3 ms·

The input sizes in the diagram appear to be wrong: https://github.com/nihit/TensorFlow101/blob/master/fig/cnn_architecture.png https://github.com/nihit/TensorFl

by p1esk 10y ago

The input sizes in the diagram appear to be wrong: https://github.com/nihit/TensorFlow101/blob/master/fig/cnn_architecture.png https://github.com/nihit/TensorFlow101/blob/master/fig/cnn_a... It should be 24x24 in the conv. layer, not 28x28, and so on.

nihitd 10y ago
Actually, spatial dimensions are maintained across convolution layers by adding appropriate padding - only the pooling layers change spatial dimensions. I believe "padding = SAME" argument takes care of this in Tensorflow(?)
- p1esk 10y ago
  Why would you want to maintain spatial dimensions? Padding here does not make any sense.