def test_training_from_pretrained_with_head_replace(pipeline: Pipeline,
dataset: Dataset,
tmp_path: str):
configuration = TrainerConfiguration(
data_bucketing=True,
batch_size=2,
num_epochs=5,
cuda_device=-1,
)
output_dir = os.path.join(tmp_path, "output")
pipeline.train(output=output_dir,
trainer=configuration,
training=dataset,
quiet=True)
pipeline.set_head(TestHead)
pipeline.config.tokenizer_config.max_nr_of_sentences = 3
copied = pipeline.copy()
assert isinstance(copied.head, TestHead)
assert copied.num_parameters == pipeline.num_parameters
assert copied.num_trainable_parameters == pipeline.num_trainable_parameters
copied_model_state = copied._model.state_dict()
original_model_state = pipeline._model.state_dict()
for key, value in copied_model_state.items():
if "backbone" in key:
assert torch.all(torch.eq(value, original_model_state[key]))
assert copied.backbone.featurizer.tokenizer.config.max_nr_of_sentences == 3
def test_training_with_data_bucketing(pipeline: Pipeline, dataset: Dataset,
tmp_path: str):
configuration = TrainerConfiguration(data_bucketing=True,
batch_size=2,
num_epochs=5)
pipeline.copy().train(
output=os.path.join(tmp_path, "output"),
trainer=configuration,
training=dataset,
validation=dataset,
lazy=False,
)
pipeline.copy().train(
output=os.path.join(tmp_path, "output"),
trainer=configuration,
training=dataset,
validation=dataset,
lazy=True,
)