chiphuyen/stanford-tensorflow-tutorials

tensorflow sampled softmax loss

adrienmialland opened this issue · 0 comments

I am using tf.sampled_softmax_loss in a RNN for translation from english to french. This is usefull to speed up the training but, is it possible to get the prediction of the samples used by tf.sampled_softmax_loss ?

I'd like to be able to compute the accuracy.

Indeed, what I understand is that tf.sampled_softmax_loss acte like a final layer that output all the logits for all the required class, and then apply the sampled algrithm to proccess a "smaller" softmax. But, I can't directly access, at least, the logits.