How to count the number of multiplications given a keras model?

Hi,

I am wondering how to count the number of multiplications given a keras model, which is well defined. Is there any off-the-shelf function or codes?

I know obtaining the number of parameters given a keras model is to use model.summary() .

[jerinkantony]

Is there any solution for this?

[busyBeaver1]

I had the same problem. Here’s my solution. It turns out that for a dense layer the number of multiplications should be the same as the number of weights. But this does not work as well for the convolution layers. For a convolution layer you need to multiply the number of weights by the number of instances of the kernel applied on the input to get the total number of multiplications. If my assumptions are right, then this should work for dense, conv and separable conv layers:

import numpy as np

/# here you need to create the model

print(‘\n’.join([ 'layer: ’ + type(l).name +
', output shape: ’ + str(l.output_shape[1:]) +
', weights: ’ + str(sum([i.size for i in l.get_weights()])) +
', multiplications: ’ + str(sum([i.size for i in l.get_weights()]) * int(np.prod(l.output_shape[1: -1])))
for l in model.layers ]))

The output for my CNN (note that you should, at least temporarily, define the input shape):

layer: SeparableConv2D, output shape: (240, 240, 128), weights: 400, multiplications: 23040000 layer: BatchNormalization, output shape: (240, 240, 128), weights: 512, multiplications: 29491200 layer: SeparableConv2D, output shape: (80, 80, 256), weights: 51456, multiplications: 329318400 layer: BatchNormalization, output shape: (80, 80, 256), weights: 1024, multiplications: 6553600 layer: SeparableConv2D, output shape: (40, 40, 256), weights: 90368, multiplications: 144588800 layer: BatchNormalization, output shape: (40, 40, 256), weights: 1024, multiplications: 1638400 layer: Dropout, output shape: (40, 40, 256), weights: 0, multiplications: 0 layer: SeparableConv2D, output shape: (40, 40, 256), weights: 84224, multiplications: 134758400 layer: Dropout, output shape: (40, 40, 256), weights: 0, multiplications: 0 layer: Conv2D, output shape: (40, 40, 5), weights: 92165, multiplications: 147464000

It does not actually work for BatchNormalization layers, so the output for those layers probably is meaningless.

P.S. idk how to add code properly