def forward(self, inputs, input_length):
#Args:
# inputs: padding data of batch frames, [batch, max_seq_len, feat_dim]
# inputs_length: length of padding batch data, [batch]
enc_mask = get_enc_padding_mask(inputs, input_length)
enc_output, enc_mask = self.embed(inputs, enc_mask)
# enc_output: with shape [batch, max_seq_len, emd_dim],
# used input for K and V of decoder src-attention,
# so its shape is same as embedding of decoder
# enc_mask: with shape [batch, 1, max_seq_len]
# [1, max_seq_len] used to mask soft_max value
# mask padding data of embed output use enc_mask
enc_output.masked_fill_(~enc_mask.transpose(1, 2), 0.0)
for _, block in enumerate(self.blocks):
enc_output, enc_mask = block(enc_output, enc_mask)
# mask padding data of embed output use enc_mask
enc_output.masked_fill_(~enc_mask.transpose(1, 2), 0.0)
if self.normalize_before:
enc_output = self.after_norm(enc_output)
return enc_output, enc_mask
def forward(self, inputs, input_length):
enc_mask = get_enc_padding_mask(inputs, input_length)
enc_output, enc_mask = self.embed(inputs, enc_mask)
enc_output.masked_fill_(~enc_mask.transpose(1, 2), 0.0)
for _, block in enumerate(self.blocks):
enc_output, enc_mask = block(enc_output, enc_mask)
enc_output.masked_fill_(~enc_mask.transpose(1, 2), 0.0)
if self.normalize_before:
enc_output = self.after_norm(enc_output)
return enc_output, enc_mask
def forward(self, inputs, input_length, streaming=False):
enc_mask = get_enc_padding_mask(inputs, input_length)
enc_output, enc_mask = self.embed(inputs, enc_mask)
enc_output.masked_fill_(~enc_mask.transpose(1, 2), 0.0)
if streaming:
length = torch.sum(enc_mask.squeeze(1), dim=-1)
enc_mask = get_streaming_mask(enc_output,
length,
left_context=20,
right_context=0)
for _, block in enumerate(self.blocks):
enc_output, enc_mask = block(enc_output, enc_mask)
enc_output.masked_fill_(~enc_mask.transpose(1, 2), 0.0)
if self.normalize_before:
enc_output = self.after_norm(enc_output)
return enc_output, enc_mask