Regularization defaults missing

[adambielski]

The models shared are great for feature extraction, but their use for deeper fine-tuning is limited due to lack of information regarding L2 kernel regularization and Dropout. What's the reason for not adding them to the model definition? They wouldn't change anything for inference, but would allow proper fine-tuning of earlier layers. There's also no information where the weights are taken from, so it's impossible to the modification on one's own. Fine-tuning the entire model with mismatching regularization (or no regularization) seems suboptimal. @fchollet is it possible to add original L2 reg and Dropout for model definition? If not, could at least the information about these values be available along with models?