search

keras-team / keras-applications

Reference implementations of popular deep learning models.
Other
2k stars 910 forks source link

VGG16/19 models do not train #93

Closed htrenquier closed 5 years ago

htrenquier commented 5 years ago

I am trying to train VGG16 and VGG19 on cifar10. However, it seems that the models are not training. The validation accuracy stagnates at 0.1 (1/10: 10 classes for the cifar10 dataset).

In the code, (line 82 for vgg19 and vgg16) it seems that the line global backend, layers, models, keras_utils is missing. In comparison to other implemented models, this line is written after the documentation/description of the definition of the model.

Keras-Applications version is 1.0.7

Thank you

taehoonlee commented 5 years ago

@htrenquier, I hope the official example will help you.

htrenquier commented 5 years ago

@taehoonlee thank you for answering. However, this example is about ResNet and I don't encounter any issue when training this model. I have written a simple example: https://github.com/htrenquier/ml-attribute-based-testing/blob/master/example.py Maybe it will be easier for you to see my problem.

taehoonlee commented 5 years ago

@htrenquier, The convergence rate depends on many factors, such as model complexity, data distribution, and hyperparameters of an optimizer. You can try a more lighter model that VGG, or data augmentation in the official example. VGG in keras-applications has no problem.