def populate_defaults(input_feature):
set_default_values(input_feature, {TIED: None, "encoder": "parallel_cnn", "level": "word"})
encoder_class = get_encoder_cls(input_feature["type"], input_feature["encoder"])
if hasattr(encoder_class, "default_params"):
set_default_values(input_feature, encoder_class.default_params)
def test_sequence_embed_encoder(enc_embedding_size: int,
input_sequence: torch.Tensor) -> None:
encoder_parameters["embedding_size"] = enc_embedding_size
encoder_obj = get_encoder_cls(SEQUENCE, "embed")(**encoder_parameters)
encoder_out = encoder_obj(input_sequence)
assert encoder_out["encoder_output"].size()[1:] == encoder_obj.output_shape
def test_passthrough_encoder(enc_reduce_output, input_sequence):
encoder_parameters = {"reduce_output": enc_reduce_output}
# retrieve encoder to test
encoder_obj = get_encoder_cls(SEQUENCE,
"passthrough")(**encoder_parameters)
encoder_out = encoder_obj(input_sequence)
assert "encoder_output" in encoder_out
assert (encoder_out["encoder_output"].shape == (BATCH_SIZE, SEQ_SIZE, 1)
if enc_reduce_output is None else (BATCH_SIZE, 1))
def test_sequence_embed_encoder(enc_embedding_size: int,
input_sequence: torch.Tensor) -> None:
encoder_parameters["embedding_size"] = enc_embedding_size
# retrieve encoder to test
encoder_obj = get_encoder_cls(SEQUENCE, "embed")(**encoder_parameters)
encoder_out = encoder_obj(input_sequence)
# check to make sure shape is correct depending on relationship
# of vocab_size and embedding_size
if enc_embedding_size > TEST_VOCAB_SIZE:
assert (BATCH_SIZE, *encoder_obj.output_shape) == (BATCH_SIZE,
TEST_VOCAB_SIZE)
else:
assert (BATCH_SIZE, *encoder_obj.output_shape) == (BATCH_SIZE,
enc_embedding_size)
def initialize_encoder(self, encoder_parameters):
return get_encoder_cls(self.type(), self.encoder)(**encoder_parameters)
def test_sequence_encoders(
enc_encoder: str,
enc_cell_type: str,
enc_dropout: float,
enc_num_layers: int,
enc_norm: Union[None, str],
enc_reduce_output: Union[None, str],
enc_should_embed: bool,
input_sequence: torch.Tensor,
):
# update encoder parameters for specific unit test case
encoder_parameters["cell_type"] = enc_cell_type
encoder_parameters["dropout"] = enc_dropout
encoder_parameters["num_layers"] = enc_num_layers
encoder_parameters["norm"] = enc_norm
encoder_parameters["reduce_output"] = enc_reduce_output
# retrieve encoder to test
encoder_obj = get_encoder_cls(SEQUENCE, enc_encoder)(**encoder_parameters)
if torch.cuda.is_available():
encoder_obj = encoder_obj.cuda()
encoder_out = encoder_obj(input_sequence)
assert "encoder_output" in encoder_out
assert isinstance(encoder_out["encoder_output"], torch.Tensor)
if enc_encoder == "parallel_cnn":
number_parallel_cnn_layers = PARALLEL_CNN_LAYERS
output_dimension = encoder_parameters[
"num_filters"] * number_parallel_cnn_layers
assert (encoder_out["encoder_output"].shape == (BATCH_SIZE, SEQ_SIZE,
output_dimension) if
enc_reduce_output is None else (BATCH_SIZE, output_dimension))
elif enc_encoder == "stacked_parallel_cnn":
number_parallel_cnn_layers = PARALLEL_CNN_LAYERS
output_dimension = encoder_parameters[
"num_filters"] * number_parallel_cnn_layers
assert (encoder_out["encoder_output"].shape == (BATCH_SIZE, SEQ_SIZE,
output_dimension) if
enc_reduce_output is None else (BATCH_SIZE, output_dimension))
elif enc_encoder == "rnn":
assert (encoder_out["encoder_output"].shape == (BATCH_SIZE, SEQ_SIZE,
TEST_STATE_SIZE) if
enc_reduce_output is None else (BATCH_SIZE, TEST_STATE_SIZE))
assert "encoder_output_state" in encoder_out
if enc_cell_type == "lstm":
assert isinstance(encoder_out["encoder_output_state"], tuple)
assert isinstance(encoder_out["encoder_output_state"][0],
torch.Tensor)
assert isinstance(encoder_out["encoder_output_state"][1],
torch.Tensor)
assert encoder_out["encoder_output_state"][0].shape == (
BATCH_SIZE, TEST_STATE_SIZE)
assert encoder_out["encoder_output_state"][1].shape == (
BATCH_SIZE, TEST_STATE_SIZE)
else:
assert isinstance(encoder_out["encoder_output_state"],
torch.Tensor)
assert encoder_out["encoder_output_state"].shape == (
BATCH_SIZE, TEST_STATE_SIZE)
elif enc_encoder == "cnnrnn":
assert encoder_out["encoder_output"].shape[
1:] == encoder_obj.output_shape
assert "encoder_output_state" in encoder_out
if enc_cell_type == "lstm":
assert isinstance(encoder_out["encoder_output_state"], tuple)
assert encoder_out["encoder_output_state"][0].shape == (
BATCH_SIZE, TEST_STATE_SIZE)
assert encoder_out["encoder_output_state"][1].shape == (
BATCH_SIZE, TEST_STATE_SIZE)
else:
assert isinstance(encoder_out["encoder_output_state"],
torch.Tensor)
assert encoder_out["encoder_output_state"].shape == (
BATCH_SIZE, TEST_STATE_SIZE)
elif enc_encoder == "stacked_cnn":
assert encoder_out["encoder_output"].shape[
1:] == encoder_obj.output_shape
elif enc_encoder == "embed":
assert (encoder_out["encoder_output"].shape == (BATCH_SIZE, SEQ_SIZE,
TEST_EMBEDDING_SIZE)
if enc_reduce_output is None else
(BATCH_SIZE, TEST_EMBEDDING_SIZE))
elif enc_encoder == "transformer":
assert encoder_out["encoder_output"].shape[
1:] == encoder_obj.output_shape
elif enc_encoder == "passthrough":
assert (encoder_out["encoder_output"].shape
== (BATCH_SIZE, SEQ_SIZE, 1) if enc_reduce_output is None else
(BATCH_SIZE, 1))
else:
raise ValueError(f"{enc_encoder} is an invalid encoder specification")
