Has anyone ever trained a deep model on ImageNet from scratch using keras (Tensorflow backend)? I have tried several deep nets, including ResNet50. The top-1 accuracy I got is way lower than the expected 75%. The best I get so far is below 70% at 60 epochs ... I could achieve much better results using caffe. I would appreciate it if you could share your experience in training on ImageNet from scratch using keras.
Asked
Active
Viewed 79 times
