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 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
def __init__(self,
word_vec_size,
hidden_size,
word_mat,
pool='max',
dropout=0.2,
no_wordvec_layer=False):
super(BOWModel, self).__init__()
self.pool = pool
self.no_wordvec_layer = no_wordvec_layer
self.embs = ibp.Embedding.from_pretrained(word_mat)
if no_wordvec_layer:
self.linear_hidden = ibp.Linear(word_vec_size, hidden_size)
else:
self.linear_input = ibp.Linear(word_vec_size, hidden_size)
self.linear_hidden = ibp.Linear(hidden_size, hidden_size)
self.linear_output = ibp.Linear(hidden_size, 1)
self.dropout = ibp.Dropout(dropout)
if self.pool == 'attn':
self.attn_pool = ibp.Linear(hidden_size, 1)
def __init__(self,
device,
num_vocabs,
x_size,
h_size,
num_classes,
dropout,
cell_type='nary',
pretrained_emb=None,
no_wordvec_layer=False,
perturbation=None):
super(TreeLSTMDP, self).__init__()
assert perturbation is not None
# TODO: to implement more discrete perturbation space
self.device = device
self.perturbation = perturbation
self.adv_attack = False
self.out_x_vecs = None
self.adv_attack_Ins = False
self.out_ioux_c = [None, None]
self.deltas = Perturbation.str2deltas(perturbation)
self.deltas_p1 = [delta + 1 for delta in self.deltas]
self.x_size = x_size
self.h_size = h_size
self.embedding = ibp.Embedding(num_vocabs, x_size)
if pretrained_emb is not None:
print('Using glove')
self.embedding.weight.data.copy_(pretrained_emb)
self.embedding.weight.requires_grad = True
self.dropout = ibp.Dropout(dropout)
self.linear = ibp.Linear(h_size, num_classes)
cell = TreeLSTMCellDP if cell_type == 'nary' else None
if cell is None:
raise NotImplementedError
self.no_wordvec_layer = no_wordvec_layer
if no_wordvec_layer:
self.cell = cell(x_size, h_size, self.deltas_p1, device)
else:
self.linear_input = ibp.Linear(x_size, h_size)
self.cell = cell(h_size, h_size, self.deltas_p1, device)
def __init__(self,
word_vec_size,
hidden_size,
word_mat,
device,
dropout=0.2,
no_wordvec_layer=False):
super(LSTMFinalStateModel, self).__init__()
self.hidden_size = hidden_size
self.no_wordvec_layer = no_wordvec_layer
self.device = device
self.embs = ibp.Embedding.from_pretrained(word_mat)
if no_wordvec_layer:
self.lstm = ibp.LSTM(word_vec_size,
hidden_size,
bidirectional=True)
else:
self.linear_input = ibp.Linear(word_vec_size, hidden_size)
self.lstm = ibp.LSTM(hidden_size, hidden_size)
self.dropout = ibp.Dropout(dropout)
self.fc_hidden = ibp.Linear(hidden_size, hidden_size)
self.fc_output = ibp.Linear(hidden_size, 1)
def __init__(self,
word_vec_size,
hidden_size,
kernel_size,
word_mat,
pool='max',
dropout=0.2,
no_wordvec_layer=False,
early_ibp=False,
relu_wordvec=True,
unfreeze_wordvec=False):
super(CNNModel, self).__init__()
cnn_padding = (kernel_size - 1) // 2 # preserves size
self.pool = pool
# Ablations
self.no_wordvec_layer = no_wordvec_layer
self.early_ibp = early_ibp
self.relu_wordvec = relu_wordvec
self.unfreeze_wordvec = False
# End ablations
self.embs = ibp.Embedding.from_pretrained(
word_mat, freeze=not self.unfreeze_wordvec)
if no_wordvec_layer:
self.conv1 = ibp.Conv1d(word_vec_size,
hidden_size,
kernel_size,
padding=cnn_padding)
else:
self.linear_input = ibp.Linear(word_vec_size, hidden_size)
self.conv1 = ibp.Conv1d(hidden_size,
hidden_size,
kernel_size,
padding=cnn_padding)
if self.pool == 'attn':
self.attn_pool = ibp.Linear(hidden_size, 1)
self.dropout = ibp.Dropout(dropout)
self.fc_hidden = ibp.Linear(hidden_size, hidden_size)
self.fc_output = ibp.Linear(hidden_size, 1)