def forward(self, inputs):
x, seq_lens = inputs
for block in self.blocks:
x, seq_lens = block(x, seq_lens)
x, _ = self.global_pool(x, seq_lens)
x = L.flatten(x) # B x F*k
# Run the fc layer if we have one
for fc_layer in self.fc_layers:
x = fc_layer(x)
self.loss_in = x
return self.loss_in
def legacy_forward(self, x):
# Apply the resampler if necessary
if self.resample:
x = self.resampler(x)
for rnn_layer in self.rnn_layers:
x = rnn_layer(x) # B x Li x H
x = self.pool(x)
# For testing purposes only
# x = self.att(x)
x = L.flatten(x) # B x k*H
for fc_layer in self.fc_layers:
x = fc_layer(x)
self.loss_in = x # B x 6
return self.loss_in
def forward(self, inputs):
x, seq_lens = inputs
# Do the conv layers
for conv in self.conv_layers:
x, seq_lens = conv(x, seq_lens)
x, seq_lens = self.pool(x, seq_lens)
# Do the global pooling
x, seq_lens = self.global_pool(x, seq_lens)
x = L.flatten(x) # B x k*H
for fc_layer in self.fc_layers:
x = fc_layer(x)
self.loss_in = x
return self.loss_in
def forward(self, x):
x, seq_lens = x
# Apply the resampler if necessary
if self.resample:
x = self.resampler(x)
# Run the rnns
for rnn_layer_ind in range(0, len(self.rnn_layers) - 1):
x = torch.cat([x, self.rnn_layers[rnn_layer_ind](x)],
dim=-1) # B x L x (Hx + Hout)
# Final rnn
x = self.rnn_layers[-1](x)
# Apply the mask
x = L.unpad_sequences(x, seq_lens)
# Run the pooling and flatten
x = self.pool(x)
x = L.flatten(x) # B x k*H
for fc_layer in self.fc_layers:
x = fc_layer(x)
self.loss_in = x # B x 6
return self.loss_in
def forward(self, x):
if self.legacy:
return self.legacy_forward(x)
x, seq_lens = x
# Apply the resampler if necessary
if self.resample:
x = self.resampler(x)
for rnn_layer in self.rnn_layers:
x = rnn_layer(x) # B x Li x H
# Apply the mask
if self.use_pool:
x = L.unpad_sequences(x, seq_lens)
# Do the pooling
if self.use_multi_pool:
x = self.concat([pool_i(x) for pool_i in self.pool])
else:
x = self.pool(x)
x = L.flatten(x) # B x k*H
for fc_layer in self.fc_layers:
x = fc_layer(x)
self.loss_in = x # B x 6
# return F.sigmoid(self.loss_in)
return self.loss_in