def __init__(self,
input_size,
num_tree_layers,
single_hidden_size,
depth_embedding_dim=None,
normalize=False,
dropout=0.):
super().__init__()
self.input_size = input_size
self.num_tree_layers = num_tree_layers
self.single_hidden_size = single_hidden_size
self.depth_embedding_dim = depth_embedding_dim
self.normalize = normalize
self.dropout = dropout
self.depths_dim = self.num_tree_layers
if self.depth_embedding_dim is not None:
self.depths_dim = self.depth_embedding_dim
self.depth_embeddings = nn.Linear(
in_features=self.num_tree_layers,
out_features=self.depth_embedding_dim,
bias=False)
init_layers_uniform(-0.05, 0.05, [self.depth_embeddings])
if self.normalize:
self.norm = NormalizationLayer(features_num=self.input_size +
self.depths_dim)
self.layered_recurrent = LSTMCellDropout(
input_size=self.input_size + self.depths_dim,
hidden_size=self.single_hidden_size,
dropout=dropout)
def __init__(self, num_embeddings, embedding_dim, sparse=False):
super().__init__()
self.sparse = sparse
self.model = nn.Embedding(num_embeddings=num_embeddings,
embedding_dim=embedding_dim,
sparse=sparse)
init_layers_uniform(min_value=-0.1, max_value=0.1, layers=[self.model])
def __init__(self, method, hidden_size):
super(Attn, self).__init__()
self.method = method
self.hidden_size = hidden_size
if self.method == 'general':
self.attn = nn.Linear(self.hidden_size, self.hidden_size)
init_layers_uniform(-0.05, 0.05, [self.attn])
def __init__(self, input_size, output_size, bias=True):
super().__init__()
self.input_size = input_size
self.output_size = output_size
self.affine = nn.Linear(in_features=self.input_size,
out_features=self.output_size,
bias=bias)
init_layers_uniform(min_value=-0.05,
max_value=0.05,
layers=[self.affine])
def __init__(self, features_num):
super().__init__()
self.features_num = features_num
self.norm = torch.nn.BatchNorm1d(features_num)
init_layers_uniform(-0.05, 0.05, [self.norm])