Hey Alireza,

In case you haven't seen this:

https://stackoverflow.com/questions/60836958/why-is-separableconv2d-slower-than-conv2d

Thank You Dear Toygar Akgün

I did some little search before asking questions and saw this link. However, my problem has not yet been solved. Convolutional Autoencoder by Conv2D has 6000 parameters and by SeparableConv2D has 1000 parameters.

I am using Keras 2.3.1 and Tensorflow 2.1 and the GPU driver (1060Ti) has updated.

Based on theory SeparableConv2D have to faster speed (because of fewer parameters and FLOPs) but when I applied them is almost twice slower.

Could you profile your code to see if separable API call indeed goes to GPU?

I suspect that it does not execute on GPU.

This "might" be due to installing a prebuilt binary.

Thank you Toygar Akgün

I found this article:

Creating depthwise separable convolutions with TensorFlow 2 and Keras – MachineCurve

in last section mentioned that this seems to be some issue with TensorFlow.