Do any fairseq CLI's support forced decoding?

[erip]

❓ Questions and Help

What is your question?

Is there some code path to enable forced decoding from the CLI?

Code

It seems like adding encoded inputs as prefix_tokens to the SequenceGenerator's forward method should do this, but I'm not sure if there's a code path for this in the CLI.

What have you tried?

Looked through documentation and code. 😄

What's your environment?

• fairseq Version (e.g., 1.0 or main): main
• PyTorch Version (e.g., 1.0) 1.12
• OS (e.g., Linux): Linux
• How you installed fairseq (pip, source): source
• Build command you used (if compiling from source): n/a
• Python version: 3.10
• CUDA/cuDNN version: n/a
• GPU models and configuration: n/a
• Any other relevant information: n/a

[erip]

It seems fairseq-generate supports a --score-references flag which seems to do what I expect.

[gmryu]

There is a --prefix-size INT, which make generation initialized by target prefix of given length. Default: 0 This argument makes prefix_tokens and as you stated prefix_tokens are passed to generator: https://github.com/facebookresearch/fairseq/blob/main/fairseq_cli/generate.py#L201-L207

--

Does --score-references helps?
I may not understand your question, but it does not generate, right? From codes here, it uses SequenceScorer instead of SequenceGenerator.

From SequenceScorer's forward L120-121, L145, it returns the answer directly.

Well, if feels like --score-references is an alternative of validation. (eval_lm can not be used with translation)