Hi, I am new to XMC. After downloading all datasets, I noticed that the labels for both test and train sets are number or index. I am wondering if there is a file storing the exact text label corresponding to each label index. Thank you so much.

Hi! Did I need to do something for label clustering when I train LightXML on my own datasets other than [eurlex4k|wiki31k|amazon13k|amazon670k|wiki500k] Why does LightXML only use label clustering on Wiki500k&670k?

Thanks for your code !!!

After running your code, I found that you just did a simply binary classification for Wiki10-31k (without label recalling , label ranking and dynamic negative sampling) :

https://github.com/kongds/LightXML/blob/b9af9443004d3bce8b9116edfe038b702d1b295c/src/model.py#L100-L107

So, I wonder whether it was all you've done for Wiki10-31k or I missed something important, thanks a lot!

The apex==0.9.10dev that you mentioned in the requirements.txt file is not a valid version of apex which you wanted to use. I got bellow error when running your code

  File "/usr/local/lib/python3.7/dist-packages/apex/__init__.py", line 13, in <module>
    from pyramid.session import UnencryptedCookieSessionFactoryConfig
ImportError: cannot import name 'UnencryptedCookieSessionFactoryConfig' from 'pyramid.session' (unknown location)

According to this discussion best way to solve this issue is to:

1. remove apex==0.9.10dev from requirements.txt
2. add bellow instructions in readme as a step of requirement installation

git clone https://github.com/NVIDIA/apex
cd apex
python setup.py install

Hello. Thank you very much for the impressive work and for sharing your work. I have one question related to the training process. In the paper, I read that LightXML can use raw text data to provide end-to-end prediction, similar to usual deep learning based approaches which use raw text.

However, when I ran the LightXML code, I found that TfIdf feature is used in the train.txt file. I wonder why this TfIdf is necessary? Would the TfIdf affect the performance? Can we swap the TfIdf with other vectorizer (e.g., bag of word feature, word piece, etc.).

Thank you very much for the help!

Only ensemble of 'bert-base', 'roberta' and 'xlnet' would get sota on dataset of Eurlex-4K, AmazonCat-13K. Is that realy？

code from 'src/model.py':

is_training = labels is not None
outs = self.bert(
    input_ids,
    attention_mask=attention_mask,
    token_type_ids=token_type_ids
)[-1]

out = torch.cat([outs[-i][:, 0] for i in range(1, self.feature_layers+1)], dim=-1)
out = self.drop_out(out)
group_logits = self.l0(out)
if self.group_y is None:
    logits = group_logits
    if is_training:
        loss_fn = torch.nn.BCEWithLogitsLoss()
        loss = loss_fn(logits, labels)
        return logits, loss
    else:
        return logits

The "requirements.txt" shows the level of the apex is 0.9.10dev. Do you mean the apex from the pypi(https://libraries.io/pypi/apex) or the Nivida apex(https://github.com/NVIDIA/apex)?

Hi! Great work! : D If the version of Python is 3.6 in your code? And Does LightXML just use Bert-base-uncased from Huggingface, or train it on specific corpus for specific task？ Thanks!

Thanks for your excellent work! I download EUR-Lex.tar but only find train.txt, train_labels.txt, train_texts.txt, test.txt ,test_labels.txt and test_texts.txt. There is no Eurlex-4K/train_raw_texts.txt in tar file. Could you please tell me where to find such file? Thanks!

Could you please tell me the CUDA version of the project cause I encounter the problem of "RuntimeError: CUDA error: CUBLAS_STATUS_EXECUTION_FAILED when calling cublasGemmEx( handle, opa, opb, m, n, k, &falpha, a, CUDA_R_16F, lda, b, CUDA_R_16F, ldb, &fbeta, c, CUDA_R_16F, ldc, CUDA_R_32F, CUBLAS_GEMM_DFALT_TENSOR_OP)". I think the description of CUDA version may help.

Hi

Thanks for releasing the code for your paper. I was trying out your code on a new dataset. Things are clear to me except the clustering part: can you please release the code to cluster your labels for a new dataset or describe the exact steps?