5x5 kernel size instead of 3x3?

[sudo-bcli]

Hi, I just read this tutorial on your official website NEURAL NETWORKS
and think according to the image and the following code, maybe the kernel size of the first convolution layer should be 5x5 instead of 3x3.
If we follow this formula and by default the argument of conv2d is padding = 0 and stride = 1, we have

• 1st conv2d with 5x5 kernel: 32x32 -> 28x28
• 1st max pooling: 28x28 -> 14x14
• 2nd conv2d with 3x3 kernel: 14x14 -> 12x12
• 2nd max pooling: 12x12 -> 6x6

Which will explain both the image and the following linear layer (6x6 image dimension) in your code.

[holly1238]

Kernel size is 5x5. This was fixed in the tutorial in #1257.