0

I'm a newbie in keras tensorflow and python. Network structure :Resnet50(with imagenet weight, all layers freezed) + 2 Bi-LSTM with 256 units

Number of images : 624000

Image size : 224x224x3

Batch size : 150

GPU : 1080Ti single

Backend or API: keras/tensorflow

ETA: 2-3 hours per epoch

Does the time it took to finish an epoch makes sense? The total amount of trainable parameters is only around 4 millions since all the layers in resnet was freezed.

If it really does makes sense should I reduce the image size? Let's say I resized it to 112x66x3 then how do I set the mean of the training set to imagenet mean? Do I use the same preprocess_input for Resnet50 offered by Keras?

Miles High
  • 37
  • 5
  • Sounds like a lot of training time, on the other hand you have a lot of images and a complicated network. Maybe try checking if you actually using your GPU: https://stackoverflow.com/questions/44544766/how-do-i-check-if-keras-is-using-gpu-version-of-tensorflow – Tinu Nov 03 '19 at 18:32
  • Yes I have already checked that I'm using GPU. – Miles High Nov 03 '19 at 18:42

0 Answers0