CODEWITHSUNDEEP
tensorflowdeep-learningcomputer-visionpytorch
bellawillrise
bellawillrise

Reputation: 85

Formula to compute the padding in convolutions pytorch (googlenet)

I am implementing googlenet (smaller version) from scratch in pytorch. The architecture is below:

enter image description here

For the Downsample Module I have the following code:

   class DownSampleModule(nn.Module):
   def __init__(self, in_channel, ch3, w):
       super(DownSampleModule, self).__init__()
       kernel_size = 3
       padding = (kernel_size-1)/2

       self.branch1 = nn.Sequential(
           ConvBlock(in_channel, ch3, kernel_size = 3,stride=2, padding=int(padding))
       )
       self.branch2 = nn.Sequential(
           nn.MaxPool2d(3, stride=2, padding=0, ceil_mode=True)
       )
   def forward(self, x):
       branch1 = self.branch1(x)
       branch2 = self.branch2(x)
      
       return torch.cat([padded_tensor, branch2], 1)

The ConvBlock is from this module

class ConvBlock(nn.Module):
   def __init__(self, in_channels, out_channels, kernel_size, stride, padding):
       super(ConvBlock, self).__init__()
       #padding = (kernel_size -1 )/2
       #print(padding)
       self.conv = nn.Conv2d(in_channels, out_channels, kernel_size, stride, padding)
       self.bn = nn.BatchNorm2d(out_channels)
       self.act = nn.ReLU()
       
   def forward(self, x):
       x = self.conv(x)
       x = self.bn(x)
       x = self.act(x)
       return x

Basically, we are creating two branches: a convolution module and a max pool. The output of these two branches are then concatenated on the channel dimension.

However, I have the following problem:

Downsample Convolution:torch.Size([1, 80, 15, 15])
Maxpool Convolution:torch.Size([1, 80, 15, 15])

Downsample Convolution:torch.Size([1, 96, 8, 8])
Maxpool Convolution:torch.Size([1, 144, 7, 7])

Does anyone know the formula to compute the correct padding? Thank you.

In Keras, you can just set the padding="same" or "valid" but it's not supported on pytorch.

Upvotes: 0

Views: 203

Answers (1)

trialNerror
trialNerror

Reputation: 3553

Your maxpool and conv branches have the same input, and will produce identically-shaped output if you give them the same parameters for kernel size, stride and padding. So just replacing your padding = 0 with padding = int(padding) should be enough to make both branches be concat-compatible.

ceil_modeshould also be set to False. When the resulting dimension is not an integer, the rounding behavior of conv2d is to use floor, so you want your maxpool to do that as well.

By the way, you can remove your nn.Sequential. Your "sequences" of layers are made of only one layer, so... not really sequential :)

Upvotes: 1

Related Questions