THUMT
THUMT copied to clipboard
THUNLP-MT
An open-source neural machine translation toolkit developed by Tsinghua Natural Language Processing Group
tensorflow版本加载训练数据时target端为什么只在结束加eos,却没有在开头加bos。 而对比pytorch版本在target前后都加了标签。 
It seems that the parameter enabling CPU inference is not correctly passed to the function `get_infer_dataset`. Actually, the function `get_eval_dataset` has the same issue.  Those who have this problem...
1. The encdec_attention is added to model/transformer.py, it is helpful to speed up inference. 2. When we convert the pytorch model (model.pt) to onnx models (like fastt5), it is necesaary...
你好，我发现即使我在同一数据集（同一个随机种子打乱）的条件下训练，可是每次训练结束的验证集的bleu值都会产生较大的波动，问一下是训练的时候还有类似seed的参数吗？ 谢谢！
When I run: thumt-trainer \ --input corpus.tc.32k.zh.shuf corpus.tc.32k.en.shuf \ --vocabulary vocab.32k.zh.txt vocab.32k.en.txt \ --model transformer \ --validation newsdev2017.tc.32k.zh \ --references newsdev2017.tc.en \ --parameters=batch_size=4096,device_list=[0,1,2,3],update_cycle=2 \ --hparam_set base I have got this...
训练报错 Providing a bool or integral fill value without setting the optional `dtype` or `out` arguments is currently unsupported. In PyTorch 1.7, 什么版本的pytoch可以使用？
When I use the dataset wmt14en-de ,I got the bleu score:24.5,which is just like the paper's score, but when I use the same way to train the model with Wmt17...
I use Py-THUMT as the framework of new NMT algorithm. I only replace Transformer with another module and other parts remain unchanged. I use THUMT add_name() to add parameters to...
Hi, How can I set params or modify "translator.py" for using **cpu** to inference?
