def forward(self, xs, xlens, batch_first=True):
"""Forward pass.
Args:
xs (FloatTensor): `[B, T, F]` or `[T, B, F]`
xlens (IntTensor): `[B]` (on CPU)
batch_first (bool): operate batch-first tensor
Returns:
xs (FloatTensor): `[B, T', F']` or `[T', B, F']`
xlens (IntTensor): `[B]` (on CPU)
"""
if self.factor == 1:
return xs, xlens
if batch_first:
xs = self.conv1d(xs.transpose(2, 1))
xs = xs.transpose(2, 1).contiguous()
else:
xs = self.conv1d(xs.permute(1, 2, 0))
xs = xs.permute(2, 0, 1).contiguous()
xs = torch.relu(xs)
xlens = update_lens_1d(xlens, self.conv1d)
return xs, xlens
def forward(self, xs, xlens):
if self.factor == 1:
return xs, xlens
xs = self.pool(xs.transpose(2, 1)).transpose(2, 1).contiguous()
xlens = update_lens_1d(xlens, self.pool)
return xs, xlens
def forward(self, xs, xlens):
"""Forward pass.
Args:
xs (FloatTensor): `[B, T, F]`
xlens (IntTensor): `[B]` (on CPU)
Returns:
xs (FloatTensor): `[B, T', F']`
xlens (IntTensor): `[B]` (on CPU)
"""
if self.subsampling_factor == 1:
return xs, xlens
xs = self.pool(xs.transpose(2, 1)).transpose(2, 1).contiguous()
xlens = update_lens_1d(xlens, self.pool)
return xs, xlens
def forward(self, xs, xlens):
"""Forward pass.
Args:
xs (FloatTensor): `[B, C_i, T, F]`
xlens (IntTensor): `[B]`
Returns:
xs (FloatTensor): `[B, C_o, T, F]`
xlens (IntTensor): `[B]`
"""
B, C, T, F = xs.size()
xs = xs.transpose(3, 2).view(B, C * F, T)
xs = self.dropout(torch.relu(self.conv1d(xs)))
xs = xs.view(B, self.C_out, F, -1).transpose(3, 2)
xs = self.norm(xs)
xlens = update_lens_1d(xlens, self.conv1d)
return xs, xlens