Passing the output from the last Convolutional layer to the FCC layer - PyTorch

Question

I am trying to pass the output from the last Convolutional Layer to FCC layer but I am struggling with dimentions. As a default, the network uses AdaptiveAvgPool2d(output_size=(6, 6)) what does not let me use torch.use_deterministic_algorithms(True) for reproducibility purpose. This is the error I am getting:

*mat1 dim 1 must match mat2 dim 0*

    (10): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (11): ReLU(inplace=True)
    (12): MaxPool2d(kernel_size=3, stride=2, padding=0, dilation=1, ceil_mode=False)
  )
  (avgpool): AdaptiveAvgPool2d(output_size=(6, 6))
  (classifier): Sequential(
    (0): Dropout(p=0.5, inplace=False)
    (1): Linear(in_features=9216, out_features=4096, bias=True)

The input tensor is: [10, 3, 350, 350]. The shape of tensor from last Conv2d/MaxPool2d layer is: torch.Size([10, 256, 9, 9]). I assume that number of inputs for FCC should be 256 x 9 x 9 = 20736 but it does not work as well.

Here is also my class for forwarding the output from CONV to FCC layer:

class Identity(nn.Module):
    def __init__(self):
        super(Identity, self).__init__()

    def forward(self, x):
        print('SHAPE', np.shape(x))
        return x

The idea has been taken from video: https://www.youtube.com/watch?v=qaDe0qQZ5AQ&t=301s. Thank you so much in advance.

Ivan Ivan · Accepted Answer · 2021-08-17T21:18:49

TLDR; The number of neurons in your fully connected layer is fine, your shape is not.

The nn.AdaptativeAveragePool2d layer between your CNN and classifier will output a tensor of shape (10, 256, 6, 6). since you've initialized it with an output_size of (6, 6). This being said, the first fully connected layer should have 256*6*6 neurons.
```
self.fc = nn.Linear(in_features=9216, out_features=4096)
```

This matches your current model's setup, not your proposed 20736...

Your classifier input shape should be flattened, this can be done by defining a flatten layer nn.Flatten (or using an inline alternative). First define your layer in the initializer:
```
self.flatten = nn.Flatten()
```
Then
```
>>> x.shape # nn.AdaptativeAveragePool2d output
torch.Size([10, 256, 6, 6])

>>> self.flatten(x)
torch.Size([10, 9216])
```

Passing the output from the last Convolutional layer to the FCC layer - PyTorch

1 Answers