def forward_sample(self, sampleSoft, sampleHard, z, c, h):
if sampleSoft is not None:
# with sampleSoftIx (mbsize x vocabsize) gradients will pass through
emb = soft_embed(self.emb, sampleSoft)
else:
# with sampleIx (mbsize) indextensor, gradients dont pass through.
emb = self.emb(sampleHard)
# mb x (embdim + zdim + cdim)
emb = torch.cat([emb, z, c], 1)
# insert seqlen 1 (mbsize x 1 x ezcdim)
emb = emb.unsqueeze(1)
# 1 x mbsize x h_dim
output, h = self.rnn(emb, h)
# mbsize x h_dim
output = output.squeeze(1)
# apply skip connection
if self.skip_connetions:
latent_code = torch.cat([z, c], 1)
output = self.skip_weight_x(output) + self.skip_weight_z(
latent_code)
# [mbsize x self.n_vocab]
logits = self.fc(output)
return logits, h
def forward_encoder(self, inputs):
'''
Inputs is batch of sentences: seq_len x mbsize
or batch of soft sentences: seq_len x mbsize x n_vocab.
'''
if inputs.dim() == 2:
inputs = self.word_emb(inputs)
else: # dim == 3.
inputs = soft_embed(self.word_emb, inputs)
return self.encoder(inputs)
def forward_classifier(self, inputs):
"""
Inputs is batch of sentences: mbsize x seq_len
or batch of soft sentences: mbsize x seq_len x n_vocab.
"""
if inputs.dim() == 2:
inputs = self.word_emb(inputs)
else: # dim == 3.
inputs = soft_embed(self.word_emb, inputs)
return self.classifier(inputs)