Negative train and validation loss as training proceeds about ranking HOT 4 CLOSED

The loss in approx_ndcg_loss is the negative of the approximate NDCG, so it should be negative by design.

from ranking.

This is weird. Can you make sure that you don't have negative labels in your data? Also, did you try other loss functions and observe negative loss as well?

from ranking.

There are no negative labels in the data.

For sigmoid_cross_entropy_loss, loss is negative only when relevance is > 1. For relevance = 1, loss converges nicely and is >0 all the time.

Yes, loss is also negative for approx_ndcg_loss.

I think there might be an issue with how the loss is calculated from the output of score_fn. Please review and help in solving this.

from ranking.

Based on the documentation here, only 0 and 1 are valid for labels of sigmoid_cross_entropy_loss: https://cs.corp.google.com/piper///depot/google3/third_party/tensorflow/python/ops/losses/losses_impl.py?q=sigmoid_cross_entropy_loss&g=0&l=655.

How about softmax_loss? If that's ok by now, you can use that.

from ranking.