def forward(self, top_buf, top_stack_1, top_stack_2):
if self.tracking_ln:
top_buf = self.buf_ln(top_buf)
top_stack_1 = self.stack1_ln(top_stack_1)
top_stack_2 = self.stack2_ln(top_stack_2)
if self.lateral_tracking:
tracker_inp = self.buf(top_buf)
tracker_inp += self.stack1(top_stack_1)
tracker_inp += self.stack2(top_stack_2)
batch_size = tracker_inp.size(0)
if self.h is not None:
tracker_inp += self.lateral(self.h)
if self.c is None:
self.c = to_gpu(
Variable(torch.from_numpy(
np.zeros((batch_size, self.state_size),
dtype=np.float32)),
volatile=tracker_inp.volatile))
# Run tracking lstm.
self.c, self.h = lstm(self.c, tracker_inp)
return self.h, self.c
else:
return torch.cat([top_buf, top_stack_1, top_stack_2], 1), None
def forward(self, top_buf, top_stack_1, top_stack_2):
tracker_inp = self.buf(top_buf.h)
tracker_inp += self.stack1(top_stack_1.h)
tracker_inp += self.stack2(top_stack_2.h)
batch_size = tracker_inp.size(0)
if self.lateral_tracking:
if self.h is not None:
tracker_inp += self.lateral(self.h)
if self.c is None:
self.c = to_gpu(Variable(torch.from_numpy(
np.zeros((batch_size, self.state_size),
dtype=np.float32)),
volatile=tracker_inp.volatile))
# Run tracking lstm.
self.c, self.h = lstm(self.c, tracker_inp)
return self.h, self.c
else:
outp = self.transform(tracker_inp)
return outp, None