• Put the link of dataset and Transformer paper in the Readme.
• More comments, such as the class SimpleLossCompute (examples/pytorch/transformer/loss/__init__.py), why the input argument "criterion" could be a list or instance (the support of Multi-GPU)?
• the function "clones" in examples/pytorch/transformer/modules/layers.py maybe not necessary.
• the class Embeddings (examples/pytorch/transformer/modules/embedding.py) could be replaced by the torch.nn.Embedding with the "max_norm" argument?
• Does the variable "N" in dgl/examples/pytorch/transformer/scripts/build_vocab.py used?
• Keep the translated text after testing, I think the user wants to see it.

1. I saw data files in the folder. You could pack them in a zip, upload it to some cloud storage (current solution, improved later) and download it on demand. You can use the download and the unzip function.
2. The data provider related codes can stay in the examples folder.
3. What are the scripts used for? If they are used to preprocess data, I will also pack them in the same zip, and run it using your python script.

@jermainewang

1. The reason i don't use download function is that one of the dataset required in my task was stored in google drive, which is not supported, thus i use another download function copied from torchtext.
2. The scripts are used for both preprocessing and evaluating.

package dependencies: requests, tqdm
The model is still training, it looks good.

Data and script related files are supposed to be downloaded from s3. I'll fix this.

Tutorial and README finished.