目标检测项目遇RuntimeError：输入通道数与卷积层预期不匹配

我正在开展一个目标检测项目，目前遇到如下错误：

RuntimeError: Given groups=1, weight of size [32, 3, 3, 3], expected input[32, 500, 800, 3] to have 3 channels, but got 500 channels instead

我认为引发该问题的代码段如下：

# cnn_model.py
import torch
import torch.nn as nn
import pdb

class CNNModel(nn.Module):
    def __init__(self, num_classes):
        super(CNNModel, self).__init__()
        self.conv1 = nn.Conv2d(3, 32, kernel_size=3, padding=1)  # Expects 3 input channels (RGB)
        self.pool = nn.MaxPool2d(2, 2)

        # No need to define expected_in_features here in the constructor

        # Adjust in_features based on the output size of your conv layers
        # Assuming your last conv layer output size is (32, 250, 400)
        self.fc1 = nn.Linear(32 * 250 * 400, 128)
        self.fc2 = nn.Linear(128, num_classes)

    def forward(self, x):
        #pdb.set_trace()
        print(f"Input shape to conv1: {x.shape}")
        x = self.pool(torch.relu(self.conv1(x)))

        # Calculate expected_in_features based on the output size of the previous layers
        expected_in_features = x.size(1) * x.size(2) * x.size(3)  # Calculate it based on the current shape

        x = x.view(-1, expected_in_features)

        x = torch.relu(self.fc1(x))
        x = self.fc2(x)
        return x

这个错误的核心是PyTorch张量维度顺序不匹配：

• 你的卷积层conv1遵循PyTorch标准格式，要求输入为(batch_size, channels, height, width)，但实际传入的张量是(32, 500, 800, 3)，也就是(batch_size, height, width, channels)（TensorFlow常用格式）。
• 卷积层误将第2个维度（500）识别为通道数，和权重定义的3输入通道不匹配，因此报错。

解决步骤 #

1. 调整输入维度顺序
   在传入模型前，把输入从(B, H, W, C)转成(B, C, H, W)，可以用permute方法实现。可以在模型的forward方法里处理，也可以在数据加载阶段完成。

修改后的forward方法示例：

def forward(self, x):
    # 转换维度顺序：从(B, H, W, C)转为(B, C, H, W)
    x = x.permute(0, 3, 1, 2)
    print(f"Input shape to conv1 after permute: {x.shape}")
    x = self.pool(torch.relu(self.conv1(x)))

    expected_in_features = x.size(1) * x.size(2) * x.size(3)
    # 动态修正全连接层输入维度，避免硬编码导致的尺寸不匹配
    if expected_in_features != self.fc1.in_features:
        self.fc1 = nn.Linear(expected_in_features, 128).to(x.device)
    
    x = x.view(-1, expected_in_features)
    x = torch.relu(self.fc1(x))
    x = self.fc2(x)
    return x

2. 数据加载阶段预处理（推荐）
   如果在Dataset的__getitem__方法中处理维度转换，能避免在模型里额外操作：

def __getitem__(self, idx):
    image = self.load_image(idx)  # 假设加载后是(H, W, C)格式
    image_tensor = torch.tensor(image).permute(2, 0, 1)  # 转为(C, H, W)
    return image_tensor, self.labels[idx]

内容的提问来源于stack exchange，提问作者Ihsan Seker