About the exercise

[Puff-Wen]

Hi,

I added a conv as following

    self.conv1 = nn.Conv2d(1, 6, kernel_size=(5,5),stride=1, padding=0)
    self.conv1_5 = nn.Conv2d(6, 12, kernel_size=(5,5),stride=1, padding=1)
    self.conv2 = nn.Conv2d(12, 16, kernel_size=(5,5),stride=1, padding=0)

and

    out = F.relu(self.conv1_5(out))

in the forward function. Then I got the error message as following.

RuntimeError: Given input size: (16x1x1). Calculated output size: (16x0x0). Output size is too small at /pytorch/torch/lib/THCUNN/generic/SpatialDilatedMaxPooling.cu:69

Would you please help to give some clues?
Thanks.

[hui-po-wang]

Hi,

I try to reproduce the error, but it seems to work well.
I am not sure what happened to your codes, it's helpful if you can provide more information.

The attachment is my codes, you can print out each size of the outputs to check the output size as I write in the forward function.
test.txt

Thanks.

[Puff-Wen]

Hi Hui-Po,

Thanks for your prompt reply. Two questions are as following.

1. I added print in my forward function as following.

   def forward(self, x):
   out = F.relu(self.conv1(x))
   print(out.size())
   out = F.max_pool2d(out, 2)
   print(out.size())
   out = F.relu(self.conv1_5(out))
   print(out.size())
   out = F.max_pool2d(out, 2, stride=1, padding=1)
   print(out.size())
   out = F.relu(self.conv2(out))
   print(out.size())
   out = F.max_pool2d(out, 2)
   print(out.size())
   out = out.view(out.size(0), -1) #flatten
   print(out.size())
   out = F.relu(self.fc1(out))
   print(out.size())
   out = F.relu(self.fc2(out))
   print(out.size())
   out = self.fc3(out)
   print(out.size())
   return out

And the output is

(128L, 6L, 24L, 24L)
(128L, 6L, 12L, 12L)
(128L, 12L, 10L, 10L)
(128L, 12L, 11L, 11L)
(128L, 16L, 7L, 7L)
(128L, 16L, 3L, 3L)
(128L, 144L)
RuntimeError: size mismatch at /pytorch/torch/lib/THC/generic/THCTensorMathBlas.cu:243

2. I found the fc1 in your code is updated from
   self.fc1 = nn.Linear(1644, 120)
   to
   self.fc1 = nn.Linear(1633, 120)

Would you please help to explain about the modification? Thanks

[JiaRenChang]

Hi,
I believe that the input dimension of your fc1 layer is wrong.
Now the output dimension before fc layer is 128(batch size) x 144 (feature dimension).
So the input of your first fc layer must be 144.

[hui-po-wang]

Hi,

The main problem here comes from the new convolution layer you added. The output size of self.conv1_5 is not the same as the input size of self.conv1_5 .
To keep the same size, the padding should be floor (kernel_size/2), that is, 2. Otherwise, you will find that your output size becomes (128L, 16, 3, 3) rather than (128, 16, 4, 4). As a result, you need to modify your self.fc1.

[Puff-Wen]

Hi all,

Thanks. It is resolved.