There are two parts in my loss. One is the dssm loss and the other is the reg loss. However even the dssm loss is much bigger than the reg loss, it does not even change while the reg loss is slowly decressing. Has anyone encountered the same problem?

how to discrimate the positive & negtive example in the code dssm_v2.py?
i don't see any label.

Hi @LiahA, thanks for sharing the code for your implementation. It is good to compare performance.
Can you share the data used for training your model? Can it be generated using another script? I did see that you've added an ignore to the data folder in your .gitignore file. Did you use the MS Marco dataset as done in the paper?
Thanks.

where to download data used in dssm?

When running dssm/single/dssm_v3.py, InvalidArgumentError happens at SparseTensorDenseMatMul with the info "Cannot multiply A and B because inner dimension does not match". I check the code and find that maybe the shape for SparseTensorValue should be np.array([BS, TRIGRAM_D], np.int64) not np.array([BS, TRIGRAM_D], np.int64). Does anyone has the same problems?

why my train loss ,after 4or5 epoch ,softmax value equal nan。

In dssm.py
temp = tf.tile(doc_y, [1, 1])

I think temp is equal to doc_y
is tile useful?

can you explain each file?
And i hope that now you can provide data sources.

I have seen your demo dssm/single/dssm_v3.py, and want to know how your data be organazed. For example, what the format of query.train.pickle ?

Hi, I see that in your code both train and test loss are the same, which is: 1. computed prob of the positive sample using softmax function, 2. compute its logloss against label (always 1).

My question is, the first step depends on randomly sampled negative samples, which makes the losses jumps during my training. I'm curious if you have tried to compute logloss using positive sample's logit only (not depend on negative samples)?

when i train this my loss fast equal 0 and auc=1？