def forward(self, data: BatchHolder):
if self.use_attention:
output = data.hidden
mask = data.masks
attn = self.attention(data.seq, output, mask)
if self.use_regulariser_attention:
data.reg_loss = 5 * self.regularizer_attention.regularise(
data.seq, output, mask, attn)
if isTrue(data, 'detach'):
attn = attn.detach()
if isTrue(data, 'permute'):
permutation = data.generate_permutation()
attn = torch.gather(attn, -1,
torch.LongTensor(permutation).to(device))
context = (attn.unsqueeze(-1) * output).sum(1)
data.attn = attn
else:
context = data.last_hidden
predict = self.decode(context)
data.predict = predict
def forward(self, data: BatchHolder):
if self.use_attention:
output = data.hidden
attn = data.generate_frozen_uniform_attn()
context = (attn.unsqueeze(-1) * output).sum(1)
data.attn = attn
else:
context = data.last_hidden
predict = self.decode(context)
data.predict = predict
def forward(self, data: BatchHolder):
if self.use_attention:
output = data.hidden
attn = data.target_attn
context = (attn.unsqueeze(-1) * output).sum(1)
data.attn = attn
else:
context = data.last_hidden
predict = self.decode(context)
data.predict = predict
def forward(self, data: BatchHolder):
if self.use_attention:
output = data.hidden
mask = data.masks
attn = self.attention(data.seq, output, mask)
context = (attn.unsqueeze(-1) * output).sum(1)
data.attn = attn
else:
context = data.last_hidden
predict = self.decode(context)
data.predict = predict