timm
/

Image Feature Extraction
timm
PyTorch
Safetensors
Transformers
rwightman HF Staff commited on
Commit
b74d1ab
·
verified ·
1 Parent(s): 8f1aa10

Update model config and README

Browse files
Files changed (2) hide show
  1. README.md +7 -6
  2. config.json +1 -1
README.md CHANGED
@@ -1,5 +1,6 @@
1
  ---
2
  tags:
 
3
  - timm
4
  - transformers
5
  pipeline_tag: image-feature-extraction
@@ -10,7 +11,7 @@ license_link: https://ai.meta.com/resources/models-and-libraries/dinov3-license
10
  datasets:
11
  - lvd-1689m
12
  ---
13
- # Model card for vit_small_plus_patch16_dinov3_qkvb.lvdm_1689m
14
 
15
  A DINOv3 ViT model image feature encoder. Distilled on LVD-1689M from the DINOv3 ViT-7B model.
16
 
@@ -19,7 +20,7 @@ A DINOv3 ViT model image feature encoder. Distilled on LVD-1689M from the DINOv3
19
  * The original models keep RoPE periods as a persistent `bfloat16` buffer. `timm` generates `float32` periods at init. This results in some numerical differences, however the `timm` approach should be less problematic running on devices without bfloat16 support, and appears to work as well if not slightly better for fine-tuning. `model.rope.periods = model.rope.periods.to(torch.bfloat16).to(torch.float32)` will truncate the periods to bfloat16 and result in matching outputs.
20
 
21
  ## Model Details
22
- - **Model Type:** Image feature encoder
23
  - **Model Stats:**
24
  - Params (M): 28.7
25
  - GMACs: 8.1
@@ -44,7 +45,7 @@ img = Image.open(urlopen(
44
  'https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/beignets-task-guide.png'
45
  ))
46
 
47
- model = timm.create_model('vit_small_plus_patch16_dinov3_qkvb.lvdm_1689m', pretrained=True)
48
  model = model.eval()
49
 
50
  # get model specific transforms (normalization, resize)
@@ -67,7 +68,7 @@ img = Image.open(urlopen(
67
  ))
68
 
69
  model = timm.create_model(
70
- 'vit_small_plus_patch16_dinov3_qkvb.lvdm_1689m',
71
  pretrained=True,
72
  features_only=True,
73
  )
@@ -100,7 +101,7 @@ img = Image.open(urlopen(
100
  ))
101
 
102
  model = timm.create_model(
103
- 'vit_small_plus_patch16_dinov3_qkvb.lvdm_1689m',
104
  pretrained=True,
105
  num_classes=0, # remove classifier nn.Linear
106
  )
@@ -190,4 +191,4 @@ See the associated paper for details on the evaluation protocols
190
  doi = {10.5281/zenodo.4414861},
191
  howpublished = {\url{https://github.com/huggingface/pytorch-image-models}}
192
  }
193
- ```
 
1
  ---
2
  tags:
3
+ - image-feature-extraction
4
  - timm
5
  - transformers
6
  pipeline_tag: image-feature-extraction
 
11
  datasets:
12
  - lvd-1689m
13
  ---
14
+ # Model card for vit_small_plus_patch16_dinov3_qkvb.lvd_1689m
15
 
16
  A DINOv3 ViT model image feature encoder. Distilled on LVD-1689M from the DINOv3 ViT-7B model.
17
 
 
20
  * The original models keep RoPE periods as a persistent `bfloat16` buffer. `timm` generates `float32` periods at init. This results in some numerical differences, however the `timm` approach should be less problematic running on devices without bfloat16 support, and appears to work as well if not slightly better for fine-tuning. `model.rope.periods = model.rope.periods.to(torch.bfloat16).to(torch.float32)` will truncate the periods to bfloat16 and result in matching outputs.
21
 
22
  ## Model Details
23
+ - **Model Type:** Image Feature Encoder
24
  - **Model Stats:**
25
  - Params (M): 28.7
26
  - GMACs: 8.1
 
45
  'https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/beignets-task-guide.png'
46
  ))
47
 
48
+ model = timm.create_model('vit_small_plus_patch16_dinov3_qkvb.lvd_1689m', pretrained=True)
49
  model = model.eval()
50
 
51
  # get model specific transforms (normalization, resize)
 
68
  ))
69
 
70
  model = timm.create_model(
71
+ 'vit_small_plus_patch16_dinov3_qkvb.lvd_1689m',
72
  pretrained=True,
73
  features_only=True,
74
  )
 
101
  ))
102
 
103
  model = timm.create_model(
104
+ 'vit_small_plus_patch16_dinov3_qkvb.lvd_1689m',
105
  pretrained=True,
106
  num_classes=0, # remove classifier nn.Linear
107
  )
 
191
  doi = {10.5281/zenodo.4414861},
192
  howpublished = {\url{https://github.com/huggingface/pytorch-image-models}}
193
  }
194
+ ```
config.json CHANGED
@@ -4,7 +4,7 @@
4
  "num_features": 384,
5
  "global_pool": "avg",
6
  "pretrained_cfg": {
7
- "tag": "lvdm_1689m",
8
  "custom_load": false,
9
  "input_size": [
10
  3,
 
4
  "num_features": 384,
5
  "global_pool": "avg",
6
  "pretrained_cfg": {
7
+ "tag": "lvd_1689m",
8
  "custom_load": false,
9
  "input_size": [
10
  3,