def test_train_model_can_instantiate_from_params(self):
params = Params.from_file(self.FIXTURES_ROOT / "simple_tagger" / "experiment.json")
# Can instantiate from base class params
TrainModel.from_params(
params=params, serialization_dir=self.TEST_DIR, local_rank=0, batch_weight_key=""
)
def test_pretrained_configs(self, path):
params = Params.from_file(
path,
params_overrides="{"
"'trainer.cuda_device': -1, "
"'trainer.use_amp': false, "
"'trainer.num_epochs': 2, "
"}",
)
# Patch max_instances in the multitask case
patch_dataset_reader(params["dataset_reader"])
if "validation_dataset_reader" in params:
# Unclear why this doesn't work for biattentive_classification_network
if "biattentive_classification_network" not in path:
patch_dataset_reader(params["validation_dataset_reader"])
# Patch any pretrained glove files with smaller fixtures.
patch_glove(params)
# Patch image_dir and feature_cache_dir keys so they point at our test fixtures instead.
patch_image_dir(params)
# Remove unnecessary keys.
for key in ("random_seed", "numpy_seed", "pytorch_seed",
"distributed"):
if key in params:
del params[key]
# Just make sure the train loop can be instantiated.
TrainModel.from_params(params=params,
serialization_dir=self.TEST_DIR,
local_rank=0)
def test_pretrained_configs(self, path):
params = Params.from_file(
path,
params_overrides="{"
"'trainer.cuda_device': -1, "
"'trainer.use_amp': false, "
"'trainer.num_epochs': 2, "
"'dataset_reader.max_instances': 4, "
"'dataset_reader.lazy': false, "
"}",
)
# Patch any pretrained glove files with smaller fixtures.
patch_glove(params)
# Remove unnecessary keys.
for key in ("random_seed", "numpy_seed", "pytorch_seed",
"distributed"):
if key in params:
del params[key]
# Just make sure the train loop can be instantiated.
TrainModel.from_params(params=params,
serialization_dir=self.TEST_DIR,
local_rank=0)
def test_train_can_fine_tune_model_from_archive(self):
params = Params.from_file(
self.FIXTURES_ROOT / "basic_classifier" / "experiment_from_archive.jsonnet"
)
train_loop = TrainModel.from_params(
params=params, serialization_dir=self.TEST_DIR, local_rank=0, batch_weight_key=""
)
train_loop.run()
model = Model.from_archive(
self.FIXTURES_ROOT / "basic_classifier" / "serialization" / "model.tar.gz"
)
# This is checking that the vocabulary actually got extended. The data that we're using for
# training is different from the data we used to produce the model archive, and we set
# parameters such that the vocab should have been extended.
assert train_loop.model.vocab.get_vocab_size() > model.vocab.get_vocab_size()
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--action",
type=Action,
choices=list(Action),
required=True)
parser.add_argument("--config", required=True)
parser.add_argument("--serialization-dir", required=True)
parser.add_argument("--batch-count", type=int, default=0)
parser.add_argument("--assume-multiprocess-types", action="store_true")
args = parser.parse_args()
params = Params.from_file(args.config)
train_model = TrainModel.from_params(params,
args.serialization_dir,
local_rank=0,
batch_weight_key="")
trainer = train_model.trainer
raw_generator = trainer.iterator(trainer.train_data,
num_epochs=1,
shuffle=True)
if args.action is Action.log:
log_iterable(raw_generator, args.assume_multiprocess_types)
elif args.action is Action.time:
time_iterable(raw_generator, args.batch_count)
elif args.action is Action.first:
time_to_first(raw_generator)
else:
raise Exception(f"Unaccounted for action {args.action}")