def __init__(self,
model_dim=None,
word_embedding_dim=None,
vocab_size=None,
initial_embeddings=None,
num_classes=None,
embedding_keep_rate=None,
use_sentence_pair=False,
classifier_keep_rate=None,
mlp_dim=None,
num_mlp_layers=None,
mlp_ln=None,
context_args=None,
gated=None,
selection_keep_rate=None,
pyramid_selection_keep_rate=None,
**kwargs):
super(Pyramid, self).__init__()
self.use_sentence_pair = use_sentence_pair
self.model_dim = model_dim
self.gated = gated
self.selection_keep_rate = selection_keep_rate
classifier_dropout_rate = 1. - classifier_keep_rate
args = Args()
args.size = model_dim
args.input_dropout_rate = 1. - embedding_keep_rate
vocab = Vocab()
vocab.size = initial_embeddings.shape[
0] if initial_embeddings is not None else vocab_size
vocab.vectors = initial_embeddings
self.embed = Embed(word_embedding_dim,
vocab.size,
vectors=vocab.vectors)
self.composition_fn = SimpleTreeLSTM(model_dim / 2,
composition_ln=False)
self.selection_fn = Linear(initializer=HeKaimingInitializer)(model_dim,
1)
# TODO: Set up layer norm.
mlp_input_dim = model_dim * 2 if use_sentence_pair else model_dim
self.mlp = MLP(mlp_input_dim, mlp_dim, num_classes, num_mlp_layers,
mlp_ln, classifier_dropout_rate)
self.encode = context_args.encoder
self.reshape_input = context_args.reshape_input
self.reshape_context = context_args.reshape_context
def __init__(self,
model_dim=None,
word_embedding_dim=None,
vocab_size=None,
initial_embeddings=None,
num_classes=None,
embedding_keep_rate=None,
use_sentence_pair=False,
classifier_keep_rate=None,
mlp_dim=None,
num_mlp_layers=None,
mlp_bn=None,
context_args=None,
**kwargs):
super(BaseModel, self).__init__()
self.use_sentence_pair = use_sentence_pair
self.model_dim = model_dim
classifier_dropout_rate = 1. - classifier_keep_rate
args = Args()
args.size = model_dim
args.input_dropout_rate = 1. - embedding_keep_rate
vocab = Vocab()
vocab.size = initial_embeddings.shape[
0] if initial_embeddings is not None else vocab_size
vocab.vectors = initial_embeddings
self.embed = Embed(word_embedding_dim,
vocab.size,
vectors=vocab.vectors)
self.rnn = nn.LSTM(args.size,
model_dim,
num_layers=1,
batch_first=True)
mlp_input_dim = model_dim * 2 if use_sentence_pair else model_dim
self.mlp = MLP(mlp_input_dim, mlp_dim, num_classes, num_mlp_layers,
mlp_bn, classifier_dropout_rate)
self.encode = context_args.encoder
self.reshape_input = context_args.reshape_input
self.reshape_context = context_args.reshape_context