PyTorch 0.4.0 Windows环境下CNN stride报错问题求助

Hey there! Let's break down the stride-related error you're facing with your CNN code. First, let's look at the most likely causes and how to fix them step by step.

Common Root Causes & Fixes #

1. Mismatch Between Feature Map Size and Fully Connected Layer Input #

Your current fully connected layer self.dense1 = nn.Linear(32 * 28 * 24, 60) expects a flattened feature map of size 32*28*24, but this depends entirely on the size of your input tensor and how the convolutions modify it.

With your current conv layers (no padding, stride=1, kernel size (1,3)):

• For an input tensor of shape (batch_size, 1, H, W):
  • After conv1, the shape becomes (batch_size, 16, H, W-2) (since W - 3 + 1 = W-2; H stays the same because kernel height is 1)
  • After conv2, the shape becomes (batch_size, 32, H, W-4)

Your code assumes H=28 and W-4=24 (so original W=28). If your input doesn't match (batch_size,1,28,28), you'll get a dimension mismatch error tied to stride/shape calculations.

Fix Steps:

• First, add print statements in your forward method to check tensor shapes at each step:

  def forward(self, input):
      print("Input shape:", input.size())
      x = F.relu(self.conv1(input))
      print("After conv1:", x.size())
      x = F.relu(self.conv2(x))
      print("After conv2:", x.size())
      x = x.view(x.size(0), -1)
      # ... rest of your forward logic
  

• Once you see the actual shape after conv layers, either:
  • Adjust your input tensor to match the expected (batch_size,1,28,28) shape, or
  • Update the dense1 layer's input dimension to match the flattened size you see in the print output.

2. Missing Padding Causing Unexpected Dimension Shrinkage #

If you want your convolutions to preserve the input width (instead of shrinking it by 2 each time), you can add padding to your conv layers. For a kernel size of (1,3), adding padding=(0,1) will keep the width unchanged:

self.conv1 = nn.Conv2d(in_channels=1, out_channels=16, kernel_size=(1,3), stride=1, padding=(0,1))
self.conv2 = nn.Conv2d(in_channels=16, out_channels=32, kernel_size=(1,3), stride=1, padding=(0,1))

This way, if your input is (batch_size,1,28,28), both conv layers will output (batch_size,16,28,28) and (batch_size,32,28,28) respectively—you'll just need to update dense1 to nn.Linear(32*28*28, 60).

3. Input Tensor Dimension Issues #

Make sure your input is a 4D tensor in the format (batch_size, channels, height, width). If you're passing a 3D tensor (e.g., (1,28,28) without a batch dimension), PyTorch will throw an error related to stride/input shape mismatch. Fix this by adding a batch dimension with input.unsqueeze(0) before passing it to the model.

4. Use Adaptive Pooling to Avoid Manual Dimension Calculation #

If you want your model to handle varying input sizes without manual shape adjustments, add an adaptive pooling layer after the convolutions to fix the feature map size:

def forward(self, input):
    x = F.relu(self.conv1(input))
    x = F.relu(self.conv2(x))
    # Fix feature map to (28,24) regardless of input size
    x = F.adaptive_avg_pool2d(x, (28, 24))
    x = x.view(x.size(0), -1)
    x = F.relu(self.dense1(x))
    x = self.out(x)
    return x

This ensures the flattened tensor always matches the 32*28*24 input expected by dense1.

内容的提问来源于stack exchange，提问作者Ddj