Spaces:
Running
Running
Create app.py
Browse files
app.py
ADDED
|
@@ -0,0 +1,151 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import gradio as gr
|
| 2 |
+
import torch
|
| 3 |
+
import timm
|
| 4 |
+
import torch.nn.functional as F
|
| 5 |
+
from timm.models import create_model
|
| 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 |
+
if name.lower().endswith('.attn_drop'):
|
| 28 |
+
module.full_name = name
|
| 29 |
+
print('hooking', name)
|
| 30 |
+
module.register_forward_hook(hook_fn)
|
| 31 |
+
|
| 32 |
+
def get_attention_maps(self) -> OrderedDict:
|
| 33 |
+
return self.attention_maps
|
| 34 |
+
|
| 35 |
+
def get_attention_models() -> List[str]:
|
| 36 |
+
"""Get a list of timm models that have attention blocks."""
|
| 37 |
+
all_models = timm.list_models()
|
| 38 |
+
attention_models = [model for model in all_models if 'vit' in model.lower()] # Focusing on ViT models for simplicity
|
| 39 |
+
return attention_models
|
| 40 |
+
|
| 41 |
+
def load_model(model_name: str) -> Tuple[torch.nn.Module, AttentionExtractor]:
|
| 42 |
+
"""Load a model from timm and prepare it for attention extraction."""
|
| 43 |
+
timm.layers.set_fused_attn(False)
|
| 44 |
+
model = create_model(model_name, pretrained=True)
|
| 45 |
+
model.eval()
|
| 46 |
+
extractor = AttentionExtractor(model)
|
| 47 |
+
return model, extractor
|
| 48 |
+
|
| 49 |
+
def process_image(image: Image.Image, model: torch.nn.Module, extractor: AttentionExtractor) -> Dict[str, torch.Tensor]:
|
| 50 |
+
"""Process the input image and get the attention maps."""
|
| 51 |
+
# Get the correct transform for the model
|
| 52 |
+
config = model.pretrained_cfg
|
| 53 |
+
transform = create_transform(
|
| 54 |
+
input_size=config['input_size'],
|
| 55 |
+
crop_pct=config['crop_pct'],
|
| 56 |
+
mean=config['mean'],
|
| 57 |
+
std=config['std'],
|
| 58 |
+
interpolation=config['interpolation'],
|
| 59 |
+
is_training=False
|
| 60 |
+
)
|
| 61 |
+
|
| 62 |
+
|
| 63 |
+
# Preprocess the image
|
| 64 |
+
tensor = transform(image).unsqueeze(0)
|
| 65 |
+
|
| 66 |
+
# Forward pass
|
| 67 |
+
with torch.no_grad():
|
| 68 |
+
_ = model(tensor)
|
| 69 |
+
|
| 70 |
+
# Extract attention maps
|
| 71 |
+
attention_maps = extractor.get_attention_maps()
|
| 72 |
+
|
| 73 |
+
return attention_maps
|
| 74 |
+
|
| 75 |
+
def apply_mask(image: np.ndarray, mask: np.ndarray, color: Tuple[float, float, float], alpha: float = 0.5) -> np.ndarray:
|
| 76 |
+
# Ensure mask and image have the same shape
|
| 77 |
+
mask = mask[:, :, np.newaxis]
|
| 78 |
+
mask = np.repeat(mask, 3, axis=2)
|
| 79 |
+
|
| 80 |
+
# Convert color to numpy array
|
| 81 |
+
color = np.array(color)
|
| 82 |
+
|
| 83 |
+
# Apply mask
|
| 84 |
+
masked_image = image * (1 - alpha * mask) + alpha * mask * color[np.newaxis, np.newaxis, :] * 255
|
| 85 |
+
return masked_image.astype(np.uint8)
|
| 86 |
+
|
| 87 |
+
def visualize_attention(image: Image.Image, model_name: str) -> List[Image.Image]:
|
| 88 |
+
"""Visualize attention maps for the given image and model."""
|
| 89 |
+
model, extractor = load_model(model_name)
|
| 90 |
+
attention_maps = process_image(image, model, extractor)
|
| 91 |
+
|
| 92 |
+
# Convert PIL Image to numpy array
|
| 93 |
+
image_np = np.array(image)
|
| 94 |
+
|
| 95 |
+
# Create visualizations
|
| 96 |
+
visualizations = []
|
| 97 |
+
for layer_name, attn_map in attention_maps.items():
|
| 98 |
+
print(f"Attention map shape for {layer_name}: {attn_map.shape}")
|
| 99 |
+
|
| 100 |
+
# Remove the CLS token attention and average over heads
|
| 101 |
+
attn_map = attn_map[0, :, 0, 1:].mean(0) # Shape: (seq_len-1,)
|
| 102 |
+
|
| 103 |
+
# Reshape the attention map to 2D
|
| 104 |
+
num_patches = int(np.sqrt(attn_map.shape[0]))
|
| 105 |
+
attn_map = attn_map.reshape(num_patches, num_patches)
|
| 106 |
+
|
| 107 |
+
# Interpolate to match image size
|
| 108 |
+
attn_map = torch.tensor(attn_map).unsqueeze(0).unsqueeze(0)
|
| 109 |
+
attn_map = F.interpolate(attn_map, size=(image_np.shape[0], image_np.shape[1]), mode='bilinear', align_corners=False)
|
| 110 |
+
attn_map = attn_map.squeeze().cpu().numpy()
|
| 111 |
+
|
| 112 |
+
# Normalize attention map
|
| 113 |
+
attn_map = (attn_map - attn_map.min()) / (attn_map.max() - attn_map.min())
|
| 114 |
+
|
| 115 |
+
# Create visualization
|
| 116 |
+
fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(20, 10))
|
| 117 |
+
|
| 118 |
+
# Original image
|
| 119 |
+
ax1.imshow(image_np)
|
| 120 |
+
ax1.set_title("Original Image")
|
| 121 |
+
ax1.axis('off')
|
| 122 |
+
|
| 123 |
+
# Attention map overlay
|
| 124 |
+
masked_image = apply_mask(image_np, attn_map, color=(1, 0, 0)) # Red mask
|
| 125 |
+
ax2.imshow(masked_image)
|
| 126 |
+
ax2.set_title(f'Attention Map for {layer_name}')
|
| 127 |
+
ax2.axis('off')
|
| 128 |
+
|
| 129 |
+
plt.tight_layout()
|
| 130 |
+
|
| 131 |
+
# Convert plot to image
|
| 132 |
+
fig.canvas.draw()
|
| 133 |
+
vis_image = Image.frombytes('RGB', fig.canvas.get_width_height(), fig.canvas.tostring_rgb())
|
| 134 |
+
visualizations.append(vis_image)
|
| 135 |
+
plt.close(fig)
|
| 136 |
+
|
| 137 |
+
return visualizations
|
| 138 |
+
|
| 139 |
+
# Create Gradio interface
|
| 140 |
+
iface = gr.Interface(
|
| 141 |
+
fn=visualize_attention,
|
| 142 |
+
inputs=[
|
| 143 |
+
gr.Image(type="pil", label="Input Image"),
|
| 144 |
+
gr.Dropdown(choices=get_attention_models(), label="Select Model")
|
| 145 |
+
],
|
| 146 |
+
outputs=gr.Gallery(label="Attention Maps"),
|
| 147 |
+
title="Attention Map Visualizer for timm Models",
|
| 148 |
+
description="Upload an image and select a timm model to visualize its attention maps."
|
| 149 |
+
)
|
| 150 |
+
|
| 151 |
+
iface.launch(debug=True)
|