Issue description

Convolutions do not raise an exception for too small input shape if stride > 1.

Code example

Suppose we have a following convolution:

conv = torch.nn.Conv2d(1, 1, kernel_size=3, dilation=2, stride=1)
tensor = torch.empty(1, 1, 4, 4)
conv(tensor).shape

This code raises a RuntimeError: Given input size per channel: (4 x 4). Calculated output size per channel: (0 x 0). Which is quite expected.

Now change the stride to 2:

conv = torch.nn.Conv2d(1, 1, kernel_size=3, dilation=2, stride=2)
tensor = torch.empty(1, 1, 4, 4)
conv(tensor).shape

Which gives torch.Size([1, 1, 1, 1]). Is this expected behavior?

I guess this is because in the formula for the output shape there is a division by the stride and the check is performed before taking the floor operation.

System Info

• How you installed PyTorch: conda
• OS: Linux
• PyTorch version: 0.4
• Python version: 3.6

The code was run on CPU.