0

Due to the nature of the layers in my network, I have to do three transpositions: 

Input to 1D convolution
1D Convolution to 2D Convolution Transpose
1D Convolution to Dense

This takes an enormous amount of time

batch_time=121.467 seconds

for a batch_size of 10.

A previous model, similar in size, without the transpositions, and on the same dataset took like 10 seconds for the first batch and 0.4 seconds per batch after that. After the first batch, my model above still takes over 4 seconds per batch.

I think I managed to narrow it down to the transpositions I do. Is there a way around this?

Qubix
  • 4,161
  • 7
  • 36
  • 73
  • 1
    add a minimal example in code. if it is python I'll certainly try and to take a look. If it isn't python still share your code because someone else will most likely be able to help. – Anton Codes Jun 01 '17 at 13:00
  • well, it's basically a 4D tensor in tesorflow where I want to replace tf.transpose(tensor, [0,2,1,3]) with a less computationally expensive operation. – Qubix Jun 02 '17 at 07:35

0 Answers0