hungpthanh / gru4rec-pytorch Goto Github PK

Hi,I have a question about how to run gru4rec and gru4rec+ with your code.From the paper,we can know these two methods vary in Loss Function and Sampling Strategy while the former is obvious in your code,I wonder how to apply the different sampling strategy.Thanks.

Hello, I found that the calculation of recall in the test set is obtained by taking topk after positive and negative sampling, which reduces the total candidate set from a large number of items to batch size. I think it may increase the recall value. May I ask why judge this way?

Hi, i have read your bench mark file https://docs.google.com/spreadsheets/d/19z6zFEY6pC0msi3wOQLk_kJsvqF8xnGOJPUGhQ36-wI/edit#gid=1577545099, and may i ask you if this code is GRU4Rec or GRU4Rec+, and can the Improved GRU4Rec bench marks be achieved with this code ? thank you.

Hi.
Thank you for sharing this code.

I found difference from original paper in regularization term of TOP1 Loss.
According to the original paper, the regularization term should be calculated over only negative samples¹.
However, this repo calculates it over all samples².
It might not be significant difference, but I just pointed it out.

https://docs.google.com/spreadsheets/d/19z6zFEY6pC0msi3wOQLk_kJsvqF8xnGOJPUGhQ36-wI/edit#gid=0

How can we use the trained model to get item predictions for a user?

Hi, I have a question about the "mask" in DataLoader(),

for i in range(minlen - 1):
     idx_input = idx_target
     idx_target = df.item_idx.values[start + i + 1]
     input = torch.LongTensor(idx_input)
     target = torch.LongTensor(idx_target)
     yield input, target,  mask

'mask' is all the same during the for loop, but I think it should be reset to [] after the first loop.
Otherwise, the reset_hidden() will always reset the hidden state.
Could you check it, Thank you!

Have a look at this. Is "hdden" a mistake for "hidden", or you just want to make recommendation through 1-step information.

        with torch.no_grad():
            hidden = self.model.init_hidden()
            for ii, (input, target, mask) in tqdm(enumerate(dataloader), total=len(dataloader.dataset.df) // dataloader.batch_size, miniters = 1000):
            #for input, target, mask in dataloader:
                input = input.to(self.device)
                target = target.to(self.device)
                logit, hidden = self.model(input, hidden)
                logit_sampled = logit[:, target.view(-1)]
                loss = self.loss_func(logit_sampled)
                recall, mrr = lib.evaluate(logit, target, k=self.topk)

Hi, lib.evaluation.Evaluation does not reset hidden staes. Is that right?