def __init__(self,
word_vec_size,
hidden_size,
word_mat,
dropout_prob=0.1,
num_layers=3,
no_wordvec_layer=False):
super(BOWModel, self).__init__()
self.embs = ibp.Embedding.from_pretrained(word_mat)
self.rotation = ibp.Linear(word_vec_size, hidden_size)
self.sum_drop = ibp.Dropout(dropout_prob) if dropout_prob else None
layers = []
for i in range(num_layers):
layers.append(ibp.Linear(2 * hidden_size, 2 * hidden_size))
layers.append(ibp.Activation(F.relu))
if dropout_prob:
layers.append(ibp.Dropout(dropout_prob))
layers.append(ibp.Linear(2 * hidden_size, len(EntailmentLabels)))
layers.append(ibp.LogSoftmax(dim=1))
self.layers = nn.Sequential(*layers)
def __init__(self,
word_vec_size,
hidden_size,
word_mat,
dropout_prob=0.1,
num_layers=2,
no_wordvec_layer=False):
super(DecompAttentionModel, self).__init__()
self.embs = ibp.Embedding.from_pretrained(word_mat)
self.null = nn.Parameter(torch.normal(mean=torch.zeros(word_vec_size)))
self.rotation = None
hidden_size = word_vec_size
self.rotation = ibp.Linear(word_vec_size, hidden_size)
def get_feedforward_layers(num_layers, input_size, hidden_size,
output_size):
layers = []
for i in range(num_layers):
layer_in_size = input_size if i == 0 else hidden_size
layer_out_size = output_size if i == num_layers - 1 else hidden_size
if dropout_prob:
layers.append(ibp.Dropout(dropout_prob))
layers.append(ibp.Linear(layer_in_size, layer_out_size))
if i < num_layers - 1:
layers.append(ibp.Activation(F.relu))
return layers
ff_layers = get_feedforward_layers(num_layers, hidden_size,
hidden_size, 1)
self.feedforward = nn.Sequential(*ff_layers)
compare_layers = get_feedforward_layers(num_layers, 2 * hidden_size,
hidden_size, hidden_size)
self.compare_ff = nn.Sequential(*compare_layers)
output_layers = get_feedforward_layers(num_layers, 2 * hidden_size,
hidden_size, hidden_size)
output_layers.append(ibp.Linear(hidden_size, len(EntailmentLabels)))
output_layers.append(ibp.LogSoftmax(dim=1))
self.output_layer = nn.Sequential(*output_layers)