Using BART as an example … this: tokenizer.prepare_seq2seq_batch( src_texts=['This is a very short text', 'This is shorter'], tgt_texts=['very short', 'much shorter than very short']) returns … {'input_ids': [[100, 19, 3, 9, 182, 710, 1499, 1], [947, 19, 10951, 1, 0, 0, 0, 0]], 'attentio…

Is there a way to return the "decoder_input_ids" from "tokenizer.prepare_seq2seq_batch"?

wgpubs December 23, 2020, 11:08pm 4

Ah cool … thanks for confirming.

Topic		Replies	Views
What is the correct form of decoder_input_ids for LEDForConditionalGeneration? 🤗Transformers	1	712	July 5, 2021
What should be shifted for decoder input for Bart Beginners	1	329	July 8, 2021
Decoder attention mask in text2text/se2seq generation encoder-decoder models 🤗Transformers	1	1642	March 22, 2022
T5 fine tuning, loss difference when using labels and decoder_input_ids 🤗Transformers	2	1178	October 12, 2020
How does T5 create the correct decoder_input_ids? 🤗Transformers	2	2684	September 21, 2020