blickwinkel1107 / making-the-most-of-context-nmt Goto Github PK

I have been experiencing this issue during the process of reproducing your model.
NaN or Inf found in input tensor.
How can I solve it

I encountered this problem during the training. I think it may be that their format is wrong. I checked the script, and you defined it like this.
I want to know what your data of valid and reference formats are. (20 of doc?sents?)
Because I can't print out the input, I won't solve this problem. I hope to get your answer. Thank you

valid_data: - "/home/user_data55/yuex/IWSLT15.zh-en/docs_data/dev.zh.doc.bpe.20" - "/home/user_data55/yuex/IWSLT15.zh-en/docs_data/dev.en.doc.bpe.20" bleu_valid_reference: "/home/user_data55/yuex/IWSLT15.zh-en/sents_data/dev.en.norm.tok"

I had this problem in training.
"where_cpu" not implemented for 'Bool'

I don't know which part of the code should be changed. Can you help me? Thank you

Hello
I'm going to reproduce your code.

1.4 Building Vocabulary
What is the role of this step, and what is the vocabulary to do after BPE？

1.5 Documental Data Format for Model Processing
Training data contains several chapters. Is each chapter a document?
In addition, in your example, there is only the symbol < BOS > of each sentence. Do you need something like < DOS >? I don't quite understand what each document means
That is to say, is the specific training data just < EOS> added to each sentence
There is one more question
you partition the original document as up to 20 sentences as a group.
What do you mean, your training data only has 20 sentences? I don't understand this training process

What else should I pay attention to? It may be because I am not familiar with it, so I didn't understand what you meant.

I am looking forward to your answer. I may need to discuss with you later. In the process of reproducing your model, thank you