def from_config(cls, model_config, feature_config, metadata: CommonMetadata):
if model_config.compositional_type == RNNGParser.Config.CompositionalType.SUM:
p_compositional = CompositionalSummationNN(
lstm_dim=model_config.lstm.lstm_dim
)
elif (
model_config.compositional_type == RNNGParser.Config.CompositionalType.BLSTM
):
p_compositional = CompositionalNN(lstm_dim=model_config.lstm.lstm_dim)
else:
raise ValueError(
"Cannot understand compositional flag {}".format(
model_config.compositional_type
)
)
return cls(
ablation=model_config.ablation,
constraints=model_config.constraints,
lstm_num_layers=model_config.lstm.num_layers,
lstm_dim=model_config.lstm.lstm_dim,
max_open_NT=model_config.max_open_NT,
dropout=model_config.dropout,
actions_vocab=metadata.actions_vocab,
shift_idx=metadata.shift_idx,
reduce_idx=metadata.reduce_idx,
ignore_subNTs_roots=metadata.ignore_subNTs_roots,
valid_NT_idxs=metadata.valid_NT_idxs,
valid_IN_idxs=metadata.valid_IN_idxs,
valid_SL_idxs=metadata.valid_SL_idxs,
embedding=Model.create_embedding(feature_config, metadata=metadata),
p_compositional=p_compositional,
)
def test_CompositionFunction(self):
lstm_dim = 100
embedding = torch.ones(1, lstm_dim)
input_sequence = [embedding for _ in range(10)]
compositionalNN = CompositionalNN(lstm_dim)
self.assertEqual(compositionalNN(input_sequence).shape, embedding.shape)
compositionalSummationNN = CompositionalSummationNN(lstm_dim)
self.assertEqual(
compositionalSummationNN(input_sequence).shape, embedding.shape
)
def from_config(
cls,
model_config,
feature_config=None,
metadata: CommonMetadata = None,
tensorizers: Dict[str, Tensorizer] = None,
):
if model_config.compositional_type == RNNGParser.Config.CompositionalType.SUM:
p_compositional = CompositionalSummationNN(
lstm_dim=model_config.lstm.lstm_dim)
elif (model_config.compositional_type ==
RNNGParser.Config.CompositionalType.BLSTM):
p_compositional = CompositionalNN(
lstm_dim=model_config.lstm.lstm_dim)
else:
raise ValueError("Cannot understand compositional flag {}".format(
model_config.compositional_type))
if tensorizers is not None:
embedding = EmbeddingList(
[
create_module(model_config.embedding,
tensorizer=tensorizers["tokens"])
],
concat=True,
)
actions_params = tensorizers["actions"]
actions_vocab = actions_params.vocab
else:
embedding = Model.create_embedding(feature_config,
metadata=metadata)
actions_params = metadata
actions_vocab = metadata.actions_vocab
return cls(
ablation=model_config.ablation,
constraints=model_config.constraints,
lstm_num_layers=model_config.lstm.num_layers,
lstm_dim=model_config.lstm.lstm_dim,
max_open_NT=model_config.max_open_NT,
dropout=model_config.dropout,
actions_vocab=actions_vocab,
shift_idx=actions_params.shift_idx,
reduce_idx=actions_params.reduce_idx,
ignore_subNTs_roots=actions_params.ignore_subNTs_roots,
valid_NT_idxs=actions_params.valid_NT_idxs,
valid_IN_idxs=actions_params.valid_IN_idxs,
valid_SL_idxs=actions_params.valid_SL_idxs,
embedding=embedding,
p_compositional=p_compositional,
)
def setUp(self):
actions_counter = Counter()
for action in [
"IN:A",
"IN:B",
"IN:UNSUPPORTED",
"REDUCE",
"SHIFT",
"SL:C",
"SL:D",
]:
actions_counter[action] += 1
actions_vocab = Vocab(actions_counter, specials=[])
self.parser = RNNGParser(
ablation=RNNGParser.Config.AblationParams(),
constraints=RNNGParser.Config.RNNGConstraints(),
lstm_num_layers=2,
lstm_dim=20,
max_open_NT=10,
dropout=0.2,
beam_size=3,
top_k=3,
actions_vocab=actions_vocab,
shift_idx=4,
reduce_idx=3,
ignore_subNTs_roots=[2],
valid_NT_idxs=[0, 1, 2, 5, 6],
valid_IN_idxs=[0, 1, 2],
valid_SL_idxs=[5, 6],
embedding=EmbeddingList(
embeddings=[
WordEmbedding(
num_embeddings=5,
embedding_dim=20,
embeddings_weight=None,
init_range=[-1, 1],
unk_token_idx=4,
mlp_layer_dims=[],
),
DictEmbedding(
num_embeddings=4, embed_dim=10, pooling_type=PoolingType.MEAN
),
],
concat=True,
),
p_compositional=CompositionalNN(lstm_dim=20),
)
self.parser.train()
def setUp(self):
contextual_emb_dim = 1
emb_module = EmbeddingList(
embeddings=[
WordEmbedding(num_embeddings=103, embedding_dim=100),
DictEmbedding(
num_embeddings=59, embed_dim=10, pooling_type=PoolingType.MEAN
),
ContextualTokenEmbedding(contextual_emb_dim),
],
concat=True,
)
self.training_model = RNNGModel(
input_for_trace=RNNGModel.get_input_for_trace(contextual_emb_dim),
embedding=emb_module,
ablation=RNNGParser.Config.AblationParams(),
constraints=RNNGParser.Config.RNNGConstraints(),
lstm_num_layers=2,
lstm_dim=32,
max_open_NT=10,
dropout=0.4,
num_actions=20,
shift_idx=0,
reduce_idx=1,
ignore_subNTs_roots=[8, 15],
valid_NT_idxs=[2, 3, 4, 5, 6, 7, 8, 9, 10, 11]
+ [12, 13, 14, 15, 16, 17, 18, 19],
valid_IN_idxs=[2, 4, 7, 8, 10, 12, 13, 14, 15],
valid_SL_idxs=[3, 5, 6, 9, 11, 16, 17, 18, 19],
embedding_dim=emb_module.embedding_dim,
p_compositional=CompositionalNN(lstm_dim=32, device="cpu"),
device="cpu",
)
self.training_model.train()
self.inference_model = RNNGInference(
self.training_model.trace_embedding(),
self.training_model.jit_model,
MockVocab(["<unk>", "foo", "bar"]),
MockVocab(["<unk>", "a", "b"]),
MockVocab(["SHIFT", "REDUCE", "IN:END_CALL", "SL:METHOD_CALL"]),
)
self.inference_model.eval()
def from_config(cls, model_config, feature_config,
metadata: CommonMetadata):
device = ("cuda:{}".format(torch.cuda.current_device())
if cuda.CUDA_ENABLED else "cpu")
if model_config.compositional_type == RNNGParser.Config.CompositionalType.SUM:
p_compositional = CompositionalSummationNN(
lstm_dim=model_config.lstm.lstm_dim)
elif (model_config.compositional_type ==
RNNGParser.Config.CompositionalType.BLSTM):
p_compositional = CompositionalNN(
lstm_dim=model_config.lstm.lstm_dim, device=device)
else:
raise ValueError("Cannot understand compositional flag {}".format(
model_config.compositional_type))
emb_module = Model.create_embedding(feature_config, metadata=metadata)
contextual_emb_dim = feature_config.contextual_token_embedding.embed_dim
return cls(
cls.get_input_for_trace(contextual_emb_dim),
embedding=emb_module,
ablation=model_config.ablation,
constraints=model_config.constraints,
lstm_num_layers=model_config.lstm.num_layers,
lstm_dim=model_config.lstm.lstm_dim,
max_open_NT=model_config.max_open_NT,
dropout=model_config.dropout,
num_actions=len(metadata.actions_vocab),
shift_idx=metadata.shift_idx,
reduce_idx=metadata.reduce_idx,
ignore_subNTs_roots=metadata.ignore_subNTs_roots,
valid_NT_idxs=metadata.valid_NT_idxs,
valid_IN_idxs=metadata.valid_IN_idxs,
valid_SL_idxs=metadata.valid_SL_idxs,
embedding_dim=emb_module.embedding_dim,
p_compositional=p_compositional,
device=device,
)
