Save perplexity on training and validation sets about nmt-keras

[lvapeab]

If you set perplexity as a keras metric (KERAS_METRICS = ['perplexity']), it should be logged in Tensorboard (by default, in path-to-model/tensorboard_logs).

[philipcori]

Got it, thanks.

[philipcori]

Sorry, even though I have tensorboard working, I only see loss and perplexity on training data now in tensorboard. I set params['EVAL_ON_SETS_KERAS'] = ['val', 'train'], however this doesn't seem to change anything.

[lvapeab]

Try updating MKW (pip install -U multimodal-keras-wrapper) and setting METRICS= ['perpleity'].

The parameter KERAS_METRICS is used for logging of the training, while METRICS evaluates the model predictions (typically on the val set). The EVAL_ON_SETS parameters refers to METRICS.

This is rather convoluted and confusing. Will try to clarify it in a future.

[philipcori]

Ok it works now after updating multi-modal-keras. Though I had to pip install directly from the repo, for some reason pip install -U multimodal-keras-wrapper did not successfully update.