CNN卷积层可视化疑问：滤波器维度转换与归一化方式探讨

1. RGB卷积滤波器维度置换的正确性 #

你的做法完全正确。PyTorch中卷积层权重的默认维度是(out_channels, in_channels, height, width)，针对输入为RGB图像的第一层卷积，in_channels=3对应RGB三通道。而Matplotlib的imshow函数要求RGB图像维度为(height, width, channels)，因此通过permute(1, 2, 0)将filters[i]的维度从(in_ch, h, w)转换为(h, w, in_ch)，完全匹配图像可视化的格式要求，能正确呈现RGB滤波器的样式。

2. 归一化方式的选择：全局vs单滤波器独立归一化 #

两种方式各有适用场景，取决于你想观察的核心信息：

• 全局归一化（当前代码的做法）：用整个权重张量的最小/最大值统一归一化所有滤波器。这种方式会保留不同滤波器间的权重幅度差异——权重整体偏大的滤波器可视化后更亮，偏小的则更暗。但如果不同滤波器的权重范围差异悬殊，部分滤波器的细节可能被压缩，导致模式辨识度降低。
• 单滤波器独立归一化：对每个滤波器单独计算最小/最大值后映射到[0,1]。这种方式能最大化单个滤波器的对比度，让你清晰看到每个滤波器的权重分布模式，但会丢失滤波器间的权重幅度信息，无法直观对比响应强度的高低。

若要切换到独立归一化，可修改代码如下：

# 替换原归一化代码段
filters = []
for weight in weights:
    w_min, w_max = weight.min(), weight.max()
    normalized = (weight - w_min) / (w_max - w_min)
    filters.append(normalized)
filters = torch.stack(filters)

原可视化代码 #

# Extract weights from the first layer
weights = model.cn1.weight.detach().cpu()

# Normalize weights to [0, 1] for visualization
weights_min, weights_max = weights.min(), weights.max()
filters = (weights - weights_min) / (weights_max - weights_min)

fig, axes = plt.subplots(4, 8, figsize=(12, 6))
fig.suptitle('Filters of the First Convolutional Layer (cn1)', fontsize=16)

for i in range(32):
    ax = axes[i // 8, i % 8]
    # Weights are (out_ch, in_ch, h, w). We permute to (h, w, in_ch) for plotting
    filter_img = filters[i].permute(1, 2, 0).numpy()
    ax.imshow(filter_img)
    ax.axis('off')

plt.tight_layout(rect=[0, 0.03, 1, 0.95])
plt.show()


def get_feature_maps(model, x):
    maps = []

    # Block 1
    x = F.relu(model.bn1(model.cn1(x)))
    x = F.max_pool2d(x, 2)
    maps.append(x.detach().cpu())

    # Block 2
    x = F.relu(model.bn2(model.cn2(x)))
    x = F.max_pool2d(x, 2)
    maps.append(x.detach().cpu())

    # Block 3
    x = F.relu(model.bn3(model.cn3(x)))
    x = F.max_pool2d(x, 2)
    maps.append(x.detach().cpu())

    return maps


test_img, test_label = next(iter(test_dl))
sample_input = test_img[0:1].to(device) # Shape (1, 3, 32, 32)

feature_maps = get_feature_maps(model, sample_input)

fig, axes = plt.subplots(3, 8, figsize=(15, 6))
for block_idx, fmap in enumerate(feature_maps):
    for i in range(8):
        ax = axes[block_idx, i]

        ax.imshow(fmap[0, i], cmap='viridis')
        ax.axis('off')
        if i == 0:
            ax.set_ylabel(f'Block {block_idx+1}', size='large')

plt.suptitle('8 Channels of Feature Maps from each Conv Block', fontsize=16)
plt.tight_layout()
plt.show()

内容的提问来源于stack exchange，提问作者Sandesh Singh