sets merger.cmap and merger.glyphOrder."""
68
+
69
+ # TODO Handle format=14.
70
+ # Only merge format 4 and 12 Unicode subtables, ignores all other subtables
71
+ # If there is a format 12 table for a font, ignore the format 4 table of it
72
+ chosenCmapTables = []
73
+ for fontIdx, table in enumerate(cmapTables):
74
+ format4 = None
75
+ format12 = None
76
+ for subtable in table.tables:
77
+ properties = (subtable.format, subtable.platformID, subtable.platEncID)
78
+ if properties in _CmapUnicodePlatEncodings.BMP:
79
+ format4 = subtable
80
+ elif properties in _CmapUnicodePlatEncodings.FullRepertoire:
81
+ format12 = subtable
82
+ else:
83
+ log.warning(
84
+ "Dropped cmap subtable from font '%s':\t"
85
+ "format %2s, platformID %2s, platEncID %2s",
86
+ fontIdx,
87
+ subtable.format,
88
+ subtable.platformID,
89
+ subtable.platEncID,
90
+ )
91
+ if format12 is not None:
92
+ chosenCmapTables.append((format12, fontIdx))
93
+ elif format4 is not None:
94
+ chosenCmapTables.append((format4, fontIdx))
95
+
96
+ # Build the unicode mapping
97
+ merger.cmap = cmap = {}
98
+ fontIndexForGlyph = {}
99
+ glyphSets = [None for f in merger.fonts] if hasattr(merger, "fonts") else None
100
+
101
+ for table, fontIdx in chosenCmapTables:
102
+ # handle duplicates
103
+ for uni, gid in table.cmap.items():
104
+ oldgid = cmap.get(uni, None)
105
+ if oldgid is None:
106
+ cmap[uni] = gid
107
+ fontIndexForGlyph[gid] = fontIdx
108
+ elif is_Default_Ignorable(uni) or uni in (0x25CC,): # U+25CC DOTTED CIRCLE
109
+ continue
110
+ elif oldgid != gid:
111
+ # Char previously mapped to oldgid, now to gid.
112
+ # Record, to fix up in GSUB 'locl' later.
113
+ if merger.duplicateGlyphsPerFont[fontIdx].get(oldgid) is None:
114
+ if glyphSets is not None:
115
+ oldFontIdx = fontIndexForGlyph[oldgid]
116
+ for idx in (fontIdx, oldFontIdx):
117
+ if glyphSets[idx] is None:
118
+ glyphSets[idx] = merger.fonts[idx].getGlyphSet()
119
+ # if _glyphsAreSame(glyphSets[oldFontIdx], glyphSets[fontIdx], oldgid, gid):
120
+ # continue
121
+ merger.duplicateGlyphsPerFont[fontIdx][oldgid] = gid
122
+ elif merger.duplicateGlyphsPerFont[fontIdx][oldgid] != gid:
123
+ # Char previously mapped to oldgid but oldgid is already remapped to a different
124
+ # gid, because of another Unicode character.
125
+ # TODO: Try harder to do something about these.
126
+ log.warning(
127
+ "Dropped mapping from codepoint %#06X to glyphId '%s'", uni, gid
128
+ )
129
+
130
+
131
+ def renameCFFCharStrings(merger, glyphOrder, cffTable):
132
+ """Rename topDictIndex charStrings based on glyphOrder."""
133
+ td = cffTable.cff.topDictIndex[0]
134
+
135
+ charStrings = {}
136
+ for i, v in enumerate(td.CharStrings.charStrings.values()):
137
+ glyphName = glyphOrder[i]
138
+ charStrings[glyphName] = v
139
+ td.CharStrings.charStrings = charStrings
140
+
141
+ td.charset = list(glyphOrder)
infer_4_47_1/lib/python3.10/site-packages/fontTools/merge/layout.py ADDED
@@ -0,0 +1,526 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2013 Google, Inc. All Rights Reserved.
2
+ #
3
+ # Google Author(s): Behdad Esfahbod, Roozbeh Pournader
4
+
5
+ from fontTools import ttLib
6
+ from fontTools.ttLib.tables.DefaultTable import DefaultTable
7
+ from fontTools.ttLib.tables import otTables
8
+ from fontTools.merge.base import add_method, mergeObjects
9
+ from fontTools.merge.util import *
10
+ import logging
11
+
12
+
13
+ log = logging.getLogger("fontTools.merge")
14
+
15
+
16
+ def mergeLookupLists(lst):
17
+ # TODO Do smarter merge.
18
+ return sumLists(lst)
19
+
20
+
21
+ def mergeFeatures(lst):
22
+ assert lst
23
+ self = otTables.Feature()
24
+ self.FeatureParams = None
25
+ self.LookupListIndex = mergeLookupLists(
26
+ [l.LookupListIndex for l in lst if l.LookupListIndex]
27
+ )
28
+ self.LookupCount = len(self.LookupListIndex)
29
+ return self
30
+
31
+
32
+ def mergeFeatureLists(lst):
33
+ d = {}
34
+ for l in lst:
35
+ for f in l:
36
+ tag = f.FeatureTag
37
+ if tag not in d:
38
+ d[tag] = []
39
+ d[tag].append(f.Feature)
40
+ ret = []
41
+ for tag in sorted(d.keys()):
42
+ rec = otTables.FeatureRecord()
43
+ rec.FeatureTag = tag
44
+ rec.Feature = mergeFeatures(d[tag])
45
+ ret.append(rec)
46
+ return ret
47
+
48
+
49
+ def mergeLangSyses(lst):
50
+ assert lst
51
+
52
+ # TODO Support merging ReqFeatureIndex
53
+ assert all(l.ReqFeatureIndex == 0xFFFF for l in lst)
54
+
55
+ self = otTables.LangSys()
56
+ self.LookupOrder = None
57
+ self.ReqFeatureIndex = 0xFFFF
58
+ self.FeatureIndex = mergeFeatureLists(
59
+ [l.FeatureIndex for l in lst if l.FeatureIndex]
60
+ )
61
+ self.FeatureCount = len(self.FeatureIndex)
62
+ return self
63
+
64
+
65
+ def mergeScripts(lst):
66
+ assert lst
67
+
68
+ if len(lst) == 1:
69
+ return lst[0]
70
+ langSyses = {}
71
+ for sr in lst:
72
+ for lsr in sr.LangSysRecord:
73
+ if lsr.LangSysTag not in langSyses:
74
+ langSyses[lsr.LangSysTag] = []
75
+ langSyses[lsr.LangSysTag].append(lsr.LangSys)
76
+ lsrecords = []
77
+ for tag, langSys_list in sorted(langSyses.items()):
78
+ lsr = otTables.LangSysRecord()
79
+ lsr.LangSys = mergeLangSyses(langSys_list)
80
+ lsr.LangSysTag = tag
81
+ lsrecords.append(lsr)
82
+
83
+ self = otTables.Script()
84
+ self.LangSysRecord = lsrecords
85
+ self.LangSysCount = len(lsrecords)
86
+ dfltLangSyses = [s.DefaultLangSys for s in lst if s.DefaultLangSys]
87
+ if dfltLangSyses:
88
+ self.DefaultLangSys = mergeLangSyses(dfltLangSyses)
89
+ else:
90
+ self.DefaultLangSys = None
91
+ return self
92
+
93
+
94
+ def mergeScriptRecords(lst):
95
+ d = {}
96
+ for l in lst:
97
+ for s in l:
98
+ tag = s.ScriptTag
99
+ if tag not in d:
100
+ d[tag] = []
101
+ d[tag].append(s.Script)
102
+ ret = []
103
+ for tag in sorted(d.keys()):
104
+ rec = otTables.ScriptRecord()
105
+ rec.ScriptTag = tag
106
+ rec.Script = mergeScripts(d[tag])
107
+ ret.append(rec)
108
+ return ret
109
+
110
+
111
+ otTables.ScriptList.mergeMap = {
112
+ "ScriptCount": lambda lst: None, # TODO
113
+ "ScriptRecord": mergeScriptRecords,
114
+ }
115
+ otTables.BaseScriptList.mergeMap = {
116
+ "BaseScriptCount": lambda lst: None, # TODO
117
+ # TODO: Merge duplicate entries
118
+ "BaseScriptRecord": lambda lst: sorted(
119
+ sumLists(lst), key=lambda s: s.BaseScriptTag
120
+ ),
121
+ }
122
+
123
+ otTables.FeatureList.mergeMap = {
124
+ "FeatureCount": sum,
125
+ "FeatureRecord": lambda lst: sorted(sumLists(lst), key=lambda s: s.FeatureTag),
126
+ }
127
+
128
+ otTables.LookupList.mergeMap = {
129
+ "LookupCount": sum,
130
+ "Lookup": sumLists,
131
+ }
132
+
133
+ otTables.Coverage.mergeMap = {
134
+ "Format": min,
135
+ "glyphs": sumLists,
136
+ }
137
+
138
+ otTables.ClassDef.mergeMap = {
139
+ "Format": min,
140
+ "classDefs": sumDicts,
141
+ }
142
+
143
+ otTables.LigCaretList.mergeMap = {
144
+ "Coverage": mergeObjects,
145
+ "LigGlyphCount": sum,
146
+ "LigGlyph": sumLists,
147
+ }
148
+
149
+ otTables.AttachList.mergeMap = {
150
+ "Coverage": mergeObjects,
151
+ "GlyphCount": sum,
152
+ "AttachPoint": sumLists,
153
+ }
154
+
155
+ # XXX Renumber MarkFilterSets of lookups
156
+ otTables.MarkGlyphSetsDef.mergeMap = {
157
+ "MarkSetTableFormat": equal,
158
+ "MarkSetCount": sum,
159
+ "Coverage": sumLists,
160
+ }
161
+
162
+ otTables.Axis.mergeMap = {
163
+ "*": mergeObjects,
164
+ }
165
+
166
+ # XXX Fix BASE table merging
167
+ otTables.BaseTagList.mergeMap = {
168
+ "BaseTagCount": sum,
169
+ "BaselineTag": sumLists,
170
+ }
171
+
172
+ otTables.GDEF.mergeMap = otTables.GSUB.mergeMap = otTables.GPOS.mergeMap = (
173
+ otTables.BASE.mergeMap
174
+ ) = otTables.JSTF.mergeMap = otTables.MATH.mergeMap = {
175
+ "*": mergeObjects,
176
+ "Version": max,
177
+ }
178
+
179
+ ttLib.getTableClass("GDEF").mergeMap = ttLib.getTableClass("GSUB").mergeMap = (
180
+ ttLib.getTableClass("GPOS").mergeMap
181
+ ) = ttLib.getTableClass("BASE").mergeMap = ttLib.getTableClass(
182
+ "JSTF"
183
+ ).mergeMap = ttLib.getTableClass(
184
+ "MATH"
185
+ ).mergeMap = {
186
+ "tableTag": onlyExisting(equal), # XXX clean me up
187
+ "table": mergeObjects,
188
+ }
189
+
190
+
191
+ @add_method(ttLib.getTableClass("GSUB"))
192
+ def merge(self, m, tables):
193
+ assert len(tables) == len(m.duplicateGlyphsPerFont)
194
+ for i, (table, dups) in enumerate(zip(tables, m.duplicateGlyphsPerFont)):
195
+ if not dups:
196
+ continue
197
+ if table is None or table is NotImplemented:
198
+ log.warning(
199
+ "Have non-identical duplicates to resolve for '%s' but no GSUB. Are duplicates intended?: %s",
200
+ m.fonts[i]._merger__name,
201
+ dups,
202
+ )
203
+ continue
204
+
205
+ synthFeature = None
206
+ synthLookup = None
207
+ for script in table.table.ScriptList.ScriptRecord:
208
+ if script.ScriptTag == "DFLT":
209
+ continue # XXX
210
+ for langsys in [script.Script.DefaultLangSys] + [
211
+ l.LangSys for l in script.Script.LangSysRecord
212
+ ]:
213
+ if langsys is None:
214
+ continue # XXX Create!
215
+ feature = [v for v in langsys.FeatureIndex if v.FeatureTag == "locl"]
216
+ assert len(feature) <= 1
217
+ if feature:
218
+ feature = feature[0]
219
+ else:
220
+ if not synthFeature:
221
+ synthFeature = otTables.FeatureRecord()
222
+ synthFeature.FeatureTag = "locl"
223
+ f = synthFeature.Feature = otTables.Feature()
224
+ f.FeatureParams = None
225
+ f.LookupCount = 0
226
+ f.LookupListIndex = []
227
+ table.table.FeatureList.FeatureRecord.append(synthFeature)
228
+ table.table.FeatureList.FeatureCount += 1
229
+ feature = synthFeature
230
+ langsys.FeatureIndex.append(feature)
231
+ langsys.FeatureIndex.sort(key=lambda v: v.FeatureTag)
232
+
233
+ if not synthLookup:
234
+ subtable = otTables.SingleSubst()
235
+ subtable.mapping = dups
236
+ synthLookup = otTables.Lookup()
237
+ synthLookup.LookupFlag = 0
238
+ synthLookup.LookupType = 1
239
+ synthLookup.SubTableCount = 1
240
+ synthLookup.SubTable = [subtable]
241
+ if table.table.LookupList is None:
242
+ # mtiLib uses None as default value for LookupList,
243
+ # while feaLib points to an empty array with count 0
244
+ # TODO: make them do the same
245
+ table.table.LookupList = otTables.LookupList()
246
+ table.table.LookupList.Lookup = []
247
+ table.table.LookupList.LookupCount = 0
248
+ table.table.LookupList.Lookup.append(synthLookup)
249
+ table.table.LookupList.LookupCount += 1
250
+
251
+ if feature.Feature.LookupListIndex[:1] != [synthLookup]:
252
+ feature.Feature.LookupListIndex[:0] = [synthLookup]
253
+ feature.Feature.LookupCount += 1
254
+
255
+ DefaultTable.merge(self, m, tables)
256
+ return self
257
+
258
+
259
+ @add_method(
260
+ otTables.SingleSubst,
261
+ otTables.MultipleSubst,
262
+ otTables.AlternateSubst,
263
+ otTables.LigatureSubst,
264
+ otTables.ReverseChainSingleSubst,
265
+ otTables.SinglePos,
266
+ otTables.PairPos,
267
+ otTables.CursivePos,
268
+ otTables.MarkBasePos,
269
+ otTables.MarkLigPos,
270
+ otTables.MarkMarkPos,
271
+ )
272
+ def mapLookups(self, lookupMap):
273
+ pass
274
+
275
+
276
+ # Copied and trimmed down from subset.py
277
+ @add_method(
278
+ otTables.ContextSubst,
279
+ otTables.ChainContextSubst,
280
+ otTables.ContextPos,
281
+ otTables.ChainContextPos,
282
+ )
283
+ def __merge_classify_context(self):
284
+ class ContextHelper(object):
285
+ def __init__(self, klass, Format):
286
+ if klass.__name__.endswith("Subst"):
287
+ Typ = "Sub"
288
+ Type = "Subst"
289
+ else:
290
+ Typ = "Pos"
291
+ Type = "Pos"
292
+ if klass.__name__.startswith("Chain"):
293
+ Chain = "Chain"
294
+ else:
295
+ Chain = ""
296
+ ChainTyp = Chain + Typ
297
+
298
+ self.Typ = Typ
299
+ self.Type = Type
300
+ self.Chain = Chain
301
+ self.ChainTyp = ChainTyp
302
+
303
+ self.LookupRecord = Type + "LookupRecord"
304
+
305
+ if Format == 1:
306
+ self.Rule = ChainTyp + "Rule"
307
+ self.RuleSet = ChainTyp + "RuleSet"
308
+ elif Format == 2:
309
+ self.Rule = ChainTyp + "ClassRule"
310
+ self.RuleSet = ChainTyp + "ClassSet"
311
+
312
+ if self.Format not in [1, 2, 3]:
313
+ return None # Don't shoot the messenger; let it go
314
+ if not hasattr(self.__class__, "_merge__ContextHelpers"):
315
+ self.__class__._merge__ContextHelpers = {}
316
+ if self.Format not in self.__class__._merge__ContextHelpers:
317
+ helper = ContextHelper(self.__class__, self.Format)
318
+ self.__class__._merge__ContextHelpers[self.Format] = helper
319
+ return self.__class__._merge__ContextHelpers[self.Format]
320
+
321
+
322
+ @add_method(
323
+ otTables.ContextSubst,
324
+ otTables.ChainContextSubst,
325
+ otTables.ContextPos,
326
+ otTables.ChainContextPos,
327
+ )
328
+ def mapLookups(self, lookupMap):
329
+ c = self.__merge_classify_context()
330
+
331
+ if self.Format in [1, 2]:
332
+ for rs in getattr(self, c.RuleSet):
333
+ if not rs:
334
+ continue
335
+ for r in getattr(rs, c.Rule):
336
+ if not r:
337
+ continue
338
+ for ll in getattr(r, c.LookupRecord):
339
+ if not ll:
340
+ continue
341
+ ll.LookupListIndex = lookupMap[ll.LookupListIndex]
342
+ elif self.Format == 3:
343
+ for ll in getattr(self, c.LookupRecord):
344
+ if not ll:
345
+ continue
346
+ ll.LookupListIndex = lookupMap[ll.LookupListIndex]
347
+ else:
348
+ assert 0, "unknown format: %s" % self.Format
349
+
350
+
351
+ @add_method(otTables.ExtensionSubst, otTables.ExtensionPos)
352
+ def mapLookups(self, lookupMap):
353
+ if self.Format == 1:
354
+ self.ExtSubTable.mapLookups(lookupMap)
355
+ else:
356
+ assert 0, "unknown format: %s" % self.Format
357
+
358
+
359
+ @add_method(otTables.Lookup)
360
+ def mapLookups(self, lookupMap):
361
+ for st in self.SubTable:
362
+ if not st:
363
+ continue
364
+ st.mapLookups(lookupMap)
365
+
366
+
367
+ @add_method(otTables.LookupList)
368
+ def mapLookups(self, lookupMap):
369
+ for l in self.Lookup:
370
+ if not l:
371
+ continue
372
+ l.mapLookups(lookupMap)
373
+
374
+
375
+ @add_method(otTables.Lookup)
376
+ def mapMarkFilteringSets(self, markFilteringSetMap):
377
+ if self.LookupFlag & 0x0010:
378
+ self.MarkFilteringSet = markFilteringSetMap[self.MarkFilteringSet]
379
+
380
+
381
+ @add_method(otTables.LookupList)
382
+ def mapMarkFilteringSets(self, markFilteringSetMap):
383
+ for l in self.Lookup:
384
+ if not l:
385
+ continue
386
+ l.mapMarkFilteringSets(markFilteringSetMap)
387
+
388
+
389
+ @add_method(otTables.Feature)
390
+ def mapLookups(self, lookupMap):
391
+ self.LookupListIndex = [lookupMap[i] for i in self.LookupListIndex]
392
+
393
+
394
+ @add_method(otTables.FeatureList)
395
+ def mapLookups(self, lookupMap):
396
+ for f in self.FeatureRecord:
397
+ if not f or not f.Feature:
398
+ continue
399
+ f.Feature.mapLookups(lookupMap)
400
+
401
+
402
+ @add_method(otTables.DefaultLangSys, otTables.LangSys)
403
+ def mapFeatures(self, featureMap):
404
+ self.FeatureIndex = [featureMap[i] for i in self.FeatureIndex]
405
+ if self.ReqFeatureIndex != 65535:
406
+ self.ReqFeatureIndex = featureMap[self.ReqFeatureIndex]
407
+
408
+
409
+ @add_method(otTables.Script)
410
+ def mapFeatures(self, featureMap):
411
+ if self.DefaultLangSys:
412
+ self.DefaultLangSys.mapFeatures(featureMap)
413
+ for l in self.LangSysRecord:
414
+ if not l or not l.LangSys:
415
+ continue
416
+ l.LangSys.mapFeatures(featureMap)
417
+
418
+
419
+ @add_method(otTables.ScriptList)
420
+ def mapFeatures(self, featureMap):
421
+ for s in self.ScriptRecord:
422
+ if not s or not s.Script:
423
+ continue
424
+ s.Script.mapFeatures(featureMap)
425
+
426
+
427
+ def layoutPreMerge(font):
428
+ # Map indices to references
429
+
430
+ GDEF = font.get("GDEF")
431
+ GSUB = font.get("GSUB")
432
+ GPOS = font.get("GPOS")
433
+
434
+ for t in [GSUB, GPOS]:
435
+ if not t:
436
+ continue
437
+
438
+ if t.table.LookupList:
439
+ lookupMap = {i: v for i, v in enumerate(t.table.LookupList.Lookup)}
440
+ t.table.LookupList.mapLookups(lookupMap)
441
+ t.table.FeatureList.mapLookups(lookupMap)
442
+
443
+ if (
444
+ GDEF
445
+ and GDEF.table.Version >= 0x00010002
446
+ and GDEF.table.MarkGlyphSetsDef
447
+ ):
448
+ markFilteringSetMap = {
449
+ i: v for i, v in enumerate(GDEF.table.MarkGlyphSetsDef.Coverage)
450
+ }
451
+ t.table.LookupList.mapMarkFilteringSets(markFilteringSetMap)
452
+
453
+ if t.table.FeatureList and t.table.ScriptList:
454
+ featureMap = {i: v for i, v in enumerate(t.table.FeatureList.FeatureRecord)}
455
+ t.table.ScriptList.mapFeatures(featureMap)
456
+
457
+ # TODO FeatureParams nameIDs
458
+
459
+
460
+ def layoutPostMerge(font):
461
+ # Map references back to indices
462
+
463
+ GDEF = font.get("GDEF")
464
+ GSUB = font.get("GSUB")
465
+ GPOS = font.get("GPOS")
466
+
467
+ for t in [GSUB, GPOS]:
468
+ if not t:
469
+ continue
470
+
471
+ if t.table.FeatureList and t.table.ScriptList:
472
+ # Collect unregistered (new) features.
473
+ featureMap = GregariousIdentityDict(t.table.FeatureList.FeatureRecord)
474
+ t.table.ScriptList.mapFeatures(featureMap)
475
+
476
+ # Record used features.
477
+ featureMap = AttendanceRecordingIdentityDict(
478
+ t.table.FeatureList.FeatureRecord
479
+ )
480
+ t.table.ScriptList.mapFeatures(featureMap)
481
+ usedIndices = featureMap.s
482
+
483
+ # Remove unused features
484
+ t.table.FeatureList.FeatureRecord = [
485
+ f
486
+ for i, f in enumerate(t.table.FeatureList.FeatureRecord)
487
+ if i in usedIndices
488
+ ]
489
+
490
+ # Map back to indices.
491
+ featureMap = NonhashableDict(t.table.FeatureList.FeatureRecord)
492
+ t.table.ScriptList.mapFeatures(featureMap)
493
+
494
+ t.table.FeatureList.FeatureCount = len(t.table.FeatureList.FeatureRecord)
495
+
496
+ if t.table.LookupList:
497
+ # Collect unregistered (new) lookups.
498
+ lookupMap = GregariousIdentityDict(t.table.LookupList.Lookup)
499
+ t.table.FeatureList.mapLookups(lookupMap)
500
+ t.table.LookupList.mapLookups(lookupMap)
501
+
502
+ # Record used lookups.
503
+ lookupMap = AttendanceRecordingIdentityDict(t.table.LookupList.Lookup)
504
+ t.table.FeatureList.mapLookups(lookupMap)
505
+ t.table.LookupList.mapLookups(lookupMap)
506
+ usedIndices = lookupMap.s
507
+
508
+ # Remove unused lookups
509
+ t.table.LookupList.Lookup = [
510
+ l for i, l in enumerate(t.table.LookupList.Lookup) if i in usedIndices
511
+ ]
512
+
513
+ # Map back to indices.
514
+ lookupMap = NonhashableDict(t.table.LookupList.Lookup)
515
+ t.table.FeatureList.mapLookups(lookupMap)
516
+ t.table.LookupList.mapLookups(lookupMap)
517
+
518
+ t.table.LookupList.LookupCount = len(t.table.LookupList.Lookup)
519
+
520
+ if GDEF and GDEF.table.Version >= 0x00010002:
521
+ markFilteringSetMap = NonhashableDict(
522
+ GDEF.table.MarkGlyphSetsDef.Coverage
523
+ )
524
+ t.table.LookupList.mapMarkFilteringSets(markFilteringSetMap)
525
+
526
+ # TODO FeatureParams nameIDs
infer_4_47_1/lib/python3.10/site-packages/fontTools/merge/options.py ADDED
@@ -0,0 +1,85 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2013 Google, Inc. All Rights Reserved.
2
+ #
3
+ # Google Author(s): Behdad Esfahbod, Roozbeh Pournader
4
+
5
+
6
+ class Options(object):
7
+ class UnknownOptionError(Exception):
8
+ pass
9
+
10
+ def __init__(self, **kwargs):
11
+ self.verbose = False
12
+ self.timing = False
13
+ self.drop_tables = []
14
+ self.input_file = None
15
+ self.output_file = "merged.ttf"
16
+ self.import_file = None
17
+
18
+ self.set(**kwargs)
19
+
20
+ def set(self, **kwargs):
21
+ for k, v in kwargs.items():
22
+ if not hasattr(self, k):
23
+ raise self.UnknownOptionError("Unknown option '%s'" % k)
24
+ setattr(self, k, v)
25
+
26
+ def parse_opts(self, argv, ignore_unknown=[]):
27
+ ret = []
28
+ opts = {}
29
+ for a in argv:
30
+ orig_a = a
31
+ if not a.startswith("--"):
32
+ ret.append(a)
33
+ continue
34
+ a = a[2:]
35
+ i = a.find("=")
36
+ op = "="
37
+ if i == -1:
38
+ if a.startswith("no-"):
39
+ k = a[3:]
40
+ v = False
41
+ else:
42
+ k = a
43
+ v = True
44
+ else:
45
+ k = a[:i]
46
+ if k[-1] in "-+":
47
+ op = k[-1] + "=" # Ops is '-=' or '+=' now.
48
+ k = k[:-1]
49
+ v = a[i + 1 :]
50
+ ok = k
51
+ k = k.replace("-", "_")
52
+ if not hasattr(self, k):
53
+ if ignore_unknown is True or ok in ignore_unknown:
54
+ ret.append(orig_a)
55
+ continue
56
+ else:
57
+ raise self.UnknownOptionError("Unknown option '%s'" % a)
58
+
59
+ ov = getattr(self, k)
60
+ if isinstance(ov, bool):
61
+ v = bool(v)
62
+ elif isinstance(ov, int):
63
+ v = int(v)
64
+ elif isinstance(ov, list):
65
+ vv = v.split(",")
66
+ if vv == [""]:
67
+ vv = []
68
+ vv = [int(x, 0) if len(x) and x[0] in "0123456789" else x for x in vv]
69
+ if op == "=":
70
+ v = vv
71
+ elif op == "+=":
72
+ v = ov
73
+ v.extend(vv)
74
+ elif op == "-=":
75
+ v = ov
76
+ for x in vv:
77
+ if x in v:
78
+ v.remove(x)
79
+ else:
80
+ assert 0
81
+
82
+ opts[k] = v
83
+ self.set(**opts)
84
+
85
+ return ret
infer_4_47_1/lib/python3.10/site-packages/fontTools/merge/tables.py ADDED
@@ -0,0 +1,341 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2013 Google, Inc. All Rights Reserved.
2
+ #
3
+ # Google Author(s): Behdad Esfahbod, Roozbeh Pournader
4
+
5
+ from fontTools import ttLib, cffLib
6
+ from fontTools.misc.psCharStrings import T2WidthExtractor
7
+ from fontTools.ttLib.tables.DefaultTable import DefaultTable
8
+ from fontTools.merge.base import add_method, mergeObjects
9
+ from fontTools.merge.cmap import computeMegaCmap
10
+ from fontTools.merge.util import *
11
+ import logging
12
+
13
+
14
+ log = logging.getLogger("fontTools.merge")
15
+
16
+
17
+ ttLib.getTableClass("maxp").mergeMap = {
18
+ "*": max,
19
+ "tableTag": equal,
20
+ "tableVersion": equal,
21
+ "numGlyphs": sum,
22
+ "maxStorage": first,
23
+ "maxFunctionDefs": first,
24
+ "maxInstructionDefs": first,
25
+ # TODO When we correctly merge hinting data, update these values:
26
+ # maxFunctionDefs, maxInstructionDefs, maxSizeOfInstructions
27
+ }
28
+
29
+ headFlagsMergeBitMap = {
30
+ "size": 16,
31
+ "*": bitwise_or,
32
+ 1: bitwise_and, # Baseline at y = 0
33
+ 2: bitwise_and, # lsb at x = 0
34
+ 3: bitwise_and, # Force ppem to integer values. FIXME?
35
+ 5: bitwise_and, # Font is vertical
36
+ 6: lambda bit: 0, # Always set to zero
37
+ 11: bitwise_and, # Font data is 'lossless'
38
+ 13: bitwise_and, # Optimized for ClearType
39
+ 14: bitwise_and, # Last resort font. FIXME? equal or first may be better
40
+ 15: lambda bit: 0, # Always set to zero
41
+ }
42
+
43
+ ttLib.getTableClass("head").mergeMap = {
44
+ "tableTag": equal,
45
+ "tableVersion": max,
46
+ "fontRevision": max,
47
+ "checkSumAdjustment": lambda lst: 0, # We need *something* here
48
+ "magicNumber": equal,
49
+ "flags": mergeBits(headFlagsMergeBitMap),
50
+ "unitsPerEm": equal,
51
+ "created": current_time,
52
+ "modified": current_time,
53
+ "xMin": min,
54
+ "yMin": min,
55
+ "xMax": max,
56
+ "yMax": max,
57
+ "macStyle": first,
58
+ "lowestRecPPEM": max,
59
+ "fontDirectionHint": lambda lst: 2,
60
+ "indexToLocFormat": first,
61
+ "glyphDataFormat": equal,
62
+ }
63
+
64
+ ttLib.getTableClass("hhea").mergeMap = {
65
+ "*": equal,
66
+ "tableTag": equal,
67
+ "tableVersion": max,
68
+ "ascent": max,
69
+ "descent": min,
70
+ "lineGap": max,
71
+ "advanceWidthMax": max,
72
+ "minLeftSideBearing": min,
73
+ "minRightSideBearing": min,
74
+ "xMaxExtent": max,
75
+ "caretSlopeRise": first,
76
+ "caretSlopeRun": first,
77
+ "caretOffset": first,
78
+ "numberOfHMetrics": recalculate,
79
+ }
80
+
81
+ ttLib.getTableClass("vhea").mergeMap = {
82
+ "*": equal,
83
+ "tableTag": equal,
84
+ "tableVersion": max,
85
+ "ascent": max,
86
+ "descent": min,
87
+ "lineGap": max,
88
+ "advanceHeightMax": max,
89
+ "minTopSideBearing": min,
90
+ "minBottomSideBearing": min,
91
+ "yMaxExtent": max,
92
+ "caretSlopeRise": first,
93
+ "caretSlopeRun": first,
94
+ "caretOffset": first,
95
+ "numberOfVMetrics": recalculate,
96
+ }
97
+
98
+ os2FsTypeMergeBitMap = {
99
+ "size": 16,
100
+ "*": lambda bit: 0,
101
+ 1: bitwise_or, # no embedding permitted
102
+ 2: bitwise_and, # allow previewing and printing documents
103
+ 3: bitwise_and, # allow editing documents
104
+ 8: bitwise_or, # no subsetting permitted
105
+ 9: bitwise_or, # no embedding of outlines permitted
106
+ }
107
+
108
+
109
+ def mergeOs2FsType(lst):
110
+ lst = list(lst)
111
+ if all(item == 0 for item in lst):
112
+ return 0
113
+
114
+ # Compute least restrictive logic for each fsType value
115
+ for i in range(len(lst)):
116
+ # unset bit 1 (no embedding permitted) if either bit 2 or 3 is set
117
+ if lst[i] & 0x000C:
118
+ lst[i] &= ~0x0002
119
+ # set bit 2 (allow previewing) if bit 3 is set (allow editing)
120
+ elif lst[i] & 0x0008:
121
+ lst[i] |= 0x0004
122
+ # set bits 2 and 3 if everything is allowed
123
+ elif lst[i] == 0:
124
+ lst[i] = 0x000C
125
+
126
+ fsType = mergeBits(os2FsTypeMergeBitMap)(lst)
127
+ # unset bits 2 and 3 if bit 1 is set (some font is "no embedding")
128
+ if fsType & 0x0002:
129
+ fsType &= ~0x000C
130
+ return fsType
131
+
132
+
133
+ ttLib.getTableClass("OS/2").mergeMap = {
134
+ "*": first,
135
+ "tableTag": equal,
136
+ "version": max,
137
+ "xAvgCharWidth": first, # Will be recalculated at the end on the merged font
138
+ "fsType": mergeOs2FsType, # Will be overwritten
139
+ "panose": first, # FIXME: should really be the first Latin font
140
+ "ulUnicodeRange1": bitwise_or,
141
+ "ulUnicodeRange2": bitwise_or,
142
+ "ulUnicodeRange3": bitwise_or,
143
+ "ulUnicodeRange4": bitwise_or,
144
+ "fsFirstCharIndex": min,
145
+ "fsLastCharIndex": max,
146
+ "sTypoAscender": max,
147
+ "sTypoDescender": min,
148
+ "sTypoLineGap": max,
149
+ "usWinAscent": max,
150
+ "usWinDescent": max,
151
+ # Version 1
152
+ "ulCodePageRange1": onlyExisting(bitwise_or),
153
+ "ulCodePageRange2": onlyExisting(bitwise_or),
154
+ # Version 2, 3, 4
155
+ "sxHeight": onlyExisting(max),
156
+ "sCapHeight": onlyExisting(max),
157
+ "usDefaultChar": onlyExisting(first),
158
+ "usBreakChar": onlyExisting(first),
159
+ "usMaxContext": onlyExisting(max),
160
+ # version 5
161
+ "usLowerOpticalPointSize": onlyExisting(min),
162
+ "usUpperOpticalPointSize": onlyExisting(max),
163
+ }
164
+
165
+
166
+ @add_method(ttLib.getTableClass("OS/2"))
167
+ def merge(self, m, tables):
168
+ DefaultTable.merge(self, m, tables)
169
+ if self.version < 2:
170
+ # bits 8 and 9 are reserved and should be set to zero
171
+ self.fsType &= ~0x0300
172
+ if self.version >= 3:
173
+ # Only one of bits 1, 2, and 3 may be set. We already take
174
+ # care of bit 1 implications in mergeOs2FsType. So unset
175
+ # bit 2 if bit 3 is already set.
176
+ if self.fsType & 0x0008:
177
+ self.fsType &= ~0x0004
178
+ return self
179
+
180
+
181
+ ttLib.getTableClass("post").mergeMap = {
182
+ "*": first,
183
+ "tableTag": equal,
184
+ "formatType": max,
185
+ "isFixedPitch": min,
186
+ "minMemType42": max,
187
+ "maxMemType42": lambda lst: 0,
188
+ "minMemType1": max,
189
+ "maxMemType1": lambda lst: 0,
190
+ "mapping": onlyExisting(sumDicts),
191
+ "extraNames": lambda lst: [],
192
+ }
193
+
194
+ ttLib.getTableClass("vmtx").mergeMap = ttLib.getTableClass("hmtx").mergeMap = {
195
+ "tableTag": equal,
196
+ "metrics": sumDicts,
197
+ }
198
+
199
+ ttLib.getTableClass("name").mergeMap = {
200
+ "tableTag": equal,
201
+ "names": first, # FIXME? Does mixing name records make sense?
202
+ }
203
+
204
+ ttLib.getTableClass("loca").mergeMap = {
205
+ "*": recalculate,
206
+ "tableTag": equal,
207
+ }
208
+
209
+ ttLib.getTableClass("glyf").mergeMap = {
210
+ "tableTag": equal,
211
+ "glyphs": sumDicts,
212
+ "glyphOrder": sumLists,
213
+ "_reverseGlyphOrder": recalculate,
214
+ "axisTags": equal,
215
+ }
216
+
217
+
218
+ @add_method(ttLib.getTableClass("glyf"))
219
+ def merge(self, m, tables):
220
+ for i, table in enumerate(tables):
221
+ for g in table.glyphs.values():
222
+ if i:
223
+ # Drop hints for all but first font, since
224
+ # we don't map functions / CVT values.
225
+ g.removeHinting()
226
+ # Expand composite glyphs to load their
227
+ # composite glyph names.
228
+ if g.isComposite():
229
+ g.expand(table)
230
+ return DefaultTable.merge(self, m, tables)
231
+
232
+
233
+ ttLib.getTableClass("prep").mergeMap = lambda self, lst: first(lst)
234
+ ttLib.getTableClass("fpgm").mergeMap = lambda self, lst: first(lst)
235
+ ttLib.getTableClass("cvt ").mergeMap = lambda self, lst: first(lst)
236
+ ttLib.getTableClass("gasp").mergeMap = lambda self, lst: first(
237
+ lst
238
+ ) # FIXME? Appears irreconcilable
239
+
240
+
241
+ @add_method(ttLib.getTableClass("CFF "))
242
+ def merge(self, m, tables):
243
+ if any(hasattr(table.cff[0], "FDSelect") for table in tables):
244
+ raise NotImplementedError("Merging CID-keyed CFF tables is not supported yet")
245
+
246
+ for table in tables:
247
+ table.cff.desubroutinize()
248
+
249
+ newcff = tables[0]
250
+ newfont = newcff.cff[0]
251
+ private = newfont.Private
252
+ newDefaultWidthX, newNominalWidthX = private.defaultWidthX, private.nominalWidthX
253
+ storedNamesStrings = []
254
+ glyphOrderStrings = []
255
+ glyphOrder = set(newfont.getGlyphOrder())
256
+
257
+ for name in newfont.strings.strings:
258
+ if name not in glyphOrder:
259
+ storedNamesStrings.append(name)
260
+ else:
261
+ glyphOrderStrings.append(name)
262
+
263
+ chrset = list(newfont.charset)
264
+ newcs = newfont.CharStrings
265
+ log.debug("FONT 0 CharStrings: %d.", len(newcs))
266
+
267
+ for i, table in enumerate(tables[1:], start=1):
268
+ font = table.cff[0]
269
+ defaultWidthX, nominalWidthX = (
270
+ font.Private.defaultWidthX,
271
+ font.Private.nominalWidthX,
272
+ )
273
+ widthsDiffer = (
274
+ defaultWidthX != newDefaultWidthX or nominalWidthX != newNominalWidthX
275
+ )
276
+ font.Private = private
277
+ fontGlyphOrder = set(font.getGlyphOrder())
278
+ for name in font.strings.strings:
279
+ if name in fontGlyphOrder:
280
+ glyphOrderStrings.append(name)
281
+ cs = font.CharStrings
282
+ gs = table.cff.GlobalSubrs
283
+ log.debug("Font %d CharStrings: %d.", i, len(cs))
284
+ chrset.extend(font.charset)
285
+ if newcs.charStringsAreIndexed:
286
+ for i, name in enumerate(cs.charStrings, start=len(newcs)):
287
+ newcs.charStrings[name] = i
288
+ newcs.charStringsIndex.items.append(None)
289
+ for name in cs.charStrings:
290
+ if widthsDiffer:
291
+ c = cs[name]
292
+ defaultWidthXToken = object()
293
+ extractor = T2WidthExtractor([], [], nominalWidthX, defaultWidthXToken)
294
+ extractor.execute(c)
295
+ width = extractor.width
296
+ if width is not defaultWidthXToken:
297
+ # The following will be wrong if the width is added
298
+ # by a subroutine. Ouch!
299
+ c.program.pop(0)
300
+ else:
301
+ width = defaultWidthX
302
+ if width != newDefaultWidthX:
303
+ c.program.insert(0, width - newNominalWidthX)
304
+ newcs[name] = cs[name]
305
+
306
+ newfont.charset = chrset
307
+ newfont.numGlyphs = len(chrset)
308
+ newfont.strings.strings = glyphOrderStrings + storedNamesStrings
309
+
310
+ return newcff
311
+
312
+
313
+ @add_method(ttLib.getTableClass("cmap"))
314
+ def merge(self, m, tables):
315
+ # TODO Handle format=14.
316
+ if not hasattr(m, "cmap"):
317
+ computeMegaCmap(m, tables)
318
+ cmap = m.cmap
319
+
320
+ cmapBmpOnly = {uni: gid for uni, gid in cmap.items() if uni <= 0xFFFF}
321
+ self.tables = []
322
+ module = ttLib.getTableModule("cmap")
323
+ if len(cmapBmpOnly) != len(cmap):
324
+ # format-12 required.
325
+ cmapTable = module.cmap_classes[12](12)
326
+ cmapTable.platformID = 3
327
+ cmapTable.platEncID = 10
328
+ cmapTable.language = 0
329
+ cmapTable.cmap = cmap
330
+ self.tables.append(cmapTable)
331
+ # always create format-4
332
+ cmapTable = module.cmap_classes[4](4)
333
+ cmapTable.platformID = 3
334
+ cmapTable.platEncID = 1
335
+ cmapTable.language = 0
336
+ cmapTable.cmap = cmapBmpOnly
337
+ # ordered by platform then encoding
338
+ self.tables.insert(0, cmapTable)
339
+ self.tableVersion = 0
340
+ self.numSubTables = len(self.tables)
341
+ return self
infer_4_47_1/lib/python3.10/site-packages/fontTools/merge/unicode.py ADDED
@@ -0,0 +1,78 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2021 Behdad Esfahbod. All Rights Reserved.
2
+
3
+
4
+ def is_Default_Ignorable(u):
5
+ # http://www.unicode.org/reports/tr44/#Default_Ignorable_Code_Point
6
+ #
7
+ # TODO Move me to unicodedata module and autogenerate.
8
+ #
9
+ # Unicode 14.0:
10
+ # $ grep '; Default_Ignorable_Code_Point ' DerivedCoreProperties.txt | sed 's/;.*#/#/'
11
+ # 00AD # Cf SOFT HYPHEN
12
+ # 034F # Mn COMBINING GRAPHEME JOINER
13
+ # 061C # Cf ARABIC LETTER MARK
14
+ # 115F..1160 # Lo [2] HANGUL CHOSEONG FILLER..HANGUL JUNGSEONG FILLER
15
+ # 17B4..17B5 # Mn [2] KHMER VOWEL INHERENT AQ..KHMER VOWEL INHERENT AA
16
+ # 180B..180D # Mn [3] MONGOLIAN FREE VARIATION SELECTOR ONE..MONGOLIAN FREE VARIATION SELECTOR THREE
17
+ # 180E # Cf MONGOLIAN VOWEL SEPARATOR
18
+ # 180F # Mn MONGOLIAN FREE VARIATION SELECTOR FOUR
19
+ # 200B..200F # Cf [5] ZERO WIDTH SPACE..RIGHT-TO-LEFT MARK
20
+ # 202A..202E # Cf [5] LEFT-TO-RIGHT EMBEDDING..RIGHT-TO-LEFT OVERRIDE
21
+ # 2060..2064 # Cf [5] WORD JOINER..INVISIBLE PLUS
22
+ # 2065 # Cn <reserved-2065>
23
+ # 2066..206F # Cf [10] LEFT-TO-RIGHT ISOLATE..NOMINAL DIGIT SHAPES
24
+ # 3164 # Lo HANGUL FILLER
25
+ # FE00..FE0F # Mn [16] VARIATION SELECTOR-1..VARIATION SELECTOR-16
26
+ # FEFF # Cf ZERO WIDTH NO-BREAK SPACE
27
+ # FFA0 # Lo HALFWIDTH HANGUL FILLER
28
+ # FFF0..FFF8 # Cn [9] <reserved-FFF0>..<reserved-FFF8>
29
+ # 1BCA0..1BCA3 # Cf [4] SHORTHAND FORMAT LETTER OVERLAP..SHORTHAND FORMAT UP STEP
30
+ # 1D173..1D17A # Cf [8] MUSICAL SYMBOL BEGIN BEAM..MUSICAL SYMBOL END PHRASE
31
+ # E0000 # Cn <reserved-E0000>
32
+ # E0001 # Cf LANGUAGE TAG
33
+ # E0002..E001F # Cn [30] <reserved-E0002>..<reserved-E001F>
34
+ # E0020..E007F # Cf [96] TAG SPACE..CANCEL TAG
35
+ # E0080..E00FF # Cn [128] <reserved-E0080>..<reserved-E00FF>
36
+ # E0100..E01EF # Mn [240] VARIATION SELECTOR-17..VARIATION SELECTOR-256
37
+ # E01F0..E0FFF # Cn [3600] <reserved-E01F0>..<reserved-E0FFF>
38
+ return (
39
+ u == 0x00AD
40
+ or u == 0x034F # Cf SOFT HYPHEN
41
+ or u == 0x061C # Mn COMBINING GRAPHEME JOINER
42
+ or 0x115F <= u <= 0x1160 # Cf ARABIC LETTER MARK
43
+ or 0x17B4 # Lo [2] HANGUL CHOSEONG FILLER..HANGUL JUNGSEONG FILLER
44
+ <= u
45
+ <= 0x17B5
46
+ or 0x180B # Mn [2] KHMER VOWEL INHERENT AQ..KHMER VOWEL INHERENT AA
47
+ <= u
48
+ <= 0x180D
49
+ or u # Mn [3] MONGOLIAN FREE VARIATION SELECTOR ONE..MONGOLIAN FREE VARIATION SELECTOR THREE
50
+ == 0x180E
51
+ or u == 0x180F # Cf MONGOLIAN VOWEL SEPARATOR
52
+ or 0x200B <= u <= 0x200F # Mn MONGOLIAN FREE VARIATION SELECTOR FOUR
53
+ or 0x202A <= u <= 0x202E # Cf [5] ZERO WIDTH SPACE..RIGHT-TO-LEFT MARK
54
+ or 0x2060 # Cf [5] LEFT-TO-RIGHT EMBEDDING..RIGHT-TO-LEFT OVERRIDE
55
+ <= u
56
+ <= 0x2064
57
+ or u == 0x2065 # Cf [5] WORD JOINER..INVISIBLE PLUS
58
+ or 0x2066 <= u <= 0x206F # Cn <reserved-2065>
59
+ or u == 0x3164 # Cf [10] LEFT-TO-RIGHT ISOLATE..NOMINAL DIGIT SHAPES
60
+ or 0xFE00 <= u <= 0xFE0F # Lo HANGUL FILLER
61
+ or u == 0xFEFF # Mn [16] VARIATION SELECTOR-1..VARIATION SELECTOR-16
62
+ or u == 0xFFA0 # Cf ZERO WIDTH NO-BREAK SPACE
63
+ or 0xFFF0 <= u <= 0xFFF8 # Lo HALFWIDTH HANGUL FILLER
64
+ or 0x1BCA0 <= u <= 0x1BCA3 # Cn [9] <reserved-FFF0>..<reserved-FFF8>
65
+ or 0x1D173 # Cf [4] SHORTHAND FORMAT LETTER OVERLAP..SHORTHAND FORMAT UP STEP
66
+ <= u
67
+ <= 0x1D17A
68
+ or u == 0xE0000 # Cf [8] MUSICAL SYMBOL BEGIN BEAM..MUSICAL SYMBOL END PHRASE
69
+ or u == 0xE0001 # Cn <reserved-E0000>
70
+ or 0xE0002 <= u <= 0xE001F # Cf LANGUAGE TAG
71
+ or 0xE0020 <= u <= 0xE007F # Cn [30] <reserved-E0002>..<reserved-E001F>
72
+ or 0xE0080 <= u <= 0xE00FF # Cf [96] TAG SPACE..CANCEL TAG
73
+ or 0xE0100 <= u <= 0xE01EF # Cn [128] <reserved-E0080>..<reserved-E00FF>
74
+ or 0xE01F0 # Mn [240] VARIATION SELECTOR-17..VARIATION SELECTOR-256
75
+ <= u
76
+ <= 0xE0FFF
77
+ or False # Cn [3600] <reserved-E01F0>..<reserved-E0FFF>
78
+ )