Little bugs of current codes, possible fix attached

[shrango]

1. The name of "phrase" and "mapping" files are wrong, you may:
   change rename tree-None.tree $src-$tgt.mapping.$src tmp/intra-data/* in preprocess_group.sh into rename tree-None.tree mapping.$src-$tgt.$src tmp/intra-data/*;
   and change rename tree-None.tree $src-$tgt.phrase.$src tmp/intra-data/* in preprocess_group.sh into rename tree-None.tree phrase.$src-$tgt.$src tmp/intra-data/*

2. Actually, "rename" operations may not be permitted, at least in my system. Here is my solution, FYI:

cp tmp/translation-data/* $tag/
cd tmp/intra-data
for split in train valid test
do
        for suffix in idx bin
        do
                mv $split.tree-None.tree.$suffix $split.mapping.$src-$tgt.$src.$suffix
        done
done
mv dict.tree.txt dict.tree.$src.txt
cp * ../../$tag

cd ../inter-data

for split in train valid test
do
        for suffix in idx bin
        do
                mv $split.$src-None.$src.$suffix $split.phrase.$src-$tgt.$src.$suffix
        done
done
mv dict.$src.txt dict.phrase.$src.txt

3. I believe you have run all your programs with --fp16, because I find the following Error.

 File "/data/UMST/fairseq/models/PhraseTransformer.py", line 1138, in forward
    x_word = self.downsampling(x, bpe2word_scale)
  File "/data/UMST/fairseq/models/PhraseTransformer.py", line 1057, in downsampling
    return ((mapping.transpose(1,2)@sub_word_representation.transpose(0,1)) / (torch.sqrt(num_matrix.unsqueeze(2)))).transpose(0,1)
RuntimeError: expected scalar type Float but found Half

The current code works smoothly with --fp16, yet you may want to strengthen its compatibility by fixing this bug.