NOTES.md

Conversion notes

Keras 2.0.6 weight shapes

I suspect the shape of weights can change from release to release, since it broke on the upgrade from 1.2.X to 2.0.Y. The following shapes are known to be valid for 2.0.6. Bias shape is only mentioned when not obvious.

Layer	How shape is calculated
Conv{1,2,3}D	kernel_size + (input_dim, filters). That means height*width*in*out for Conv2D.
Dense	(input_dim, units)

Convolution layer output sizes: Caffe vs. Keras

Here is Caffe's code for computing a convolution layer's output shape. It does something like this for each output dimension (independently):

dilated_extent = dilation * (kernel_shape - 1) + 1
# Original was C++; just using // to show that it rounds down
out_shape = (input_shape + 2 * pad - dilated_extent) // stride + 1

In contrast, this is what Keras does to each axis:

# real dilated_extent is kenel_shape + (kernel_shape - 1) * (dilation - 1),
# this is equivalent
dilated_extent = dilation * (kernel_shape - 1) + 1
if padding_mode in {'same', 'causal'}:
    padded_shape = input_shape
elif padding_mode == 'valid':
    padded_shape = input_shape - dilated_extent + 1
elif padding_mode == 'full':
    padded_shape = input_shape + dilated_extent - 1
out_shape = (padded_shape + stride - 1) // stride

Two important differences in the two implementations: firstly, padding is handled by specification of a padding "mode" instead of a number of pixels to pad by. Secondly, the default mode is 'valid', in which case the code above can be simplified to:

dilated_extent = dilation * (kernel_shape - 1) + 1
out_shape = (input_shape - dilated_extent + stride) // stride

As far as I can tell, the Keras' code for 'valid'`` mode is equivalent to Caffe's code when pad = 0.