Question about resnet implementation

[harshita1804]

Question about resnet implementation

[suvoooo]

I think if you read the medium post along with the paper, you will see that the convolution block is acting like pooling layer because of the stride=2. In the Resnet-34 architecture you will see layers like 3x3 conv 128/2 (this is conv block) and then many 3x3 conv 128 layers (these are identity blocks). We calculate features via both identity and conv block but in conv block we reduce the input dimension. So we start a stage with convolution block and then stack the identity blocks.