def __init__(self, *, config: Dict[str, Any]):
"""
Initialize AttentionLSTMDecoder.
:param config: Configuration of AttentionLSTMDecoder.
Example:
{
"hidden_dimension": 300,
"embedding_dimension": 300,
"vocabulary_size": 100000,
"encoder_output_dimension": 2048,
"number_of_lstm_layers": 1,
"unk_token_id": 0,
"bos_token_id": 1,
"eos_token_id": 2,
"language": "en"
}
"""
super().__init__()
self.config = config
self.tokenizer = bpemb.BPEmb(lang=self.config["language"],
vs=self.config["vocabulary_size"],
dim=self.config["embedding_dimension"])
self.embedding = torch.nn.Embedding.from_pretrained(
torch.tensor(self.tokenizer.vectors))
self.lstm = LSTM(
input_size=self.config["embedding_dimension"],
hidden_size=self.config["hidden_dimension"],
num_layers=self.config["number_of_lstm_layers"],
vocabulary_size=self.config["vocabulary_size"],
context_vector_size=self.config["encoder_output_dimension"])
self.hidden_initializer = torch.nn.Linear(
in_features=self.config["encoder_output_dimension"],
out_features=self.config["hidden_dimension"])
self.cell_initializer = torch.nn.Linear(
in_features=self.config["encoder_output_dimension"],
out_features=self.config["hidden_dimension"])
self.energy_function = torch.nn.Linear(
in_features=(self.config["encoder_output_dimension"] +
self.config["hidden_dimension"]),
out_features=1)
self.style_embedding = torch.nn.Embedding(
num_embeddings=3, embedding_dim=self.config["embedding_dimension"])
self.initialize_parameters()
@log_init
def __init__(self,
hdf5_path: Path,
raw_data_group_names: Dict[str, str],
language: str = "en",
vocabulary_size: int = 100_000,
embedding_dimensionality: int = 300,
force_update: bool = False):
self.logger = logging.getLogger(self.__class__.__name__)
self.hdf5_path = hdf5_path
self.raw_data_group_names = raw_data_group_names
self.language = language
self.vocabulary_size = vocabulary_size
self.embedding_dimensionality = embedding_dimensionality
self.tokenizer = bpemb.BPEmb(lang=self.language,
vs=self.vocabulary_size,
dim=self.embedding_dimensionality)
self.force_update = force_update
@log_run
def run(self) -> None:
if self.cache_exists() and not self.force_update:
self.logger.info(
"Cached version of tokenized data already exists. " +
"Skipping tokenization.")
return None
for hdf5_group_name in self.raw_data_group_names.values():
self.tokenize(
input_hdf5_group=hdf5_group_name,
input_hdf5_dataset="caption_cleaned",
def __init__(self, dim, vocab_size, lang='en'):
self.bpemb = bpemb.BPEmb(lang=lang, dim=dim, vs=vocab_size)
self.dim = dim
self.vectors = torch.from_numpy(self.bpemb.vectors)