'train' and 'val' losses in the json file

[sakerlrf]

Printing Losses

Hello,
I am trying to understand the guidelines about hooks. In the Eval_hook, we see the rule about greater and less_keys, and we see the 'loss' in the latter one.

In the main model script, eg. I3D config, I want to state that I need train and val losses to be generated, and added in the json file. Eg we can use the following for the evaluation metrics:
evaluation = dict(interval=5, metrics=['top_k_accuracy', 'mean_class_accuracy'])

My question is:
If one needs to have the losses added in the json file, how will this be done?

Thank you in advance.

[dreamerlin]

Try --cfg-options workflow="[(train,1),(val,1)]", you can get the loss in json file.

But we still recommend to use --validate

[sakerlrf]

Much appreciated!
Thank you.

[caiocvsilva]

Hello,

The option to run the code with --validate does not exist anymore? I am trying to get the train accuracy to be printed, as well the validation loss. But most of the answers here and in the documentation seem to relate to an older version of the code.

Any idea how to do it?

Thank you.