(Input Output dim size question) A pass through Conv2d convolves over multiple batches instead of on a single batch

nlp
1

Hope you all are havign a fine day.

I have been experimenting with using BERT output embeddings as inputs to a CNN for text classification.

I have set the max number of tokenizers to 256, and batch size is 4. Bert output layer gives a 1d tensor with 768 values.

(Essentially, I am just stacking (1x768) embedding vectors 256 times for each of the 256 tokens in a single sentence that will be used for classification. Each batch has 4 sentences )
Perhaps the following image is going to get the point across:

1*-vDcDVuM2DoDnA7BsDONCQ
1*-vDcDVuM2DoDnA7BsDONCQ612×615 40.1 KB

The following are the dimensions of a single batch:

IN

batch = torch.stack([torch.unsqueeze(torch.rand((256,768)),dim=0) for i in range(4)])

OUT

torch.Size([4, 1, 256, 768])
#a single batch with 4 samples, 1 input channel, 256 tokens with 768 rowed embedding tensors

The following is just a snippet of my model and forward method:

class EmbeddingCNN(nn.Module):
    def __init__(self):
        super(EmbeddingCNN,self).__init__()
        self.conv1 = nn.Conv2d(
            1,
            3, 
            (5, 768))
        
        self.pool1 = nn.MaxPool2d(kernel_size=(4,))
        
    def forward(self,embedding):
        
        conv1out = self.conv1(embedding).squeeze(3)
        print(conv1out.shape)
        
        return conv1out

Instead of convolving with square kernels like in image classification, in the case of text, I am convolving with a 5x768 kernel (which is essentially going from top to bottom if 256 was the height and 768 was the width. The intuition is to apply filters over 5 consecutive words in a sentence) There are 3 output channels in first conv2d layer.

This results in the following dimensions

#conv1out.shape
[1,3,252,1]

Here we come to my question. Why isn’t the length of the first dimension 4 (num of batches)? Why has it reduced to 1? I would assume this convolution assume the following are the meanings of the dimension sizes

1 : ???
3 : Output channel count
4 : "Height" of the convolved layer
5 : "Width" of the convolved layer.

Is this understanding correct? If not, what do the dimensions mean?

2

Here the three input params mean: (in_channels, out_channels, kernel_size). Since your in_channels=1,
conv1out.shape = [1,3,252,1] is expected.

3

Thanks for your input.

My input to the Conv2d layer was torch.Size([4, 1, 256, 768])
where 4 is the number of samples in a single batch. Why isn’t the output: [4,3,252,1] ?
(4 samples, 3 output channels, 252 height, 1 width>)