(
self,
embed_dim: int = 1024,
num_layers: int = 12,
num_heads: int = 16,
dropout: float = 0.1,
)
| 224 | # Streaming Vocos backbone based on Transformer layers |
| 225 | class CausalVocosBackbone(nn.Module): |
| 226 | def __init__( |
| 227 | self, |
| 228 | embed_dim: int = 1024, |
| 229 | num_layers: int = 12, |
| 230 | num_heads: int = 16, |
| 231 | dropout: float = 0.1, |
| 232 | ): |
| 233 | super().__init__() |
| 234 | self.in_proj = CausalConv1d(embed_dim, embed_dim, kernel_size=7) |
| 235 | self.prior_net = nn.Sequential( |
| 236 | CausalResnetBlock(embed_dim, embed_dim, dropout=dropout), |
| 237 | CausalResnetBlock(embed_dim, embed_dim, dropout=dropout), |
| 238 | ) |
| 239 | self.transformers = nn.ModuleList( |
| 240 | [WhisperEncoderLayer(embed_dim, num_heads) for _ in range(num_layers)] |
| 241 | ) |
| 242 | self.post_net = nn.Sequential( |
| 243 | CausalResnetBlock(embed_dim, embed_dim, dropout=dropout), |
| 244 | CausalResnetBlock(embed_dim, embed_dim, dropout=dropout), |
| 245 | ) |
| 246 | self.final_norm = nn.LayerNorm(embed_dim, eps=1e-6) |
| 247 | |
| 248 | def forward( |
| 249 | self, |
nothing calls this directly
no test coverage detected