我打不开谷歌，尝试了很多办法，更改hosts文件，使用谷歌镜像网站都试过了，仍没有办法打开，请问有别的链接可以下载数据集吗，万分感谢，或者发给我一份也好，我的qq:945900130。期待您的帮助

你好，我想训练中文的多标签文本分类模型，请问有什么预训练模型可以用呢？

您代码里面的解码的前向传播没怎么看懂
`< def forward(self, inputs, init_state, contexts):
if not self.config.global_emb:
embs = self.embedding(inputs)
outputs, state, attns = [], init_state, []
for emb in embs.split(1):
output, state = self.rnn(emb.squeeze(0), state)
output, attn_weights = self.attention(output, contexts)
output = self.dropout(output)
outputs += [output]
attns += [attn_weights]
outputs = torch.stack(outputs)
attns = torch.stack(attns)
return outputs, state
else:
outputs, state, attns = [], init_state, []
embs = self.embedding(inputs).split(1)
max_time_step = len(embs)
emb = embs[0]
output, state = self.rnn(emb.squeeze(0), state)
output, attn_weights = self.attention(output, contexts)
output = self.dropout(output)
soft_score = F.softmax(self.linear(output))
outputs += [output]
attns += [attn_weights]

        batch_size = soft_score.size(0)
        a, b = self.embedding.weight.size()

        for i in range(max_time_step-1):
            emb1 = torch.bmm(soft_score.unsqueeze(1), self.embedding.weight.expand((batch_size, a, b)))
            emb2 = embs[i+1]
            gamma = F.sigmoid(self.gated1(emb1.squeeze())+self.gated2(emb2.squeeze()))
            emb = gamma * emb1.squeeze() + (1 - gamma) * emb2.squeeze()
            output, state = self.rnn(emb, state)
            output, attn_weights = self.attention(output, contexts)
            output = self.dropout(output)
            soft_score = F.softmax(self.linear(output))
            outputs += [output]
            attns += [attn_weights]
        outputs = torch.stack(outputs)
        attns = torch.stack(attns)
        return outputs, state>`

作者，你好，我在1080TI上按照论文中的参数设置运行程序，但结果达不到论文中精度效果。想问下，作者你是在什么设备上运行的？

[dhy@mu01 SGM-master]$ python preprocess.py
Building source vocabulary...
max: 64, min: 1, avg: 5.17
Traceback (most recent call last):
File "preprocess.py", line 261, in
main()
File "preprocess.py", line 234, in main
opt.src_vocab_size)
File "preprocess.py", line 118, in initVocabulary
genWordVocab = makeVocabulary(dataFile, vocabSize, char=char)
File "preprocess.py", line 99, in makeVocabulary
vocab = vocab.prune(size)
File "/home/dhy/multi_label/SGM-master/data/dict.py", line 123, in prune
newDict.add(self.idxToLabel[i])
KeyError: tensor(46)
这是什么原因呢

您好，十分感谢您可以将这份工作的数据及代码开源出来。 👍
但是，在使用您的代码过程中，发现了一些小问题 #20 （python train.py时抛出错误）。

另外，我注意到在SGM这篇论文中提到了GE这个组件，且该组件发挥了很重要的作用。但该仓库只提供了一个config.yaml，我不太清楚这个配置是SGM模型的，还是SGM+GE的。我有漏掉什么东西么？您可以同时提供这两个模型的配置文件么？

再次谢谢您百忙之中回答我的问题。:smile:

Hi, Thanks for sharing the code for the paper https://arxiv.org/pdf/1806.04822.pdf!

But I'm unable to find AAPD dataset, can you point me to the right place?

Thanks in advance.

Is it meaning that data have different label length.
If my dataset's label is:

I only need to let label train is:
-2 -2 -2 -2 1 -2 -2 -2 -2 1 -2 -2 -2 -2 -2 -2 1 -2 1 -2
Thx

Hi,

Thank for your coding contribution. I am not sure why you reverse the source sequence in the collate_function. The reference link is :

能提供下 AAPD 数据集的下载地址吗？ 非常感谢

可不可以添加个license

多标签训练的时候 beam需要设置吗，我设置成非1，代码报错
candidate += [tgt_vocab.convertToLabels(s.tolist(), utils.EOS) for s in samples]
显示 s 是list对象，无法进行to list操作

I checked the source code of hamming loss in sklearn src of hamming loss
The two inputs y_pred and y_true must have the same shape. So if the lengths are not equal, just pad them with a special token?
Besides, the order of predicted sequence also matters. For example, hamming loss of [1,2,3,4] and [2,3,4,1] is 1. So we should sort y_pred and y_true before calculating the loss?

Is that normal?That is definitely not a sentence.

您好！
请问在损失函数计算中预测的标签在代码哪个位置，我想修改下 loss.py 中计算损失的函数，可以解决标签顺序的问题。
多谢了！

Hello, can you explain that what is the meaning of the 'sim_score' parameter in the 'cross_entropy_loss' function of the 'models/loss.py' file. I am confused about this. thanks.

为什么运行您的代码后从第一次eval()了一下结果：hamming_loss: 0.03198517 | micro_f1: 0.2195。后面一直是这个结果 不变化了呢？

The code for attention layer is a bit different that the one that you reported in the paper?
Could you please let me know if it has any reference ?

mldl@ub1604:~/ub16_prj/SGM$ python3 predict.py -gpus 0 -log log_name
loading checkpoint...

loading data...

loading time cost: 3.669
building model...

Traceback (most recent call last):
File "/usr/local/lib/python3.5/dist-packages/torch/nn/modules/module.py", line 482, in load_state_dict
own_state[name].copy_(param)
RuntimeError: invalid argument 2: sizes do not match at /pytorch/torch/lib/THC/generic/THCTensorCopy.c:101

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
File "predict.py", line 91, in
model.load_state_dict(checkpoints['model'])
File "/usr/local/lib/python3.5/dist-packages/torch/nn/modules/module.py", line 487, in load_state_dict
.format(name, own_state[name].size(), param.size()))
RuntimeError: While copying the parameter named decoder.embedding.weight, whose dimensions in the model are torch.Size([58, 256]) and whose dimensions in the checkpoint are torch.Size([107, 256]).
mldl@ub1604:~/ub16_prj/SGM$

In the paper, you mentioned that "In addition, the proposed models are trained using the maximum likelihood estimation method and the cross-entropy loss function, which requires humans to predefine the order of the output labels. "
but this doesn't explain for me... It means that order matters because we use seq2seq model(instead of the mentioned 'MLE and cross-entropy'), which is the intrinsic property of model. The prediction order should follow the true order. Thus human-defined order has impact.
Is the reason that high-frequency labels are more likely the root label in a hierarchical category taxonomy and thus more likely to have correlations with other labels?

Hi, thanks for your nice paper and code!
I have noticed that the standard split for RCV1 train/test in the original paper is 23,149/781,265. But from the data downloaded from your link, I found the train file is much bigger than test.
I wonder is this correct? Thanks in advance.

您好！
请问预测的部分如何能预测为序列呢？而不是直接计算分数。
例如 MCAT ECAT M13 M131 E12
代码没太看懂。
多谢了！

The actual size of the dict will be size + len(self.special), thus may cause index out of range when embedding.

您好!
最近我在知乎上看到这篇关于您论文的解读,就是代码和图1有一些不一致.我检查了代码也是,请问方便交流一下吗?
https://zhuanlan.zhihu.com/p/57098077

请问在考虑正确率的时候也考虑也预测的顺序了吗？
假设应该在y1预测出sports这个标签,但是y1预测出tennis,y2才预测出sports，
这样还算正确吗？

您好，我对您论文中有两个小疑问，因为不知道怎么打公式，我把问题放在了图片中，希望您能解答，谢谢

Hi, Lanco!
I downloaded the dataset you provided on Google Drive https://drive.google.com/drive/folders/1lBt2MjEoh4CG2jjt4nMgHro2h3k2gwUH
and viewed the src text. But the text seems meaningless and strange, eg "june bean myr smc smc board cocoa cocoa malays aver aver ton pric high high prev prev prev tawau lahad datu sabak low low bernam".
Would you let me know is there something wrong on the dataset? Thank you very much.