SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

[deepspeed] fix load_best_model test (#17550)

Browse Source
This commit is contained in:
Stas Bekman
2022-06-03 11:19:03 -07:00
committed by GitHub
parent 72f5b94984
commit 26e5e129b4
1 changed files with 52 additions and 52 deletions
Download Patch File Download Diff File Expand all files Collapse all files

104
tests/deepspeed/test_deepspeed.py
View File

@@ -752,60 +752,60 @@ class TrainerIntegrationDeepSpeed(TrainerIntegrationDeepSpeedWithCustomConfig, T
# must use this setting to get the reload path exercised # must use this setting to get the reload path exercised
ds_config_dict["zero_optimization"]["stage3_gather_16bit_weights_on_model_save"] = True ds_config_dict["zero_optimization"]["stage3_gather_16bit_weights_on_model_save"] = True
tokenizer = T5Tokenizer.from_pretrained(T5_TINY)
model = T5ForConditionalGeneration.from_pretrained(T5_TINY)
def _add_eos_to_examples(example):
example["input_text"] = f"question: {example['question']} context: {example['context']}"
example["target_text"] = example["answers"]["text"][0] if len(example["answers"]["text"]) > 0 else ""
return example
def _convert_to_features(example_batch):
input_encodings = tokenizer.batch_encode_plus(
example_batch["input_text"], pad_to_max_length=True, max_length=512, truncation=True
)
target_encodings = tokenizer.batch_encode_plus(
example_batch["target_text"], pad_to_max_length=True, max_length=16, truncation=True
)
encodings = {
"input_ids": input_encodings["input_ids"],
"attention_mask": input_encodings["attention_mask"],
"labels": target_encodings["input_ids"],
}
return encodings
def get_dataset():
data_file = str(self.tests_dir / "fixtures/tests_samples/SQUAD/sample.json")
data_files = dict(train=data_file, validation=data_file)
raw_datasets = datasets.load_dataset("json", data_files=data_files, field="data")
train_dataset = raw_datasets["train"].map(_add_eos_to_examples).map(_convert_to_features, batched=True)
valid_dataset = deepcopy(train_dataset)
return train_dataset, valid_dataset
train_dataset, eval_dataset = get_dataset()
args_dict = {
"per_gpu_train_batch_size": 1,
"per_gpu_eval_batch_size": 1,
"gradient_accumulation_steps": 1,
"learning_rate": 1e-4,
"num_train_epochs": 1,
"do_train": True,
"do_eval": True,
"optim": "adafactor",
"evaluation_strategy": "steps",
"eval_steps": 1,
"save_strategy": "steps",
"save_steps": 1,
"load_best_model_at_end": True,
"max_steps": 1,
"deepspeed": ds_config_dict,
}
with mockenv_context(**self.dist_env_1_gpu): with mockenv_context(**self.dist_env_1_gpu):
tokenizer = T5Tokenizer.from_pretrained(T5_TINY)
model = T5ForConditionalGeneration.from_pretrained(T5_TINY)
def _add_eos_to_examples(example):
example["input_text"] = f"question: {example['question']} context: {example['context']}"
example["target_text"] = example["answers"]["text"][0] if len(example["answers"]["text"]) > 0 else ""
return example
def _convert_to_features(example_batch):
input_encodings = tokenizer.batch_encode_plus(
example_batch["input_text"], pad_to_max_length=True, max_length=512, truncation=True
)
target_encodings = tokenizer.batch_encode_plus(
example_batch["target_text"], pad_to_max_length=True, max_length=16, truncation=True
)
encodings = {
"input_ids": input_encodings["input_ids"],
"attention_mask": input_encodings["attention_mask"],
"labels": target_encodings["input_ids"],
}
return encodings
def get_dataset():
data_file = str(self.tests_dir / "fixtures/tests_samples/SQUAD/sample.json")
data_files = dict(train=data_file, validation=data_file)
raw_datasets = datasets.load_dataset("json", data_files=data_files, field="data")
train_dataset = raw_datasets["train"].map(_add_eos_to_examples).map(_convert_to_features, batched=True)
valid_dataset = deepcopy(train_dataset)
return train_dataset, valid_dataset
train_dataset, eval_dataset = get_dataset()
args_dict = {
"per_gpu_train_batch_size": 1,
"per_gpu_eval_batch_size": 1,
"gradient_accumulation_steps": 1,
"learning_rate": 1e-4,
"num_train_epochs": 1,
"do_train": True,
"do_eval": True,
"optim": "adafactor",
"evaluation_strategy": "steps",
"eval_steps": 1,
"save_strategy": "steps",
"save_steps": 1,
"load_best_model_at_end": True,
"max_steps": 1,
"deepspeed": ds_config_dict,
}
training_args = TrainingArguments(output_dir, **args_dict) training_args = TrainingArguments(output_dir, **args_dict)
trainer = Trainer( trainer = Trainer(