Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
@@ -1,37 +1,17 @@
|
|
|
|
|
|
|
|
1 |
import gradio as gr
|
2 |
import torch
|
3 |
-
import timm
|
4 |
import torch.nn.functional as F
|
5 |
-
|
6 |
from timm.data import create_transform
|
|
|
|
|
7 |
from PIL import Image
|
8 |
import numpy as np
|
9 |
import matplotlib.pyplot as plt
|
10 |
-
from typing import List, Tuple, Dict
|
11 |
-
from collections import OrderedDict
|
12 |
|
13 |
-
class AttentionExtractor:
|
14 |
-
def __init__(self, model: torch.nn.Module):
|
15 |
-
self.model = model
|
16 |
-
self.attention_maps = OrderedDict()
|
17 |
-
self._register_hooks()
|
18 |
-
|
19 |
-
def _register_hooks(self):
|
20 |
-
def hook_fn(module, input, output):
|
21 |
-
if isinstance(output, tuple):
|
22 |
-
self.attention_maps[module.full_name] = output[1] # attention_probs
|
23 |
-
else:
|
24 |
-
self.attention_maps[module.full_name] = output
|
25 |
-
|
26 |
-
for name, module in self.model.named_modules():
|
27 |
-
# FIXME need to make more generic outside of vit
|
28 |
-
if name.lower().endswith('.attn_drop'):
|
29 |
-
module.full_name = name
|
30 |
-
print('hooking', name)
|
31 |
-
module.register_forward_hook(hook_fn)
|
32 |
-
|
33 |
-
def get_attention_maps(self) -> OrderedDict:
|
34 |
-
return self.attention_maps
|
35 |
|
36 |
def get_attention_models() -> List[str]:
|
37 |
"""Get a list of timm models that have attention blocks."""
|
@@ -45,7 +25,7 @@ def load_model(model_name: str) -> Tuple[torch.nn.Module, AttentionExtractor]:
|
|
45 |
timm.layers.set_fused_attn(False)
|
46 |
model = create_model(model_name, pretrained=True)
|
47 |
model.eval()
|
48 |
-
extractor =
|
49 |
return model, extractor
|
50 |
|
51 |
def process_image(image: Image.Image, model: torch.nn.Module, extractor: AttentionExtractor) -> Dict[str, torch.Tensor]:
|
@@ -61,16 +41,11 @@ def process_image(image: Image.Image, model: torch.nn.Module, extractor: Attenti
|
|
61 |
is_training=False
|
62 |
)
|
63 |
|
64 |
-
|
65 |
# Preprocess the image
|
66 |
tensor = transform(image).unsqueeze(0)
|
67 |
-
|
68 |
-
# Forward pass
|
69 |
-
with torch.no_grad():
|
70 |
-
_ = model(tensor)
|
71 |
-
|
72 |
# Extract attention maps
|
73 |
-
attention_maps = extractor
|
74 |
|
75 |
return attention_maps
|
76 |
|
|
|
1 |
+
from typing import List, Tuple, Dict
|
2 |
+
from collections import OrderedDict
|
3 |
+
|
4 |
import gradio as gr
|
5 |
import torch
|
|
|
6 |
import torch.nn.functional as F
|
7 |
+
import timm
|
8 |
from timm.data import create_transform
|
9 |
+
from timm.models import create_model
|
10 |
+
from timm.utils import AttentionExtract
|
11 |
from PIL import Image
|
12 |
import numpy as np
|
13 |
import matplotlib.pyplot as plt
|
|
|
|
|
14 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
15 |
|
16 |
def get_attention_models() -> List[str]:
|
17 |
"""Get a list of timm models that have attention blocks."""
|
|
|
25 |
timm.layers.set_fused_attn(False)
|
26 |
model = create_model(model_name, pretrained=True)
|
27 |
model.eval()
|
28 |
+
extractor = AttentionExtract(model, method='fx') # can use 'hooks', can also allow specifying matching names for attention nodes or modules...
|
29 |
return model, extractor
|
30 |
|
31 |
def process_image(image: Image.Image, model: torch.nn.Module, extractor: AttentionExtractor) -> Dict[str, torch.Tensor]:
|
|
|
41 |
is_training=False
|
42 |
)
|
43 |
|
|
|
44 |
# Preprocess the image
|
45 |
tensor = transform(image).unsqueeze(0)
|
46 |
+
|
|
|
|
|
|
|
|
|
47 |
# Extract attention maps
|
48 |
+
attention_maps = extractor(tensor)
|
49 |
|
50 |
return attention_maps
|
51 |
|