flake8 testing of https://github.com/salesforce/nonauto-nmt on Python 3.6.3

$ flake8 . --count --select=E901,E999,F821,F822,F823 --show-source --statistics

./model.py:92:75: F821 undefined name 'the_mask'
        return targets[input_mask], out[out_mask].view(-1, out.size(-1)), the_mask
                                                                          ^
./self_learn.py:626:48: F821 undefined name 'align_index'
                decoding1 = unsorted(decoding, align_index)
                                               ^
./self_learn.py:905:45: F821 undefined name 'loss_alter'
            loss_alter, loss_worse = export(loss_alter), export(loss_worse)
                                            ^
./self_learn.py:905:65: F821 undefined name 'loss_worse'
            loss_alter, loss_worse = export(loss_alter), export(loss_worse)
                                                                ^
./self_learn.py:1342:78: F821 undefined name 'fertility_mode'
    names = ['dev.src.b{}={}.{}'.format(args.beam_size, args.load_from, args,fertility_mode),
                                                                             ^
./self_learn.py:1343:77: F821 undefined name 'fertility_mode'
            'dev.trg.b{}={}.{}'.format(args.beam_size, args.load_from, args,fertility_mode),
                                                                            ^
./self_learn.py:1344:77: F821 undefined name 'fertility_mode'
            'dev.dec.b{}={}.{}'.format(args.beam_size, args.load_from, args,fertility_mode)]
                                                                            ^
./train.py:35:17: F821 undefined name 'copy'
    new_batch = copy.copy(batch)
                ^
8     F821 undefined name 'the_mask'
8

Thank you for sharing your code. I have a problem with running the code, when I run the code there is no error, but in debug mode, there is an issue in line 142 of "train.py" file. There is a "for" loop that its condition doesn't satisfy because of the "enumerate" function. how can I fix this? It seems that's because of "train" variable and enumerate function can not convert it to numbers or something like this.

I would be appreciated for any help.

In this implementation transformer decoder layer i is attending to encoder layer i output, which seems to be different from Google's original transformer implementation which always attends to the last layer output.

https://github.com/salesforce/nonauto-nmt/blob/efcbe4f2329b140ac3ce06abb6409457cebc8e49/model.py#L601

Plus, the provided scripts seem to contain options not supported by run.py, and many train_fast.sh relies on -load_from old_fast_model, which further makes the paper's results hard to reproduce.

Thank you for sharing your codes. I have a question about how to preprocess the data. For example, for the iwslt en-de dataset, you use a file named train.tags.en-de.bpe.dev.en2 in the script run_alignment_iwslt.sh. It seems to not belong to the original dataset. Where does it come from?