mirror of
https://github.com/huggingface/transformers.git
synced 2025-07-31 02:02:21 +06:00
Trainer: let generate pick its inputs (#22108)
* Let generate pick its inputs * fix squad seq2seq example
This commit is contained in:
parent
d979cf6efd
commit
e16cbe88ae
@ -182,23 +182,11 @@ class Seq2SeqTrainer(Trainer):
|
||||
gen_kwargs["synced_gpus"] if gen_kwargs.get("synced_gpus") is not None else default_synced_gpus
|
||||
)
|
||||
|
||||
if "attention_mask" in inputs:
|
||||
gen_kwargs["attention_mask"] = inputs.get("attention_mask", None)
|
||||
if "global_attention_mask" in inputs:
|
||||
gen_kwargs["global_attention_mask"] = inputs.get("global_attention_mask", None)
|
||||
# TODO (Joao): the following line is needed to keep a consistent result on SQUAD. Ideally, we should not block
|
||||
# users from preparing a dataset with `decoder_input_ids`.
|
||||
inputs = {k: v for k, v in inputs.items() if k != "decoder_input_ids"}
|
||||
generated_tokens = self.model.generate(**inputs, **gen_kwargs)
|
||||
|
||||
# prepare generation inputs
|
||||
# some encoder-decoder models can have varying encoder's and thus
|
||||
# varying model input names
|
||||
if hasattr(self.model, "encoder") and self.model.encoder.main_input_name != self.model.main_input_name:
|
||||
generation_inputs = inputs[self.model.encoder.main_input_name]
|
||||
else:
|
||||
generation_inputs = inputs[self.model.main_input_name]
|
||||
|
||||
generated_tokens = self.model.generate(
|
||||
generation_inputs,
|
||||
**gen_kwargs,
|
||||
)
|
||||
# Temporary hack to ensure the generation config is not initialized for each iteration of the evaluation loop
|
||||
# TODO: remove this hack when the legacy code that initializes generation_config from a model config is
|
||||
# removed in https://github.com/huggingface/transformers/blob/98d88b23f54e5a23e741833f1e973fdf600cc2c5/src/transformers/generation/utils.py#L1183
|
||||
|
Loading…
Reference in New Issue
Block a user