| | |
|---|
| | | import funasr.models.lora.layers as lora |
|---|
| | | |
|---|
| | | |
|---|
| | | |
|---|
| | | class CosineDistanceAttention(nn.Module): |
|---|
| | | """ Compute Cosine Distance between spk decoder output and speaker profile |
|---|
| | | Args: |
|---|
| | |
|---|
| | | if profile_lens is not None: |
|---|
| | | |
|---|
| | | mask = (make_pad_mask(profile_lens)[:, None, :]).to(profile.device) |
|---|
| | | min_value = float( |
|---|
| | | numpy.finfo(torch.tensor(0, dtype=x.dtype).numpy().dtype).min |
|---|
| | | min_value = float(numpy.finfo(torch.tensor(0, dtype=x.dtype).numpy().dtype).min) |
|---|
| | | weights_not_softmax = F.cosine_similarity(x, profile.unsqueeze(1), dim=-1).masked_fill( |
|---|
| | | mask, min_value |
|---|
| | | ) |
|---|
| | | weights_not_softmax=F.cosine_similarity(x, profile.unsqueeze(1), dim=-1).masked_fill(mask, min_value) |
|---|
| | | weights = self.softmax(weights_not_softmax).masked_fill(mask, 0.0) # (B, L, N) |
|---|
| | | else: |
|---|
| | | x = x[:, -1:, :, :] |
|---|
