Question about "prev_output_tokens" in class "FairseqEncoderDecoderModel"

[XueMoonLit]

❓ Questions and Help

Before asking:

1. search the issues.
2. search the docs.

What is your question?

Code

What have you tried?

What's your environment?

• fairseq Version (e.g., 1.0 or main): 0.12
• PyTorch Version (e.g., 1.0): 1.9
• OS (e.g., Linux):Linux
• How you installed fairseq (pip, source):pip
• Build command you used (if compiling from source):
• Python version:3.8
• CUDA/cuDNN version:11.1
• GPU models and configuration:
• Any other relevant information:

I am debugging the code of Transformer model and have found that "prev_output_tokens" is almost the same as "target". I noticed that "collate" function in "fairseq/data/language_pair_dataset.py" will move "eos" token to the beginning of "target" to create "prev_output_tokens". So, during trainning stage, the samples fed into decoder are in the form of "[eos] [token1] [token2] ...". However, shouldn't "bos" token be at the begining of samples instead of "eos" token like "[bos] [token1] [token2] ...".