MCPcopy Create free account
hub / github.com/FireRedTeam/FireRedTTS2 / forward

Method forward

fireredtts2/codec/decoder.py:198–221  ·  view source on GitHub ↗

Args: x: shape (b, t, c) x_lens: shape (b,)

(
        self,
        x: torch.Tensor,
        x_lens: torch.Tensor,
    )

Source from the content-addressed store, hash-verified

196 self.final_norm = nn.LayerNorm(embed_dim, eps=1e-6)
197
198 def forward(
199 self,
200 x: torch.Tensor,
201 x_lens: torch.Tensor,
202 ):
203 """
204 Args:
205 x: shape (b, t, c)
206 x_lens: shape (b,)
207 """
208 x = x.transpose(1, 2)
209 x = self.in_proj(x)
210 x = self.prior_net(x)
211 x = x.transpose(1, 2)
212
213 attention_mask = make_nonpad_mask(x_lens).unsqueeze(1) # (b, 1, t)
214 # NOTE(sfy): I think positional embedding is unnecessary
215 for layer in self.transformers:
216 x = layer(x, attention_mask)
217 x = x.transpose(1, 2)
218 x = self.post_net(x)
219 x = x.transpose(1, 2)
220 x = self.final_norm(x)
221 return x
222
223
224# Streaming Vocos backbone based on Transformer layers

Callers

nothing calls this directly

Calls 1

make_nonpad_maskFunction · 0.90

Tested by

no test coverage detected