File size: 11,807 Bytes
a01dc02
 
 
be5c319
 
 
a01dc02
be5c319
 
 
 
a01dc02
 
 
 
 
 
be5c319
a01dc02
 
 
be5c319
a01dc02
 
 
 
 
be5c319
a01dc02
 
be5c319
a01dc02
 
be5c319
a01dc02
 
 
 
 
 
 
be5c319
a01dc02
 
 
 
 
 
be5c319
a01dc02
 
 
be5c319
0101a8b
a01dc02
be5c319
0101a8b
 
a01dc02
0101a8b
 
a01dc02
0101a8b
 
 
 
 
 
 
 
 
 
 
 
 
 
be5c319
a01dc02
 
 
be5c319
a01dc02
 
be5c319
a01dc02
 
be5c319
a01dc02
 
 
be5c319
a01dc02
 
be5c319
a01dc02
 
be5c319
a01dc02
be5c319
 
 
 
 
 
 
 
 
 
a01dc02
 
be5c319
a01dc02
 
be5c319
a01dc02
 
 
 
be5c319
 
 
 
 
 
 
 
 
 
a01dc02
 
be5c319
a01dc02
 
 
 
 
 
be5c319
a01dc02
 
be5c319
 
a01dc02
 
 
 
be5c319
0101a8b
 
be5c319
a01dc02
 
be5c319
a01dc02
 
be5c319
a01dc02
 
be5c319
0101a8b
a01dc02
 
be5c319
0101a8b
 
 
 
 
 
 
 
 
be5c319
a01dc02
 
be5c319
 
 
a01dc02
 
be5c319
a01dc02
 
0101a8b
a01dc02
 
be5c319
a01dc02
 
be5c319
a01dc02
 
be5c319
 
 
a01dc02
be5c319
 
 
 
a01dc02
 
be5c319
 
 
 
a01dc02
be5c319
a01dc02
 
 
 
be5c319
a01dc02
be5c319
 
a01dc02
be5c319
 
 
a01dc02
 
 
be5c319
 
 
 
 
 
 
 
 
 
a01dc02
 
be5c319
0101a8b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a01dc02
 
 
 
 
 
be5c319
a01dc02
 
be5c319
 
a01dc02
 
 
 
be5c319
a01dc02
 
be5c319
a01dc02
 
be5c319
a01dc02
 
be5c319
a01dc02
 
be5c319
a01dc02
be5c319
a01dc02
 
be5c319
a01dc02
 
be5c319
 
 
a01dc02
 
be5c319
a01dc02
 
be5c319
a01dc02
 
be5c319
 
 
a01dc02
be5c319
 
 
a01dc02
be5c319
a01dc02
 
be5c319
 
 
a01dc02
be5c319
a01dc02
 
be5c319
a01dc02
 
 
be5c319
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
# src/explainer.py

import captum
import matplotlib.pyplot as plt
import numpy as np
import torch
import torch.nn.functional as F
from captum.attr import GradientShap, LayerGradCam
from captum.attr import visualization as viz
from PIL import Image


class ViTWrapper(torch.nn.Module):
    """
    Wrapper class to make Hugging Face ViT compatible with Captum.
    This returns raw tensors instead of Hugging Face output objects.
    """

    def __init__(self, model):
        super().__init__()
        self.model = model

    def forward(self, x):
        # Hugging Face models expect pixel_values key
        outputs = self.model(pixel_values=x)
        return outputs.logits


class AttentionHook:
    """Hook to capture attention weights from ViT model"""

    def __init__(self):
        self.attention_weights = None

    def __call__(self, module, input, output):
        # For ViT, attention weights are usually the second output
        if len(output) >= 2:
            self.attention_weights = output[1]  # attention weights
        else:
            self.attention_weights = None


def explain_attention(model, processor, image, layer_index=6, head_index=0):
    """
    Extract and visualize attention weights using hooks.
    """
    try:
        device = next(model.parameters()).device

        # Preprocess image
        inputs = processor(images=image, return_tensors="pt")
        inputs = {k: v.to(device) for k, v in inputs.items()}

        # Register hook to capture attention (supported for ViT/DeiT only)
        hook = AttentionHook()

        # Only support attention visualization for ViT-like architectures
        if hasattr(model, "vit"):
            try:
                # For standard ViT structure
                target_layer = model.vit.encoder.layer[layer_index].attention.attention
                handle = target_layer.register_forward_hook(hook)
            except Exception:
                try:
                    # Alternative structure
                    target_layer = model.vit.encoder.layers[layer_index].attention.attention
                    handle = target_layer.register_forward_hook(hook)
                except Exception:
                    raise ValueError(
                        f"Could not access layer {layer_index} for attention hook"
                    )
        else:
            raise ValueError(
                "Attention visualization currently supports ViT/DeiT models only. "
                "Please select a ViT model or use GradCAM/GradientSHAP."
            )

        # Forward pass to capture attention
        with torch.no_grad():
            _ = model(**inputs)

        # Remove hook
        handle.remove()

        if hook.attention_weights is None:
            raise ValueError("No attention weights captured by hook")

        # Get attention weights
        attention_weights = hook.attention_weights  # Shape: (batch, heads, seq_len, seq_len)
        attention_map = attention_weights[0, head_index]  # Shape: (seq_len, seq_len)

        # Remove CLS token attention to other tokens
        patch_attention = attention_map[1:, 1:]  # Remove CLS token rows and columns

        # Create visualization
        fig, ax = plt.subplots(figsize=(8, 6))

        # Display attention matrix
        im = ax.imshow(patch_attention.cpu().numpy(), cmap="viridis", aspect="auto")

        ax.set_title(
            f"Attention Map - Layer {layer_index}, Head {head_index}",
            fontsize=14,
            fontweight="bold",
        )
        ax.set_xlabel("Key Patches")
        ax.set_ylabel("Query Patches")

        # Add colorbar
        plt.colorbar(im, ax=ax)

        plt.tight_layout()
        return fig

    except Exception as e:
        print(f"Error in attention visualization: {str(e)}")
        # Return a simple error plot
        fig, ax = plt.subplots(figsize=(8, 6))
        ax.text(
            0.5,
            0.5,
            f"Attention visualization failed:\n{str(e)}",
            ha="center",
            va="center",
            transform=ax.transAxes,
            fontsize=10,
        )
        ax.set_title("Attention Visualization Error")
        return fig


def explain_gradcam(model, processor, image, target_layer_index=-2):
    """
    Generate GradCAM heatmap for the predicted class.
    """
    try:
        device = next(model.parameters()).device

        # Preprocess image
        inputs = processor(images=image, return_tensors="pt")
        input_tensor = inputs["pixel_values"].to(device)

        # Get prediction
        with torch.no_grad():
            outputs = model(input_tensor)
            predicted_class = outputs.logits.argmax(dim=1).item()

        # Get the target layer adaptively across architectures
        target_layer = _select_gradcam_target_layer(model, target_layer_index)

        # Create wrapped model for Captum compatibility
        wrapped_model = ViTWrapper(model)

        # Initialize GradCAM with wrapped model
        gradcam = LayerGradCam(wrapped_model, target_layer)

        # Generate attribution - handle tuple output
        attribution = gradcam.attribute(input_tensor, target=predicted_class)

        # Handle tuple output by taking the first element
        if isinstance(attribution, tuple):
            attribution = attribution[0]

        # If attribution has channel dimension, aggregate over channels
        if isinstance(attribution, torch.Tensor):
            att = attribution.detach().cpu()
            if att.dim() == 4:  # (B, C, H, W)
                att = att.sum(dim=1)  # (B, H, W)
            att = att.squeeze(0)  # (H, W)
            attribution = att.numpy()
        else:
            attribution = np.array(attribution)

        # Normalize attribution
        if attribution.max() > attribution.min():
            attribution = (attribution - attribution.min()) / (
                attribution.max() - attribution.min()
            )
        else:
            attribution = np.zeros_like(attribution)

        # Resize heatmap to match original image
        original_size = image.size
        heatmap = Image.fromarray((np.clip(attribution, 0, 1) * 255).astype(np.uint8))
        heatmap = heatmap.resize(original_size, Image.Resampling.LANCZOS)
        heatmap = np.array(heatmap)

        # Create visualization figure
        fig, (ax1, ax2, ax3) = plt.subplots(1, 3, figsize=(15, 5))

        # Original image
        ax1.imshow(image)
        ax1.set_title("Original Image")
        ax1.axis("off")

        # Heatmap
        ax2.imshow(heatmap, cmap="hot")
        ax2.set_title("GradCAM Heatmap")
        ax2.axis("off")

        # Overlay
        ax3.imshow(image)
        ax3.imshow(heatmap, cmap="hot", alpha=0.5)
        ax3.set_title("Overlay")
        ax3.axis("off")

        plt.tight_layout()

        # Create overlay image for dashboard
        heatmap_rgb = (plt.cm.hot(heatmap / 255.0)[:, :, :3] * 255).astype(np.uint8)
        overlay_img = Image.fromarray(heatmap_rgb)
        overlay_img = overlay_img.resize(original_size, Image.Resampling.LANCZOS)

        # Blend with original
        original_rgba = image.convert("RGBA")
        overlay_rgba = overlay_img.convert("RGBA")
        blended = Image.blend(original_rgba, overlay_rgba, alpha=0.5)

        return fig, blended.convert("RGB")

    except Exception as e:
        print(f"Error in GradCAM: {str(e)}")
        fig, ax = plt.subplots(figsize=(8, 6))
        ax.text(
            0.5,
            0.5,
            f"GradCAM failed:\n{str(e)}",
            ha="center",
            va="center",
            transform=ax.transAxes,
            fontsize=10,
        )
        ax.set_title("GradCAM Error")
        return fig, image


def _select_gradcam_target_layer(model, target_layer_index):
    """Best-effort selection of a target layer for GradCAM across architectures."""
    # 1) ViT / DeiT: use attention layer as before
    if hasattr(model, "vit"):
        try:
            return model.vit.encoder.layer[target_layer_index].attention.attention
        except Exception:
            return model.vit.encoder.layers[target_layer_index].attention.attention

    # 2) ResNet (HF): try final bottleneck/conv in layer4
    if hasattr(model, "resnet"):
        res = model.resnet
        try:
            blk = res.layer4[-1]
            # Prefer the last conv if exists
            for attr in ["conv3", "conv2", "conv1"]:
                if hasattr(blk, attr):
                    return getattr(blk, attr)
            return blk
        except Exception:
            pass

    # 3) Swin (HF): try last attention block; fallback to patch embedding conv
    if hasattr(model, "swin"):
        try:
            # Common pattern: encoder.layers[-1].blocks[-1].attention
            return model.swin.encoder.layers[-1].blocks[-1].attention
        except Exception:
            try:
                return model.swin.embeddings.patch_embeddings.projection
            except Exception:
                pass

    # 4) Generic fallback: last Conv2d found in the model
    last_conv = None
    for m in model.modules():
        if isinstance(m, torch.nn.Conv2d):
            last_conv = m
    if last_conv is not None:
        return last_conv

    # As a final fallback, just return the model (may not work with GradCAM, but avoids attribute errors)
    return model


def explain_gradient_shap(model, processor, image, n_samples=5):
    """
    Generate GradientSHAP explanations.
    """
    try:
        device = next(model.parameters()).device

        # Preprocess image
        inputs = processor(images=image, return_tensors="pt")
        input_tensor = inputs["pixel_values"].to(device)

        # Get prediction
        with torch.no_grad():
            outputs = model(input_tensor)
            predicted_class = outputs.logits.argmax(dim=1).item()

        # Create baseline (black image)
        baseline = torch.zeros_like(input_tensor)

        # Create wrapped model for Captum compatibility
        wrapped_model = ViTWrapper(model)

        # Initialize GradientSHAP with wrapped model
        gradient_shap = GradientShap(wrapped_model)

        # Generate attribution
        attribution = gradient_shap.attribute(
            input_tensor, baselines=baseline, n_samples=n_samples, target=predicted_class
        )

        # Summarize attribution across channels
        attribution = attribution.squeeze().sum(dim=0).cpu().detach().numpy()

        # Normalize
        if attribution.max() > attribution.min():
            attribution = (attribution - attribution.min()) / (
                attribution.max() - attribution.min()
            )
        else:
            attribution = np.zeros_like(attribution)

        # Create visualization
        fig, (ax1, ax2, ax3) = plt.subplots(1, 3, figsize=(15, 5))

        # Original image
        ax1.imshow(image)
        ax1.set_title("Original Image")
        ax1.axis("off")

        # SHAP attribution
        im = ax2.imshow(attribution, cmap="coolwarm")
        ax2.set_title("GradientSHAP Attribution")
        ax2.axis("off")
        plt.colorbar(im, ax=ax2)

        # Overlay
        ax3.imshow(image, alpha=0.7)
        im_overlay = ax3.imshow(attribution, cmap="coolwarm", alpha=0.5)
        ax3.set_title("Attribution Overlay")
        ax3.axis("off")
        plt.colorbar(im_overlay, ax=ax3)

        plt.tight_layout()
        return fig

    except Exception as e:
        print(f"Error in GradientSHAP: {str(e)}")
        fig, ax = plt.subplots(figsize=(8, 6))
        ax.text(
            0.5,
            0.5,
            f"GradientSHAP failed:\n{str(e)}",
            ha="center",
            va="center",
            transform=ax.transAxes,
            fontsize=10,
        )
        ax.set_title("GradientSHAP Error")
        return fig