list some problems suffered in this project

[liuliu8622]

1. This is an farseq extension version based on the simple seq2seq framework, so i will not note problems which also happen on the original version. but one problem i must to note is the [error 13] Permission denial in code segment

"bleu_out = subprocess.check_output(bleu_cmd, stdin=read_pred, stderr=subprocess.STDOUT)", in the module seq2seq\metrics\bleu.py

cause the solution shown by others did not work in my case. Basically it's a authority problem which corresponds to the python framework but not to this project. the reason of the problem is unauthorized of the module bin\tools\multi-bleu.perl. we need to get the full authorization of this file, then the problem resolved, without change any code.

2. In seq2seq\data\input_pipeline.py, author create a new class named "ParallelTextInputPipelineFairseq", but set the value of parameter "prepend_token" ="SEQUENCE_END" (which originally should be "SEQUENCE_START"). i guess he just copy the next line but forgot to revise the corresponding value. And this can explain why he does not adopt BLEU metric_spec in the configuration file. This problem will let the reference text sliced to empty string and then the BLEU score is always being 0 in all the test process.

3. In seq2seq\metrics\bleu.py the processor firstly try to download the multi-bleu.perl script from the internet, and then load the local version if the internet is not available. this is wisdom, but for the client which can not connect to the internet. this link will cost very confusing delay(in my case, it cost about 7 minutes, instead of less than 1 seconds loading in the local). make sure the net is available or comments the corresponding lines.