haojihu / sets2sets Goto Github PK

When predicting the scores for every item in the set, you use softmax to normarlize the vector, but it will result in sum 1.
In the multi-label classification, should we use the sigmoid function instead?
Actually, in my own multi-label classification experiment, i use the softmax on my output scores, the performance drops than no softmax.

can you give more details about the OPTUM dataset used in this paper?
i wan t to know where can i get this dataset. Thank you very much.

It semms that when calculating the WMSE loss, we calculate the mse between a softmax probability vector generated by decoder and the groundtruth multi-hot vector.
When testing, we choose top-k to get the multi-hot prediction vector.
Have you ever tried when training, we also choose the topk from o(vi) generated by the decoder, and then calculate the distance between two multi-hot vector?
In this way, the operations of training and testing is consistent.

Waiting for your response, thank you!

I get a bug in T-mall datasets called:

(base) wzk@ddst:~/work/Sets2Sets$ python Sets2Sets.py ./data/alibaba_history.csv ./data/alibaba_future.csv 1 2 1
start dictionary generation...
{'MATERIAL_NUMBER': 9531}
# dimensions of final vector: 9531 | 2962
finish dictionary generation*****
num of vectors having entries more than 1: 16462
num of vectors having entries more than 1: 15275
Traceback (most recent call last):
  File "Sets2Sets.py", line 990, in <module>
    main(sys.argv)
  File "Sets2Sets.py", line 955, in main
    codes_freq = get_codes_frequency_no_vector(data_chunk[past_chunk],input_size,data_chunk[future_chunk].keys())
  File "Sets2Sets.py", line 935, in get_codes_frequency_no_vector
    for idx in X[pid]:
KeyError: '371250'

Have anyone met this before? I'd be really appreciated if anyone can help.

Hi, in your loss function, it contains the WSME and PSE parts, have you ever done the experiments about the single part? Can i just use the WSME part to do the multi-label classification task?
Waiting for your response, thank you!