AssertionError while evaluating 'translation'

[saichandrapandraju]

Hi @wasiahmad ,

I am trying 'translation' capabilities of PLBART and started finetuning as mentioned. But I'm getting below error in evaluation -

File "calc_code_bleu.py", line 34, in <module>
    assert len(hypothesis) == len(pre_references[i])
AssertionError

Here is a bit detailed traceback -

2021-07-15 13:57:58 | INFO | fairseq_cli.train | early stop since valid performance hasn't improved for last 10 runs
2021-07-15 13:57:58 | INFO | fairseq_cli.train | begin save checkpoint
2021-07-15 13:58:19 | INFO | fairseq.checkpoint_utils | saved checkpoint /content/PLBART/scripts/code_to_code/translation/java_cs/checkpoint_last.pt (epoch 22 @ 14168 updates, score 80.08) (writing took 20.417050701000335 seconds)
2021-07-15 13:58:19 | INFO | fairseq_cli.train | end of epoch 22 (average epoch stats below)
2021-07-15 13:58:19 | INFO | train | {"epoch": 22, "train_loss": "2.08", "train_nll_loss": "0.177", "train_ppl": "1.13", "train_wps": "1562.9", "train_ups": "1.76", "train_wpb": "890.1", "train_bsz": "16", "train_num_updates": "14168", "train_lr": "4.93409e-05", "train_gnorm": "0.534", "train_train_wall": "255", "train_wall": "8414"}
2021-07-15 13:58:19 | INFO | fairseq_cli.train | done training in 8412.8 seconds
  0% 0/250 [00:00<?, ?it/s]/usr/local/lib/python3.7/dist-packages/fairseq_cli/generate.py:172: UserWarning: --sacrebleu is deprecated. Please use --scoring sacrebleu instead.
  scorer = scoring.build_scorer(args, tgt_dict)
Traceback (most recent call last):
  File "/usr/local/bin/fairseq-generate", line 8, in <module>
    sys.exit(cli_main())
  File "/usr/local/lib/python3.7/dist-packages/fairseq_cli/generate.py", line 379, in cli_main
    main(args)
  File "/usr/local/lib/python3.7/dist-packages/fairseq_cli/generate.py", line 41, in main
    return _main(args, sys.stdout)
  File "/usr/local/lib/python3.7/dist-packages/fairseq_cli/generate.py", line 172, in _main
    scorer = scoring.build_scorer(args, tgt_dict)
  File "/usr/local/lib/python3.7/dist-packages/fairseq/scoring/__init__.py", line 54, in build_scorer
    return _build_scorer(args)
  File "/usr/local/lib/python3.7/dist-packages/fairseq/registry.py", line 54, in build_x
    return builder(args, *extra_args, **extra_kwargs)
  File "/usr/local/lib/python3.7/dist-packages/fairseq/scoring/bleu.py", line 40, in __init__
    character_tokenization=self.args.sacrebleu_char_level,
AttributeError: 'Namespace' object has no attribute 'sacrebleu_char_level'
Traceback (most recent call last):
  File "/content/PLBART/evaluation/evaluator.py", line 36, in <module>
    main()
  File "/content/PLBART/evaluation/evaluator.py", line 20, in main
    assert len(refs) == len(pres)
AssertionError
Traceback (most recent call last):
  File "calc_code_bleu.py", line 34, in <module>
    assert len(hypothesis) == len(pre_references[i])
AssertionError

Could you plz suggest how to proceed further..

[wasiahmad]

The error is due to

AttributeError: 'Namespace' object has no attribute 'sacrebleu_char_level'

I am using the following version of sacrebleu.

Name: sacrebleu
Version: 1.2.11
Summary: Hassle-free computation of shareable, comparable, and reproducible BLEU scores
Home-page: https://github.com/awslabs/sockeye

[saichandrapandraju]

Hi @wasiahmad ,

I found below workaround -

• For translation, mbart_base is not present in fairseq==0.9.0 so I installed fairseq==0.10.0.
• For fairseq==0.10.0, sacrebleu version should be >=1.4.x . So I installed sacrebleu==1.5.1.
• Replace --sacrebleu with --scoring sacrebleu in 112 line of scripts/code_to_code/translation/run.sh.
• Commented 28 line in source/sentence_prediction.py as it's erroring ModuleNotFoundError.
• Treesitter parser gave some error. so ran below:

%cd evaluation/CodeBLEU/parser
!bash build.sh
%cd ../../..

With these changes everything ran successfully.

I have couple of queries though -

• What can be the maximum input and output lengths for PLBART?
• Seems the translation data focused on 'functions'. Can this be used for all kinds of programs which involve classes, test-cases, etc..?

[wasiahmad]

This is really helpful.

• I forgot documenting about running build.sh.
• May I know what did you use instead of using maybe_shorten_dataset in line 28 in source/sentence_prediction.py?

Answers to questions

• The maximum length should be 512.
• If you have some data to fine-tune, it would work. We did such experiments and it worked.

[saichandrapandraju]

Thanks @wasiahmad ,

As I'm working only on translation part , I thought there's no need of source/sentence_prediction.py. So got away with it (atleast for now) :-)

[wasiahmad]

We are working on a release where we will release all the fine-tuned checkpoints. Also, there are bugs that are also fixed. We will also provide scripts to set up conda environment so that experiments can be run without any issue.

[saichandrapandraju]

That would be great !!!