Scores in generate()

Hi,

I was wondering why the length of the output_scores is always +1 longer than the max_length in the output of generate()? This seems to be not consistent with the documentation https://huggingface.co/transformers/internal/generation_utils.html#transformers.generation_utils.BeamSampleEncoderDecoderOutput

I found the scores from the output of the generate() function when setting output_scores to be True is (max_length+1,)-shaped tensors or shorter due to the early eos_token_id with each element of shape (batch_size*num_beams, config.vocab_size). The shape of the output.sequence is (batch_size, max_length).

Hey @Kylie,

good observation! So output_scores should max_length - 1. The reason is that the first token, the decoder_start_token_id is not generated, meaning that no scores can be calculated.

Here an example:

#!/usr/bin/env python3
from transformers import AutoModelForSeq2SeqLM
import torch

model = AutoModelForSeq2SeqLM.from_pretrained('facebook/bart-large')

out = model.generate(torch.tensor([10 * [1]]), return_dict_in_generate=True, output_scores=True, max_length = 10)

print("len scores:", len(out.scores))  # should give 9

Would you be interested in correcting the documentation in a PR for Transformers?

Hi @patrickvonplaten ,

Thanks for your reply! That makes more sense. Sure, I can correct that in a PR.

1 Like

Hi @patrickvonplaten

Any reason that the decoder_start_token_id is concatenated to the beginning of the generated token ids?

Hi @ad26kr, BART uses <bos> at the beginning of the decoder input to indicate the start of decoding. This is how the model pretrained (see Figure 1(c) of the paper).

Hi @Kylie
I know that several seq2seq models such as BART, T5 use some special tokens as the first input token to let the decoder start decoding. However, I can’t understand why ‘the generated text/tokens from the model.generate()’ includes the special tokens (in the beginning of the generated token ids)