def forward(self, x):
x0 = self.linear0(x[0])
x1 = self.linear1(x[1])
if self.dropout_input:
x0 = F.dropout(x0, p=self.dropout_input, training=self.training)
x1 = F.dropout(x1, p=self.dropout_input, training=self.training)
x0_chunks = get_chunks(x0, self.sizes_list)
x1_chunks = get_chunks(x1, self.sizes_list)
zs = []
for chunk_id, bilinear in enumerate(self.bilinears):
x0_c = x0_chunks[chunk_id]
x1_c = x1_chunks[chunk_id]
z = bilinear(x0_c, x1_c)
if self.pos_norm == "before_cat":
z = torch.sqrt(F.relu(z)) - torch.sqrt(F.relu(-z))
z = F.normalize(z, p=2)
zs.append(z)
z = torch.cat(zs, 1)
if self.pos_norm == "after_cat":
z = torch.sqrt(F.relu(z)) - torch.sqrt(F.relu(-z))
z = F.normalize(z, p=2)
if self.dropout_pre_lin > 0:
z = F.dropout(z, p=self.dropout_pre_lin, training=self.training)
z = self.linear_out(z)
if self.dropout_output > 0:
z = F.dropout(z, p=self.dropout_output, training=self.training)
return z
def forward(self, x):
x0 = self.linear0(x[0])
x1 = self.linear1(x[1])
bsize = x1.size(0)
if self.dropout_input > 0:
x0 = F.dropout(x0, p=self.dropout_input, training=self.training)
x1 = F.dropout(x1, p=self.dropout_input, training=self.training)
x0_chunks = get_chunks(x0, self.sizes_list)
x1_chunks = get_chunks(x1, self.sizes_list)
zs = []
for chunk_id, m0, m1 in zip(range(len(self.sizes_list)),
self.merge_linears0, self.merge_linears1):
x0_c = x0_chunks[chunk_id]
x1_c = x1_chunks[chunk_id]
m = m0(x0_c) * m1(x1_c) # bsize x split_size*rank
m = m.view(bsize, self.rank, -1)
z = torch.sum(m, 1)
if self.pos_norm == "before_cat":
z = torch.sqrt(F.relu(z)) - torch.sqrt(F.relu(-z))
z = F.normalize(z, p=2)
zs.append(z)
z = torch.cat(zs, 1)
if self.pos_norm == "after_cat":
z = torch.sqrt(F.relu(z)) - torch.sqrt(F.relu(-z))
z = F.normalize(z, p=2)
if self.dropout_pre_lin > 0:
z = F.dropout(z, p=self.dropout_pre_lin, training=self.training)
z = self.linear_out(z)
if self.dropout_output > 0:
z = F.dropout(z, p=self.dropout_output, training=self.training)
return z