yhs968 / pygru4rec Goto Github PK

It seems the recall degrade in this way
thank you

hidden = reset_hidden(hidden, mask).detach()

1. The code in the model.py, makes that loss of time t can only propogate to time t-1, which make the model more like a Markov Chain Model rather than a RNN modle?
2. Supposing a session of m timesteps and an usual RNN model, loss of time t can propogate to all timesteps <=t-1;
3. On this point, The pyGRU4REC is different from the theano version of GRU4REC?

Hi, Yhs, I am interested in your code. But I do not know where can I obtain the training data and test data. Help please.

Training GRU4REC...
epoch: 1/loss:5.118/recall:0.502/mrr:0.221/time:22.532
epoch: 2/loss:4.908/recall:0.599/mrr:0.267/time:22.467
epoch: 3/loss:4.867/recall:0.613/mrr:0.274/time:22.493
epoch: 4/loss:4.848/recall:0.619/mrr:0.276/time:22.521
epoch: 5/loss:4.836/recall:0.622/mrr:0.277/time:22.520
Test result: loss:5.878/recall:0.177/mrr:0.074/time:0.029

the preprocess is very slow

mask = [] # indicator for the sessions to be terminated  
        finished = False          
        while not finished:  
            minlen = (end - start).min()

Here, mask = [] should be within the while loop while not finished: ?

Dear Younghun Song,
First of all, I really thank you for your useful implementation. It is simpler than original code except a point. In function BPRLoss, I don't understand why the shape of logit is BxB. I think that it should be B x n_item.
Can you explain it to me?
Thank you for your help :)

Hi, thank you for your work.

Here I have a question for your top-1 loss implementation.

At modules/loss.py: line 56
loss = F.sigmoid(diff).mean() + F.sigmoid(logit ** 2).mean()

If I understood top-1 loss and your code, logit have already contained positive sample, but F.sigmoid(logit ** 2).mean() should be applied for only negative samples

Could you please check whether I understand right?

Thanks!

if len(hits) == 0: return 0

It is right. I mistake nonzero of numpy and torch, sorry...