def stepwise_embeddings(self, x):
if self.lstm.bidirectional:
sz = x.size()
batch = torch.zeros(sz[0],
sz[0],
sz[2],
dtype=torch.float,
device='cuda' if self.use_cuda else 'cpu')
lengths = np.array(np.arange(sz[0])[::-1] + 1)
for i, length in enumerate(lengths):
batch[:length, i, :] = x[:length, 0, :]
batch = torch.nn.utils.rnn.pack_padded_sequence(batch, lengths)
_, (h, _) = self.lstm(batch)
h_stepwise = hidden2fc_input(self.lstm, h)
else:
h_stepwise, (_, _) = self.lstm(x)
x = h_stepwise
for fc in self.fc:
x = fc(x)
# the flipping here comes not from the bidirectionality of the net,
# but because the batch must be ordered longest first for PyTorch
if self.lstm.bidirectional:
x = torch.flip(x, dims=[0])
return self.loss_delegate.mod_forward(x)
def embedding(self, x):
_, (h, _) = self.lstm(x) # hidden state defaults to zero
x = hidden2fc_input(self.lstm, h)
for fc in self.fc[:-1]:
x = fc(x)
return x
def embedding(self, x):
_, h = self.gru(x) # hidden state defaults to zero
x = hidden2fc_input(self.gru, h)
for fc in self.fc[:-1]:
x = fc(x)
return x
def forward(self, x):
data, lengths = x
net_input = torch.nn.utils.rnn.pack_padded_sequence(
data, lengths, batch_first=self.batch_first)
_, h = self.gru(net_input) # hidden state defaults to zero
x = hidden2fc_input(self.gru, h)
for fc in self.fc:
x = fc(x)
return x
def stepwise_embeddings(self, x):
if self.gru.bidirectional:
sz = x.size()
batch = torch.zeros(sz[0], sz[0], sz[2], dtype=torch.float, device='cuda' if self.use_cuda else 'cpu')
lengths = np.array(np.arange(sz[0])[::-1] + 1)
for i, length in enumerate(lengths):
batch[:length, i, :] = x[:length, 0, :]
batch = torch.nn.utils.rnn.pack_padded_sequence(batch, lengths)
_, h = self.gru(batch)
h_stepwise = hidden2fc_input(self.gru, h)
else:
# TODO: packed sequence handled here?
h_stepwise, _ = self.gru(x)
x = h_stepwise
for fc in self.fc:
x = fc(x)
# the batch must be ordered longest first for PyTorch, flipping it to restore the expected order
if self.gru.bidirectional:
x = torch.flip(x, dims=[0])
return self.loss_delegate.mod_forward(x)