A programming framework for agentic AI
fixing a few bugs in nlp (#503)
Xueqing Liu72301b856814
diff --git a/flaml/model.py b/flaml/model.pyindex 7c4b8e861..168942379 100644--- a/flaml/model.py+++ b/flaml/model.py@@ -407,14 +407,16 @@ class TransformersEstimator(BaseEstimator):def _init_hf_args(self, automl_fit_kwargs: dict = None):from .nlp.utils import HFArgs- hf_args = HFArgs()- for key, val in automl_fit_kwargs["hf_args"].items():- assert (- key in hf_args.__dict__- ), "The specified key {} is not in the argument list of flaml.nlp.utils::HFArgs".format(- key- )- setattr(hf_args, key, val)+ hf_args = HFArgs(task=self._task)+ fit_kwargs = automl_fit_kwargs.get("hf_args", None)+ if fit_kwargs:+ for key, val in fit_kwargs.items():+ assert (+ key in hf_args.__dict__+ ), "The specified key {} is not in the argument list of flaml.nlp.utils::HFArgs".format(+ key+ )+ setattr(hf_args, key, val)self.hf_args = hf_argsdef _update_hf_args(self, automl_pred_kwargs: dict = None):diff --git a/flaml/nlp/utils.py b/flaml/nlp/utils.pyindex 8604be28d..489d00b96 100644--- a/flaml/nlp/utils.py+++ b/flaml/nlp/utils.py@@ -261,10 +261,6 @@ def tokenize_row(hf_args=None,return_column_name=False,):- assert (- "max_seq_length" in hf_args.__dict__- ), "max_seq_length must be provided for glue"-if prefix:this_row = tuple(["".join(x) for x in zip(prefix, this_row)])@@ -272,12 +268,13 @@ def tokenize_row(tokenized_example = tokenizer(*tuple(this_row),padding="max_length",- max_length=hf_args.max_seq_length,+ max_length=hf_args.max_seq_length if hf_args else None,truncation=True,)if task in NLG_TASKS:tokenized_example["decoder_input_ids"] = tokenized_example["input_ids"]tmp_column_names = sorted(tokenized_example.keys())+if return_column_name:return [tokenized_example[x] for x in tmp_column_names], tmp_column_nameselse:@@ -324,7 +321,7 @@ def tokenize_swag(this_row, tokenizer, hf_args=None, return_column_name=False):tokenized_example = tokenizer(*tuple([first_sentences, second_sentences]),truncation=True,- max_length=hf_args.max_seq_length,+ max_length=hf_args.max_seq_length if hf_args else None,padding=False,)tmp_column_names = sorted(tokenized_example.keys())@@ -442,7 +439,7 @@ class Counter:return logdir-def load_model(checkpoint_path, task, num_labels, per_model_config=None):+def load_model(checkpoint_path, task, num_labels=None, per_model_config=None):import transformerstransformers.logging.set_verbosity_error()@@ -454,7 +451,7 @@ def load_model(checkpoint_path, task, num_labels, per_model_config=None):)from ..data import SEQCLASSIFICATION, SEQREGRESSION, TOKENCLASSIFICATION- def get_this_model(task, model_config):+ def get_this_model(checkpoint_path, task, model_config):from transformers import AutoModelForSequenceClassificationfrom transformers import AutoModelForSeq2SeqLMfrom transformers import AutoModelForMultipleChoice@@ -519,7 +516,7 @@ def load_model(checkpoint_path, task, num_labels, per_model_config=None):if is_pretrained_model_in_classification_head_list(this_model_type):if num_labels != num_labels_old:- this_model = get_this_model(task, new_config)+ this_model = get_this_model(checkpoint_path, task, new_config)new_config.num_labels = num_labelsthis_model.num_labels = num_labelsthis_model.classifier = (@@ -528,9 +525,9 @@ def load_model(checkpoint_path, task, num_labels, per_model_config=None):))else:- this_model = get_this_model(task, new_config)+ this_model = get_this_model(checkpoint_path, task, new_config)else:- this_model = get_this_model(task, new_config)+ this_model = get_this_model(checkpoint_path, task, new_config)this_model.resize_token_embeddings(this_vocab_size)return this_modelelse:@@ -539,7 +536,7 @@ def load_model(checkpoint_path, task, num_labels, per_model_config=None):elif task == TOKENCLASSIFICATION:model_config_num_labels = num_labelsmodel_config = _set_model_config(checkpoint_path)- this_model = get_this_model(task, model_config)+ this_model = get_this_model(checkpoint_path, task, model_config)return this_model@@ -574,6 +571,8 @@ class HFArgs:ckpt_per_epoch (int, optional, defaults to 1): An integer, the number of checkpoints per epoch."""+ task: str = field(default="seq-classification")+output_dir: str = field(default="data/output/", metadata={"help": "data dir", "required": True})@@ -608,8 +607,12 @@ class HFArgs:metadata={"help": "per gpu evaluation batch size"},)+ def __post_init__(self):+ if self.task in NLG_TASKS:+ HFArgs.model_path = "t5-small"+@staticmethod- def load_args():+ def load_args_from_console():from dataclasses import fieldsarg_parser = argparse.ArgumentParser()diff --git a/test/load_args.py b/test/load_args.pyindex 3ad467c23..50b3a72b4 100644--- a/test/load_args.py+++ b/test/load_args.py@@ -1,7 +1,7 @@def test_load_args_sub():from flaml.nlp.utils import HFArgs- HFArgs.load_args()+ HFArgs.load_args_from_console()if __name__ == "__main__":diff --git a/test/nlp/test_autohf_classificationhead.py b/test/nlp/test_autohf_classificationhead.pyindex 8fc5fc474..826e6b553 100644--- a/test/nlp/test_autohf_classificationhead.py+++ b/test/nlp/test_autohf_classificationhead.py@@ -69,3 +69,7 @@ def test_classification_head():)except requests.exceptions.HTTPError:return+++if __name__ == "__main__":+ test_classification_head()