riaz.somc/transformers
1
0
Fork 0
mirror of https://github.com/huggingface/transformers.git synced 2025-07-04 21:30:07 +06:00
Code Issues Actions 28 Packages Projects Releases Wiki Activity

replace directly-specified-test-dirs with tmp_dir

Browse Source
This commit is contained in:
ryohei-a-shimizu 2024-12-14 01:09:15 +09:00
parent d911458b59
commit 310a6d962e
1 changed files with 141 additions and 110 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

99
tests/trainer/test_trainer.py
View File

@ -552,10 +552,10 @@ if is_torch_available():
compute_metrics = kwargs.pop("compute_metrics", None) compute_metrics = kwargs.pop("compute_metrics", None)
data_collator = kwargs.pop("data_collator", None) data_collator = kwargs.pop("data_collator", None)
optimizers = kwargs.pop("optimizers", (None, None)) optimizers = kwargs.pop("optimizers", (None, None))
output_dir = kwargs.pop("output_dir", "./regression")
preprocess_logits_for_metrics = kwargs.pop("preprocess_logits_for_metrics", None) preprocess_logits_for_metrics = kwargs.pop("preprocess_logits_for_metrics", None)
kwargs.pop("output_dir") # remove output_dir from kwargs
args = RegressionTrainingArguments(output_dir, a=a, b=b, keep_report_to=keep_report_to, **kwargs) with tempfile.TemporaryDirectory() as tmp_dir:
args = RegressionTrainingArguments(tmp_dir, a=a, b=b, keep_report_to=keep_report_to, **kwargs)
return Trainer( return Trainer(
model, model,
args, args,
@ -713,7 +713,8 @@ class TrainerIntegrationPrerunTest(TestCasePlus, TrainerIntegrationCommon):
# Base training. Should have the same results as test_reproducible_training # Base training. Should have the same results as test_reproducible_training
model = RegressionModel() model = RegressionModel()
args = TrainingArguments("./regression", learning_rate=0.1, report_to="none") with tempfile.TemporaryDirectory() as tmp_dir:
args = TrainingArguments(tmp_dir, learning_rate=0.1, report_to="none")
trainer = Trainer(model, args, train_dataset=train_dataset) trainer = Trainer(model, args, train_dataset=train_dataset)
trainer.train() trainer.train()
self.check_trained_model(trainer.model) self.check_trained_model(trainer.model)
@ -735,7 +736,8 @@ class TrainerIntegrationPrerunTest(TestCasePlus, TrainerIntegrationCommon):
def test_model_init(self): def test_model_init(self):
train_dataset = RegressionDataset() train_dataset = RegressionDataset()
args = TrainingArguments("./regression", learning_rate=0.1, report_to="none") with tempfile.TemporaryDirectory() as tmp_dir:
args = TrainingArguments(tmp_dir, learning_rate=0.1, report_to="none")
trainer = Trainer(args=args, train_dataset=train_dataset, model_init=lambda: RegressionModel()) trainer = Trainer(args=args, train_dataset=train_dataset, model_init=lambda: RegressionModel())
trainer.train() trainer.train()
self.check_trained_model(trainer.model) self.check_trained_model(trainer.model)
@ -782,8 +784,9 @@ class TrainerIntegrationPrerunTest(TestCasePlus, TrainerIntegrationCommon):
"disable_tqdm": True, "disable_tqdm": True,
} }
with tempfile.TemporaryDirectory() as tmp_dir:
args = TrainingArguments( args = TrainingArguments(
"./generation", tmp_dir,
**args_kwargs, **args_kwargs,
) )
trainer = Trainer( trainer = Trainer(
@ -797,8 +800,9 @@ class TrainerIntegrationPrerunTest(TestCasePlus, TrainerIntegrationCommon):
trainer.train() trainer.train()
grad_accum_loss_callback = StoreLossCallback() grad_accum_loss_callback = StoreLossCallback()
with tempfile.TemporaryDirectory() as tmp_dir:
args = TrainingArguments( args = TrainingArguments(
"./generation", tmp_dir,
**args_kwargs, **args_kwargs,
gradient_accumulation_steps=2, gradient_accumulation_steps=2,
per_device_train_batch_size=4, per_device_train_batch_size=4,
@ -879,8 +883,9 @@ class TrainerIntegrationPrerunTest(TestCasePlus, TrainerIntegrationCommon):
"disable_tqdm": True, "disable_tqdm": True,
} }
with tempfile.TemporaryDirectory() as tmp_dir:
args = TrainingArguments( args = TrainingArguments(
"./generation", tmp_dir,
**args_kwargs, **args_kwargs,
) )
trainer = Trainer( trainer = Trainer(
@ -894,8 +899,9 @@ class TrainerIntegrationPrerunTest(TestCasePlus, TrainerIntegrationCommon):
trainer.train() trainer.train()
grad_accum_loss_callback = StoreLossCallback() grad_accum_loss_callback = StoreLossCallback()
with tempfile.TemporaryDirectory() as tmp_dir:
args = TrainingArguments( args = TrainingArguments(
"./generation", tmp_dir,
**args_kwargs, **args_kwargs,
gradient_accumulation_steps=2, gradient_accumulation_steps=2,
per_device_train_batch_size=4, per_device_train_batch_size=4,
@ -987,7 +993,8 @@ class TrainerIntegrationPrerunTest(TestCasePlus, TrainerIntegrationCommon):
def test_custom_optimizer(self): def test_custom_optimizer(self):
train_dataset = RegressionDataset() train_dataset = RegressionDataset()
args = TrainingArguments("./regression", report_to="none") with tempfile.TemporaryDirectory() as tmp_dir:
args = TrainingArguments(tmp_dir, report_to="none")
model = RegressionModel() model = RegressionModel()
optimizer = torch.optim.SGD(model.parameters(), lr=1.0) optimizer = torch.optim.SGD(model.parameters(), lr=1.0)
lr_scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer, lr_lambda=lambda x: 1.0) lr_scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer, lr_lambda=lambda x: 1.0)
@ -1005,8 +1012,9 @@ class TrainerIntegrationPrerunTest(TestCasePlus, TrainerIntegrationCommon):
model = RegressionModel() model = RegressionModel()
num_steps, num_warmup_steps = 10, 2 num_steps, num_warmup_steps = 10, 2
extra_kwargs = {"power": 5.0, "lr_end": 1e-5} # Non-default arguments extra_kwargs = {"power": 5.0, "lr_end": 1e-5} # Non-default arguments
with tempfile.TemporaryDirectory() as tmp_dir:
args = TrainingArguments( args = TrainingArguments(
"./regression", tmp_dir,
lr_scheduler_type="polynomial", lr_scheduler_type="polynomial",
lr_scheduler_kwargs=extra_kwargs, lr_scheduler_kwargs=extra_kwargs,
learning_rate=0.2, learning_rate=0.2,
@ -1032,8 +1040,9 @@ class TrainerIntegrationPrerunTest(TestCasePlus, TrainerIntegrationCommon):
model = RegressionModel() model = RegressionModel()
num_steps, num_warmup_steps = 10, 2 num_steps, num_warmup_steps = 10, 2
extra_kwargs = {"min_lr": 1e-5} # Non-default arguments extra_kwargs = {"min_lr": 1e-5} # Non-default arguments
with tempfile.TemporaryDirectory() as tmp_dir:
args = TrainingArguments( args = TrainingArguments(
"./regression", tmp_dir,
lr_scheduler_type="cosine_with_min_lr", lr_scheduler_type="cosine_with_min_lr",
lr_scheduler_kwargs=extra_kwargs, lr_scheduler_kwargs=extra_kwargs,
learning_rate=0.2, learning_rate=0.2,
@ -1055,8 +1064,9 @@ class TrainerIntegrationPrerunTest(TestCasePlus, TrainerIntegrationCommon):
# test passed arguments for a custom ReduceLROnPlateau scheduler # test passed arguments for a custom ReduceLROnPlateau scheduler
train_dataset = RegressionDataset(length=64) train_dataset = RegressionDataset(length=64)
eval_dataset = RegressionDataset(length=64) eval_dataset = RegressionDataset(length=64)
with tempfile.TemporaryDirectory() as tmp_dir:
args = TrainingArguments( args = TrainingArguments(
"./regression", tmp_dir,
eval_strategy="epoch", eval_strategy="epoch",
metric_for_best_model="eval_loss", metric_for_best_model="eval_loss",
report_to="none", report_to="none",
@ -1087,8 +1097,9 @@ class TrainerIntegrationPrerunTest(TestCasePlus, TrainerIntegrationCommon):
train_dataset = RegressionDataset(length=64) train_dataset = RegressionDataset(length=64)
eval_dataset = RegressionDataset(length=64) eval_dataset = RegressionDataset(length=64)
with tempfile.TemporaryDirectory() as tmp_dir:
args = TrainingArguments( args = TrainingArguments(
"./regression", tmp_dir,
lr_scheduler_type="reduce_lr_on_plateau", lr_scheduler_type="reduce_lr_on_plateau",
eval_strategy="epoch", eval_strategy="epoch",
metric_for_best_model="eval_loss", metric_for_best_model="eval_loss",
@ -1127,7 +1138,8 @@ class TrainerIntegrationPrerunTest(TestCasePlus, TrainerIntegrationCommon):
from transformers.optimization import Adafactor, AdafactorSchedule from transformers.optimization import Adafactor, AdafactorSchedule
train_dataset = RegressionDataset() train_dataset = RegressionDataset()
args = TrainingArguments("./regression", report_to="none") with tempfile.TemporaryDirectory() as tmp_dir:
args = TrainingArguments(tmp_dir, report_to="none")
model = RegressionModel() model = RegressionModel()
optimizer = Adafactor(model.parameters(), scale_parameter=True, relative_step=True, warmup_init=True, lr=None) optimizer = Adafactor(model.parameters(), scale_parameter=True, relative_step=True, warmup_init=True, lr=None)
lr_scheduler = AdafactorSchedule(optimizer) lr_scheduler = AdafactorSchedule(optimizer)
@ -1179,7 +1191,8 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
train_dataset = RegressionDataset() train_dataset = RegressionDataset()
eval_dataset = RegressionDataset() eval_dataset = RegressionDataset()
model = RegressionDictModel() model = RegressionDictModel()
args = TrainingArguments("./regression", report_to="none") with tempfile.TemporaryDirectory() as tmp_dir:
args = TrainingArguments(tmp_dir, report_to="none")
trainer = Trainer(model, args, train_dataset=train_dataset, eval_dataset=eval_dataset) trainer = Trainer(model, args, train_dataset=train_dataset, eval_dataset=eval_dataset)
trainer.train() trainer.train()
_ = trainer.evaluate() _ = trainer.evaluate()
@ -1190,7 +1203,8 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
tiny_gpt2 = GPT2LMHeadModel(config) tiny_gpt2 = GPT2LMHeadModel(config)
x = torch.randint(0, 100, (128,)) x = torch.randint(0, 100, (128,))
eval_dataset = RepeatDataset(x) eval_dataset = RepeatDataset(x)
args = TrainingArguments("./test", report_to="none") with tempfile.TemporaryDirectory() as tmp_dir:
args = TrainingArguments(tmp_dir, report_to="none")
trainer = Trainer(tiny_gpt2, args, eval_dataset=eval_dataset) trainer = Trainer(tiny_gpt2, args, eval_dataset=eval_dataset)
# By default the past_key_values are removed # By default the past_key_values are removed
result = trainer.predict(eval_dataset) result = trainer.predict(eval_dataset)
@ -1203,7 +1217,8 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
def test_training_arguments_are_left_untouched(self): def test_training_arguments_are_left_untouched(self):
trainer = get_regression_trainer() trainer = get_regression_trainer()
trainer.train() trainer.train()
args = TrainingArguments("./regression", report_to=[]) with tempfile.TemporaryDirectory() as tmp_dir:
args = TrainingArguments(tmp_dir, report_to=[])
dict1, dict2 = args.to_dict(), trainer.args.to_dict() dict1, dict2 = args.to_dict(), trainer.args.to_dict()
for key in dict1.keys(): for key in dict1.keys():
# Logging dir can be slightly different as they default to something with the time. # Logging dir can be slightly different as they default to something with the time.
@ -1450,8 +1465,9 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
train_dataset = RepeatDataset(x) train_dataset = RepeatDataset(x)
# Trainer without inf/nan filter # Trainer without inf/nan filter
with tempfile.TemporaryDirectory() as tmp_dir:
args = TrainingArguments( args = TrainingArguments(
"./test", tmp_dir,
learning_rate=1e-9, learning_rate=1e-9,
logging_steps=5, logging_steps=5,
logging_nan_inf_filter=False, logging_nan_inf_filter=False,
@ -1472,8 +1488,9 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
# redefine the model # redefine the model
tiny_gpt2 = GPT2LMHeadModel(config) tiny_gpt2 = GPT2LMHeadModel(config)
# Trainer without inf/nan filter # Trainer without inf/nan filter
with tempfile.TemporaryDirectory() as tmp_dir:
args = TrainingArguments( args = TrainingArguments(
"./test", tmp_dir,
learning_rate=1e-9, learning_rate=1e-9,
logging_steps=5, logging_steps=5,
logging_nan_inf_filter=False, logging_nan_inf_filter=False,
@ -1504,16 +1521,18 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
train_dataset = RepeatDataset(x) train_dataset = RepeatDataset(x)
# Trainer without inf/nan filter # Trainer without inf/nan filter
with tempfile.TemporaryDirectory() as tmp_dir:
args = TrainingArguments( args = TrainingArguments(
"./test", learning_rate=1e9, logging_steps=5, logging_nan_inf_filter=False, report_to="none" tmp_dir, learning_rate=1e9, logging_steps=5, logging_nan_inf_filter=False, report_to="none"
) )
trainer = Trainer(tiny_gpt2, args, train_dataset=train_dataset) trainer = Trainer(tiny_gpt2, args, train_dataset=train_dataset)
trainer.train() trainer.train()
log_history_no_filter = trainer.state.log_history log_history_no_filter = trainer.state.log_history
# Trainer with inf/nan filter # Trainer with inf/nan filter
with tempfile.TemporaryDirectory() as tmp_dir:
args = TrainingArguments( args = TrainingArguments(
"./test", learning_rate=1e9, logging_steps=5, logging_nan_inf_filter=True, report_to="none" tmp_dir, learning_rate=1e9, logging_steps=5, logging_nan_inf_filter=True, report_to="none"
) )
trainer = Trainer(tiny_gpt2, args, train_dataset=train_dataset) trainer = Trainer(tiny_gpt2, args, train_dataset=train_dataset)
trainer.train() trainer.train()
@ -1576,7 +1595,8 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
train_dataset = RegressionDataset() train_dataset = RegressionDataset()
config = GPT2Config(vocab_size=100, n_positions=128, n_embd=32, n_layer=3, n_head=4) config = GPT2Config(vocab_size=100, n_positions=128, n_embd=32, n_layer=3, n_head=4)
tiny_gpt2 = GPT2LMHeadModel(config) tiny_gpt2 = GPT2LMHeadModel(config)
args = TrainingArguments("./test", report_to="none", dataloader_persistent_workers=False) with tempfile.TemporaryDirectory() as tmp_dir:
args = TrainingArguments(tmp_dir, report_to="none", dataloader_persistent_workers=False)
# Single evaluation dataset # Single evaluation dataset
eval_dataset = RegressionDataset() eval_dataset = RegressionDataset()
@ -1619,8 +1639,9 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
train_dataset = RegressionDataset() train_dataset = RegressionDataset()
config = GPT2Config(vocab_size=100, n_positions=128, n_embd=32, n_layer=3, n_head=4) config = GPT2Config(vocab_size=100, n_positions=128, n_embd=32, n_layer=3, n_head=4)
tiny_gpt2 = GPT2LMHeadModel(config) tiny_gpt2 = GPT2LMHeadModel(config)
with tempfile.TemporaryDirectory() as tmp_dir:
args = TrainingArguments( args = TrainingArguments(
"./test", tmp_dir,
report_to="none", report_to="none",
dataloader_persistent_workers=True, dataloader_persistent_workers=True,
dataloader_num_workers=2, dataloader_num_workers=2,
@ -1678,8 +1699,9 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
self.assertNotEqual(modeling_llama.apply_rotary_pos_emb, liger_rotary_pos_emb) self.assertNotEqual(modeling_llama.apply_rotary_pos_emb, liger_rotary_pos_emb)
self.assertFalse(isinstance(tiny_llama.model.norm, LigerRMSNorm)) self.assertFalse(isinstance(tiny_llama.model.norm, LigerRMSNorm))
with tempfile.TemporaryDirectory() as tmp_dir:
args = TrainingArguments( args = TrainingArguments(
"./test", tmp_dir,
use_liger_kernel=True, use_liger_kernel=True,
) )
Trainer(tiny_llama, args) Trainer(tiny_llama, args)
@ -2162,8 +2184,9 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
# Make the Trainer believe it's a parallelized model # Make the Trainer believe it's a parallelized model
model.is_parallelizable = True model.is_parallelizable = True
model.model_parallel = True model.model_parallel = True
with tempfile.TemporaryDirectory() as tmp_dir:
args = TrainingArguments( args = TrainingArguments(
"./regression", per_device_train_batch_size=16, per_device_eval_batch_size=16, report_to="none" tmp_dir, per_device_train_batch_size=16, per_device_eval_batch_size=16, report_to="none"
) )
trainer = Trainer(model, args, train_dataset=RegressionDataset(), eval_dataset=RegressionDataset()) trainer = Trainer(model, args, train_dataset=RegressionDataset(), eval_dataset=RegressionDataset())
# Check the Trainer was fooled # Check the Trainer was fooled
@ -2518,7 +2541,8 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
def test_dynamic_shapes(self): def test_dynamic_shapes(self):
eval_dataset = DynamicShapesDataset(batch_size=self.batch_size) eval_dataset = DynamicShapesDataset(batch_size=self.batch_size)
model = RegressionModel(a=2, b=1) model = RegressionModel(a=2, b=1)
args = TrainingArguments("./regression", report_to="none") with tempfile.TemporaryDirectory() as tmp_dir:
args = TrainingArguments(tmp_dir, report_to="none")
trainer = Trainer(model, args, eval_dataset=eval_dataset) trainer = Trainer(model, args, eval_dataset=eval_dataset)
# Check evaluation can run to completion # Check evaluation can run to completion
@ -2535,7 +2559,8 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
self.assertTrue(np.all(seen[expected.shape[0] :] == -100)) self.assertTrue(np.all(seen[expected.shape[0] :] == -100))
# Same tests with eval accumulation # Same tests with eval accumulation
args = TrainingArguments("./regression", eval_accumulation_steps=2, report_to="none") with tempfile.TemporaryDirectory() as tmp_dir:
args = TrainingArguments(tmp_dir, eval_accumulation_steps=2, report_to="none")
trainer = Trainer(model, args, eval_dataset=eval_dataset) trainer = Trainer(model, args, eval_dataset=eval_dataset)
# Check evaluation can run to completion # Check evaluation can run to completion
@ -3185,7 +3210,8 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
) )
eval_dataset = GlueDataset(data_args, tokenizer=tokenizer, mode="dev") eval_dataset = GlueDataset(data_args, tokenizer=tokenizer, mode="dev")
training_args = TrainingArguments(output_dir="./examples", use_cpu=True, report_to="none") with tempfile.TemporaryDirectory() as tmp_dir:
training_args = TrainingArguments(output_dir=tmp_dir, use_cpu=True, report_to="none")
trainer = Trainer(model=model, args=training_args, eval_dataset=eval_dataset) trainer = Trainer(model=model, args=training_args, eval_dataset=eval_dataset)
result = trainer.evaluate() result = trainer.evaluate()
self.assertLess(result["eval_loss"], 0.2) self.assertLess(result["eval_loss"], 0.2)
@ -3202,8 +3228,9 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
) )
for example in dataset.examples: for example in dataset.examples:
example["labels"] = example["input_ids"] example["labels"] = example["input_ids"]
with tempfile.TemporaryDirectory() as tmp_dir:
training_args = TrainingArguments( training_args = TrainingArguments(
output_dir="./examples", output_dir=tmp_dir,
use_cpu=True, use_cpu=True,
per_device_eval_batch_size=1, per_device_eval_batch_size=1,
report_to="none", report_to="none",
@ -3237,7 +3264,8 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
# Adding one column not used by the model should have no impact # Adding one column not used by the model should have no impact
train_dataset = SampleIterableDataset(label_names=["labels", "extra"]) train_dataset = SampleIterableDataset(label_names=["labels", "extra"])
args = RegressionTrainingArguments(output_dir="./examples", max_steps=4) with tempfile.TemporaryDirectory() as tmp_dir:
args = RegressionTrainingArguments(output_dir=tmp_dir, max_steps=4)
trainer = Trainer(model=model, args=args, train_dataset=train_dataset) trainer = Trainer(model=model, args=args, train_dataset=train_dataset)
trainer.train() trainer.train()
self.assertEqual(trainer.state.global_step, 4) self.assertEqual(trainer.state.global_step, 4)
@ -3252,7 +3280,8 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
# Adding one column not used by the model should have no impact # Adding one column not used by the model should have no impact
eval_dataset = SampleIterableDataset(label_names=["labels", "extra"]) eval_dataset = SampleIterableDataset(label_names=["labels", "extra"])
args = RegressionTrainingArguments(output_dir="./examples") with tempfile.TemporaryDirectory() as tmp_dir:
args = RegressionTrainingArguments(output_dir=tmp_dir)
trainer = Trainer(model=model, args=args, eval_dataset=eval_dataset, compute_metrics=AlmostAccuracy()) trainer = Trainer(model=model, args=args, eval_dataset=eval_dataset, compute_metrics=AlmostAccuracy())
results = trainer.evaluate() results = trainer.evaluate()
@ -3279,7 +3308,8 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
model = RegressionPreTrainedModel(config) model = RegressionPreTrainedModel(config)
eval_dataset = SampleIterableDataset() eval_dataset = SampleIterableDataset()
args = RegressionTrainingArguments(output_dir="./examples") with tempfile.TemporaryDirectory() as tmp_dir:
args = RegressionTrainingArguments(output_dir=tmp_dir)
trainer = Trainer(model=model, args=args, eval_dataset=eval_dataset, compute_metrics=AlmostAccuracy()) trainer = Trainer(model=model, args=args, eval_dataset=eval_dataset, compute_metrics=AlmostAccuracy())
preds = trainer.predict(trainer.eval_dataset).predictions preds = trainer.predict(trainer.eval_dataset).predictions
@ -4052,7 +4082,8 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
train_dataset = RegressionDataset() train_dataset = RegressionDataset()
eval_dataset = RegressionDataset() eval_dataset = RegressionDataset()
model = RegressionDictModel() model = RegressionDictModel()
args = TrainingArguments("./regression", report_to="none", eval_use_gather_object=True) with tempfile.TemporaryDirectory() as tmp_dir:
args = TrainingArguments(tmp_dir, report_to="none", eval_use_gather_object=True)
trainer = Trainer(model, args, train_dataset=train_dataset, eval_dataset=eval_dataset) trainer = Trainer(model, args, train_dataset=train_dataset, eval_dataset=eval_dataset)
trainer.train() trainer.train()
_ = trainer.evaluate() _ = trainer.evaluate()