0

I'm doing a benchmark of keras pre-trained models (vgg,resnet,inception,...) for image classification on personal data (electrical equipments), and I was wondering if there are best practices to have a relevant benchmark. I have 120 labeled images. I already tried data augmentation, checkpoints, early stopping,

Best pratices about :

  • getting reproducible results : every time I train my model, I have different results. I tried the tips in this post in vain : How to get reproducible results in keras

  • fully-connected layers : do we need to have complex FC layers ? What are the best pratices in transfer learning?

  • other tips?

Thank you in advance !

Louis Fontaine
  • 1
  • 1
  • Weight initialization is not the only thing you need to worry about to get exact results. You need to use the same optimizer, batch size, learning rate. You need to provide the same images in a fixed order with fixed augmentations. – Susmit Agrawal Mar 31 '20 at 16:31
  • @SusmitAgrawal thanks for your answer. All these parameters were fixed. Even for my ImageDataGenerator, i fixed a seed and disabled shuffling. – Louis Fontaine Mar 31 '20 at 18:03

0 Answers0