Spaces:
Running
Running
Refactor to use hf_hub_download instead of torch.hub.load
Browse files- app.py +15 -5
- timmfrv2.py +84 -0
app.py
CHANGED
@@ -12,9 +12,12 @@ import numpy as np
|
|
12 |
import torch
|
13 |
import torch.nn.functional as F
|
14 |
from torchvision import transforms
|
|
|
15 |
|
16 |
from utils import align_crop
|
17 |
from title import title_css, title_with_logo
|
|
|
|
|
18 |
|
19 |
# ───────────────────────────────
|
20 |
# Data & models
|
@@ -233,11 +236,18 @@ _tx = transforms.Compose(
|
|
233 |
|
234 |
def get_edge_model(name: str) -> torch.nn.Module:
|
235 |
if name not in get_edge_model.cache:
|
236 |
-
|
237 |
-
|
238 |
-
|
239 |
-
|
240 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
241 |
return get_edge_model.cache[name]
|
242 |
|
243 |
|
|
|
12 |
import torch
|
13 |
import torch.nn.functional as F
|
14 |
from torchvision import transforms
|
15 |
+
from huggingface_hub import hf_hub_download
|
16 |
|
17 |
from utils import align_crop
|
18 |
from title import title_css, title_with_logo
|
19 |
+
from timmfrv2 import TimmFRWrapperV2, model_configs
|
20 |
+
|
21 |
|
22 |
# ───────────────────────────────
|
23 |
# Data & models
|
|
|
236 |
|
237 |
def get_edge_model(name: str) -> torch.nn.Module:
|
238 |
if name not in get_edge_model.cache:
|
239 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
240 |
+
model_path = hf_hub_download(
|
241 |
+
repo_id=model_configs[name]["repo"],
|
242 |
+
filename=model_configs[name]["filename"],
|
243 |
+
local_dir="models",
|
244 |
+
)
|
245 |
+
model = TimmFRWrapperV2(model_configs[name]["timm_model"], batchnorm=False)
|
246 |
+
model = model_configs[name]["post_setup"](model)
|
247 |
+
model.load_state_dict(torch.load(model_path, map_location="cpu"))
|
248 |
+
model = model.eval()
|
249 |
+
model.to(device)
|
250 |
+
get_edge_model.cache[name] = model
|
251 |
return get_edge_model.cache[name]
|
252 |
|
253 |
|
timmfrv2.py
ADDED
@@ -0,0 +1,84 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch.nn as nn
|
2 |
+
import timm
|
3 |
+
|
4 |
+
|
5 |
+
class TimmFRWrapperV2(nn.Module):
|
6 |
+
"""
|
7 |
+
Wraps timm model
|
8 |
+
"""
|
9 |
+
|
10 |
+
def __init__(self, model_name="edgenext_x_small", featdim=512, batchnorm=False):
|
11 |
+
super().__init__()
|
12 |
+
self.featdim = featdim
|
13 |
+
self.model_name = model_name
|
14 |
+
|
15 |
+
self.model = timm.create_model(self.model_name)
|
16 |
+
self.model.reset_classifier(self.featdim)
|
17 |
+
|
18 |
+
def forward(self, x):
|
19 |
+
x = self.model(x)
|
20 |
+
return x
|
21 |
+
|
22 |
+
|
23 |
+
class LoRaLin(nn.Module):
|
24 |
+
def __init__(self, in_features, out_features, rank, bias=True):
|
25 |
+
super(LoRaLin, self).__init__()
|
26 |
+
self.in_features = in_features
|
27 |
+
self.out_features = out_features
|
28 |
+
self.rank = rank
|
29 |
+
self.linear1 = nn.Linear(in_features, rank, bias=False)
|
30 |
+
self.linear2 = nn.Linear(rank, out_features, bias=bias)
|
31 |
+
|
32 |
+
def forward(self, input):
|
33 |
+
x = self.linear1(input)
|
34 |
+
x = self.linear2(x)
|
35 |
+
return x
|
36 |
+
|
37 |
+
|
38 |
+
def replace_linear_with_lowrank_recursive_2(model, rank_ratio=0.2):
|
39 |
+
for name, module in model.named_children():
|
40 |
+
if isinstance(module, nn.Linear) and "head" not in name:
|
41 |
+
in_features = module.in_features
|
42 |
+
out_features = module.out_features
|
43 |
+
rank = max(2, int(min(in_features, out_features) * rank_ratio))
|
44 |
+
bias = False
|
45 |
+
if module.bias is not None:
|
46 |
+
bias = True
|
47 |
+
lowrank_module = LoRaLin(in_features, out_features, rank, bias)
|
48 |
+
|
49 |
+
setattr(model, name, lowrank_module)
|
50 |
+
else:
|
51 |
+
replace_linear_with_lowrank_recursive_2(module, rank_ratio)
|
52 |
+
|
53 |
+
|
54 |
+
def replace_linear_with_lowrank_2(model, rank_ratio=0.2):
|
55 |
+
replace_linear_with_lowrank_recursive_2(model, rank_ratio)
|
56 |
+
return model
|
57 |
+
|
58 |
+
|
59 |
+
model_configs = {
|
60 |
+
"edgeface_base": {
|
61 |
+
"repo": "idiap/EdgeFace-Base",
|
62 |
+
"filename": "edgeface_base.pt",
|
63 |
+
"timm_model": "edgenext_base",
|
64 |
+
"post_setup": lambda x: x,
|
65 |
+
},
|
66 |
+
"edgeface_s_gamma_05": {
|
67 |
+
"repo": "idiap/EdgeFace-S-GAMMA",
|
68 |
+
"filename": "edgeface_s_gamma_05.pt",
|
69 |
+
"timm_model": "edgenext_small",
|
70 |
+
"post_setup": lambda x: replace_linear_with_lowrank_2(x, rank_ratio=0.5),
|
71 |
+
},
|
72 |
+
"edgeface_xs_gamma_06": {
|
73 |
+
"repo": "idiap/EdgeFace-XS-GAMMA",
|
74 |
+
"filename": "edgeface_xs_gamma_06.pt",
|
75 |
+
"timm_model": "edgenext_x_small",
|
76 |
+
"post_setup": lambda x: replace_linear_with_lowrank_2(x, rank_ratio=0.6),
|
77 |
+
},
|
78 |
+
"edgeface_xxs": {
|
79 |
+
"repo": "idiap/EdgeFace-XXS",
|
80 |
+
"filename": "edgeface_xxs.pt",
|
81 |
+
"timm_model": "edgenext_xx_small",
|
82 |
+
"post_setup": lambda x: x,
|
83 |
+
},
|
84 |
+
}
|