It seems "top1" loss function do not work in this tf implementation?

[ownership-xyz]

It seems "top1" loss function do not work in this tf implementation?
I follow the parameters setting like that it theano implementation. However, I just get a result of 0.48 and 0.17. (Original is 0.59 and 0.23)

I use tensorflow 1.2.

[Songweiping]

Hi @IcyLiGit ,
Sorry about that you can't reproduce the results. First, I didn't test the code with "top1" loss, thus I have no idea how to set "good" parameters, either. Second, did you get the numbers 0.59 and 0.23 by running theano implementation?
BTW, I use Adam optimizer defaultly and you may try RMSProp(default optimizer in theano implementation) or others. Tuning initial learning rate and dropout rate would also help.
Good Luck!

Weiping

[ownership-xyz]

@Songweiping
I had ran the original theano code and it produced the right result as that in the paper.Actually, in theano code, the optimizer is adagrad. However, I found adagrad and adadelta do not work in tensorflow implementation.

RMSprop and adam with cross-entropy loss and softmax activation function may work in your implementation. However, top1 and bpr can just produce a result of 0.48 (not 0.6 in paper) , and it seems loss value decrease quicker in tf. (Maybe caused by overfitting? But I can not find the differences of these two implementation....)

[ownership-xyz]

I use the same parameter settings in these two implementations. (Softmax + cross-entropy + 0.5 drop + 0.001 lr without decay). However, the losses reported are different.

theano:

tensorflow:

[Songweiping]

It seems that TF converges faster than Theano. So how about:

1. decease training steps.
2. more concretely, use validation data to prevent over-fitting(early stop).

Weiping

[gds123]

I find the similar issue too
I add dynamic_rnn to Weiping's code, the recall then drop to 0.43 for Softmax + cross-entropy
the recall is 0.43 for top1

and it's not overfitting, I have check the recall on training data
@Songweiping
@IcyLiGit