Why setting label to 0 for all data in gen_data_set_youtube()?

[Wp-Zhang]

I'm new to this area so it may be a dumb question.
When generating the dataset for YoutubeDNN here, why do we need to set the label for the whole dataset as 0 as the code below instead of setting 1 to positive items and 0 to sampled negative items which is the method in the Tensorflow version of DeepMatch?
train_set.append((reviewerID, hist[::-1], [pos_list[i]] + [neg_list[item_idx] for item_idx in np.random.choice(neg_list, negsample)], 0, len(hist[::-1]),rating_list[i]))

[bbruceyuan]

hi, it is a good question and I think about this question a long time too.

English Version:
As we all know, there is no sampled_softmax_loss like tf. Then I use CrossEntropyLoss instead of sampled_softmax_loss through constructing sample label [1, 0, 0, 0, 0], and set postivate sample displayed in first position, all 0 is sampled negative sample.

Additionally, in CrossEntropyLoss, Target shape is (N), so label should be 0 (the index of this positive sample).

my code can see: PLBaseModel.py _step function. https://github.com/bbruceyuan/DeepMatch-Torch/blob/322589cbe7f6d4efb1c4094bd6fb62f599f62c29/deepmatch_torch/models/PLBaseModel.py

If you have a better pytorch version sampled_softmax_loss implementation，can you seed me a pr to construct it better?

Chinese Version:
在 PyTorch 里面，没有 sampledsoftmax_loss 函数，所以我通过 自己构造样本 [1, 0, 0, 0, 0] 进行 CrossEntropyLoss 来代替 sampled_softmax_loss, 1是正样本， 0 是采样的, 让 第一个元素用于是 正样本，这样进行交叉熵可以起到类似 sampledsoftmax 的作用。

那么在 PyTorch 里面的 CrossEntropyLoss 的输入参数是 Target, shape 是 (N)；因为我们用交叉熵做 loss, 所以 label 是第一个位置，因此 label 是 0。

代码部分可以查看：https://github.com/bbruceyuan/DeepMatch-Torch/blob/322589cbe7f6d4efb1c4094bd6fb62f599f62c29/deepmatch_torch/models/PLBaseModel.py 中 _step 函数；