CNN Visualization Techniques

# Feature Map Visualization import torch import torch.nn as nn import matplotlib.pyplot as plt import numpy as np class FeatureMapVisualizer: def __init__(self, model): self.model = model self.feature_maps = {} self.hooks = [] def register_hooks(self, layers_to_visualize): """Register forward hooks to capture feature maps""" def hook_fn(name): def hook(module, input, output): self.feature_maps[name] = output.detach() return hook for name, module in self.model.named_modules(): if isinstance(module, nn.Conv2d) and name in layers_to_visualize: handle = module.register_forward_hook(hook_fn(name)) self.hooks.append(handle) def visualize_feature_maps(self, input_tensor, layer_name, max_filters=16): """Visualize feature maps from a specific layer""" # Clear previous feature maps self.feature_maps.clear() # Forward pass with torch.no_grad(): _ = self.model(input_tensor) if layer_name not in self.feature_maps: print(f"Layer {layer_name} not found in captured feature maps") return feature_map = self.feature_maps[layer_name] # Select first image from batch and limit number of filters feature_map = feature_map[0] # Shape: [channels, height, width] num_filters = min(max_filters, feature_map.shape[0]) # Create subplot grid cols = 4 rows = (num_filters + cols - 1) // cols fig, axes = plt.subplots(rows, cols, figsize=(12, 3*rows)) if rows == 1: axes = axes.reshape(1, -1) for i in range(num_filters): row = i // cols col = i % cols # Normalize feature map for visualization fmap = feature_map[i].cpu().numpy() fmap = (fmap - fmap.min()) / (fmap.max() - fmap.min() + 1e-8) axes[row, col].imshow(fmap, cmap='viridis') axes[row, col].set_title(f'Filter {i}') axes[row, col].axis('off') # Hide unused subplots for i in range(num_filters, rows * cols): row = i // cols col = i % cols axes[row, col].axis('off') plt.suptitle(f'Feature Maps from {layer_name}') plt.tight_layout() plt.show() def compare_layers(self, input_tensor, layer_names): """Compare feature maps across different layers""" self.feature_maps.clear() with torch.no_grad(): _ = self.model(input_tensor) fig, axes = plt.subplots(1, len(layer_names), figsize=(5*len(layer_names), 5)) if len(layer_names) == 1: axes = [axes] for i, layer_name in enumerate(layer_names): if layer_name in self.feature_maps: # Take mean across all channels for overview feature_map = self.feature_maps[layer_name][0] # First image mean_activation = torch.mean(feature_map, dim=0).cpu().numpy() # Normalize mean_activation = (mean_activation - mean_activation.min()) / \ (mean_activation.max() - mean_activation.min() + 1e-8) axes[i].imshow(mean_activation, cmap='hot') axes[i].set_title(f'{layer_name}\n({feature_map.shape[0]} channels)') axes[i].axis('off') plt.suptitle('Feature Map Evolution Across Layers') plt.tight_layout() plt.show() def cleanup(self): """Remove all hooks""" for hook in self.hooks: hook.remove() self.hooks.clear() # Usage example model = torchvision.models.resnet50(pretrained=True) model.eval() visualizer = FeatureMapVisualizer(model) # Register hooks for specific layers layers_to_visualize = ['layer1.0.conv1', 'layer2.0.conv1', 'layer3.0.conv1', 'layer4.0.conv1'] visualizer.register_hooks(layers_to_visualize) # Load and preprocess image image = load_and_preprocess_image('path/to/image.jpg') # Your preprocessing function input_tensor = image.unsqueeze(0) # Add batch dimension # Visualize feature maps from different layers visualizer.visualize_feature_maps(input_tensor, 'layer1.0.conv1') visualizer.compare_layers(input_tensor, layers_to_visualize) # Cleanup visualizer.cleanup()