def _get_pytext_config(
self,
test_file_name: TestFileName,
task_class: Type[NewTask],
model_class: Type[Model],
) -> PyTextConfig:
test_file_metadata = get_test_file_metadata(test_file_name)
return PyTextConfig(
task=task_class.Config(
data=Data.Config(
source=TSVDataSource.Config(
train_filename=test_file_metadata.filename,
eval_filename=test_file_metadata.filename,
test_filename=test_file_metadata.filename,
field_names=test_file_metadata.field_names,
),
batcher=Batcher.Config(
), # Use Batcher to avoid shuffling.
),
trainer=TaskTrainer.Config(epochs=1),
model=model_class.Config(
inputs=type(model_class.Config.inputs)(
dense=FloatListTensorizer.Config(
column=test_file_metadata.dense_col_name,
dim=test_file_metadata.dense_feat_dim,
))),
),
use_tensorboard=False,
use_cuda_if_available=False,
version=LATEST_VERSION,
)
def get_tensorizers(add_dict_feat=False, add_contextual_feat=False):
schema = {"source_sequence": str, "dict_feat": Gazetteer, "target_sequence": str}
data_source = TSVDataSource.from_config(
TSVDataSource.Config(
train_filename=TEST_FILE_NAME,
field_names=["source_sequence", "dict_feat", "target_sequence"],
),
schema,
)
src_tensorizer = TokenTensorizer.from_config(
TokenTensorizer.Config(
column="source_sequence", add_eos_token=True, add_bos_token=True
)
)
tgt_tensorizer = TokenTensorizer.from_config(
TokenTensorizer.Config(
column="target_sequence", add_eos_token=True, add_bos_token=True
)
)
tensorizers = {"src_seq_tokens": src_tensorizer, "trg_seq_tokens": tgt_tensorizer}
initialize_tensorizers(tensorizers, data_source.train)
if add_dict_feat:
tensorizers["dict_feat"] = GazetteerTensorizer.from_config(
GazetteerTensorizer.Config(
text_column="source_sequence", dict_column="dict_feat"
)
)
initialize_tensorizers(
{"dict_feat": tensorizers["dict_feat"]}, data_source.train
)
return tensorizers
def test_load_saved_model(self):
with tempfile.NamedTemporaryFile() as snapshot_file:
train_data = tests_module.test_file("train_data_tiny.tsv")
eval_data = tests_module.test_file("test_data_tiny.tsv")
config = PyTextConfig(
task=DocumentClassificationTask.Config(data=Data.Config(
source=TSVDataSource.Config(
train_filename=train_data,
eval_filename=eval_data,
field_names=["label", "slots", "text"],
))),
version=LATEST_VERSION,
save_snapshot_path=snapshot_file.name,
)
task = create_task(config.task)
model = task.model
save(config, model, meta=None, tensorizers=task.data.tensorizers)
task2, config2, training_state_none = load(snapshot_file.name)
self.assertEqual(config.export, config2.export)
self.assertEqual(config.export_list, config2.export_list)
self.assertEqual(config.task, config2.task)
self.assertEqual(config, config2)
self.assertModulesEqual(model, task2.model)
self.assertIsNone(training_state_none)
model.eval()
task2.model.eval()
inputs = torch.LongTensor([[1, 2, 3]]), torch.LongTensor([3])
self.assertEqual(
model(*inputs).tolist(),
task2.model(*inputs).tolist())
def test_read_data_source_with_utf8_issues(self):
schema = {"text": str, "label": str}
data_source = TSVDataSource.from_config(
TSVDataSource.Config(
train_filename=tests_module.test_file("test_utf8_errors.tsv"),
field_names=["label", "text"],
),
schema,
)
list(data_source.train)
def test_reset_incremental_states(self):
"""
This test might seem trivial. However, interacting with the scripted
sequence generator crosses the Torchscript boundary, which can lead
to weird behavior. If the incremental states don't get properly
reset, the model will produce garbage _after_ the first call, which
is a pain to debug when you only catch it after training.
"""
tensorizers = get_tensorizers()
# Avoid numeric issues with quantization by setting a known seed.
torch.manual_seed(42)
model = Seq2SeqModel.from_config(
Seq2SeqModel.Config(
source_embedding=WordEmbedding.Config(embed_dim=512),
target_embedding=WordEmbedding.Config(embed_dim=512),
),
tensorizers,
)
# Get sample inputs using a data source.
schema = {
"source_sequence": str,
"dict_feat": Gazetteer,
"target_sequence": str,
}
data = Data.from_config(
Data.Config(source=TSVDataSource.Config(
train_filename=TEST_FILE_NAME,
field_names=[
"source_sequence", "dict_feat", "target_sequence"
],
)),
schema,
tensorizers,
)
data.batcher = Batcher(1, 1, 1)
raw_batch, batch = next(
iter(data.batches(Stage.TRAIN, load_early=True)))
inputs = model.arrange_model_inputs(batch)
model.eval()
outputs = model(*inputs)
pred, scores = model.get_pred(outputs, {"stage": Stage.TEST})
# Verify that the incremental states reset correctly.
decoder = model.sequence_generator.beam_search.decoder_ens
decoder.reset_incremental_states()
self.assertDictEqual(decoder.incremental_states, {"0": {}})
# Verify that the model returns the same predictions.
new_pred, new_scores = model.get_pred(outputs, {"stage": Stage.TEST})
self.assertEqual(new_scores, scores)
def test_load_checkpoint(self):
with tempfile.NamedTemporaryFile() as checkpoint_file:
train_data = tests_module.test_file("train_data_tiny.tsv")
eval_data = tests_module.test_file("test_data_tiny.tsv")
config = PyTextConfig(
task=DocumentClassificationTask.Config(data=Data.Config(
source=TSVDataSource.Config(
train_filename=train_data,
eval_filename=eval_data,
field_names=["label", "slots", "text"],
))),
version=LATEST_VERSION,
save_snapshot_path=checkpoint_file.name,
)
task = create_task(config.task)
model = task.model
# test checkpoint saving and loading
optimizer = create_optimizer(Adam.Config(), model)
scheduler = create_scheduler(Scheduler.Config(), optimizer)
training_state = TrainingState(
model=model,
optimizer=optimizer,
scheduler=scheduler,
start_time=0,
epoch=0,
rank=0,
stage=Stage.TRAIN,
epochs_since_last_improvement=0,
best_model_state=None,
best_model_metric=None,
tensorizers=task.data.tensorizers,
)
id = "epoch-1"
saved_path = save(config, model, None, task.data.tensorizers,
training_state, id)
# TODO: fix get_latest_checkpoint_path T53664139
# self.assertEqual(saved_path, get_latest_checkpoint_path())
task_restored, config_restored, training_state_restored = load(
saved_path)
self.assertCheckpointEqual(
model,
config,
training_state,
task_restored.model,
config_restored,
training_state_restored,
)
def test_force_predictions_on_eval(self):
tensorizers = get_tensorizers()
model = Seq2SeqModel.from_config(
Seq2SeqModel.Config(
source_embedding=WordEmbedding.Config(embed_dim=512),
target_embedding=WordEmbedding.Config(embed_dim=512),
),
tensorizers,
)
# Get sample inputs using a data source.
schema = {
"source_sequence": str,
"dict_feat": Gazetteer,
"target_sequence": str,
}
data = Data.from_config(
Data.Config(source=TSVDataSource.Config(
train_filename=TEST_FILE_NAME,
field_names=[
"source_sequence", "dict_feat", "target_sequence"
],
)),
schema,
tensorizers,
)
data.batcher = Batcher(1, 1, 1)
raw_batch, batch = next(
iter(data.batches(Stage.TRAIN, load_early=True)))
inputs = model.arrange_model_inputs(batch)
# Verify that model does not run sequence generation on prediction.
outputs = model(*inputs)
pred = model.get_pred(outputs, {"stage": Stage.EVAL})
self.assertEqual(pred, (None, None))
# Verify that attempting to set force_eval_predictions is correctly
# accounted for.
model.force_eval_predictions = True
with self.assertRaises(AssertionError):
_ = model.get_pred(outputs, {"stage": Stage.EVAL})
def _get_tensorizers(self):
schema = {"source_sequence": str, "target_sequence": str}
data_source = TSVDataSource.from_config(
TSVDataSource.Config(
train_filename=tests_module.test_file(
"compositional_seq2seq_unit.tsv"),
field_names=["source_sequence", "target_sequence"],
),
schema,
)
src_tensorizer = TokenTensorizer.from_config(
TokenTensorizer.Config(column="source_sequence",
add_eos_token=True,
add_bos_token=True))
tgt_tensorizer = TokenTensorizer.from_config(
TokenTensorizer.Config(column="target_sequence",
add_eos_token=True,
add_bos_token=True))
tensorizers = {
"src_seq_tokens": src_tensorizer,
"trg_seq_tokens": tgt_tensorizer,
}
initialize_tensorizers(tensorizers, data_source.train)
return tensorizers
