Merge pull request #2269 from modelscope/dev_sx2

Shi Xian

2024-12-05 22b928dd3ff37ccee57ab2b5c2e4fcda4d33d24d

 funasr/models/sond/pooling/pooling_layers.py

@@ -59,8 +59,8 @@


class ASTP(nn.Module):
    """ Attentive statistics pooling: Channel- and context-dependent
        statistics pooling, first used in ECAPA_TDNN.
    """Attentive statistics pooling: Channel- and context-dependent
    statistics pooling, first used in ECAPA_TDNN.
    """

    def __init__(self, in_dim, bottleneck_dim=128, global_context_att=False):
@@ -71,14 +71,15 @@
        # need to transpose inputs.
        if global_context_att:
            self.linear1 = nn.Conv1d(
                in_dim * 3, bottleneck_dim,
                kernel_size=1)  # equals W and b in the paper
                in_dim * 3, bottleneck_dim, kernel_size=1
            )  # equals W and b in the paper
        else:
            self.linear1 = nn.Conv1d(
                in_dim, bottleneck_dim,
                kernel_size=1)  # equals W and b in the paper
        self.linear2 = nn.Conv1d(bottleneck_dim, in_dim,
                                 kernel_size=1)  # equals V and k in the paper
                in_dim, bottleneck_dim, kernel_size=1
            )  # equals W and b in the paper
        self.linear2 = nn.Conv1d(
            bottleneck_dim, in_dim, kernel_size=1
        )  # equals V and k in the paper

    def forward(self, x):
        """
@@ -92,17 +93,15 @@

        if self.global_context_att:
            context_mean = torch.mean(x, dim=-1, keepdim=True).expand_as(x)
            context_std = torch.sqrt(
                torch.var(x, dim=-1, keepdim=True) + 1e-10).expand_as(x)
            context_std = torch.sqrt(torch.var(x, dim=-1, keepdim=True) + 1e-10).expand_as(x)
            x_in = torch.cat((x, context_mean, context_std), dim=1)
        else:
            x_in = x

        # DON'T use ReLU here! ReLU may be hard to converge.
        alpha = torch.tanh(
            self.linear1(x_in))  # alpha = F.relu(self.linear1(x_in))
        alpha = torch.tanh(self.linear1(x_in))  # alpha = F.relu(self.linear1(x_in))
        alpha = torch.softmax(self.linear2(alpha), dim=2)
        mean = torch.sum(alpha * x, dim=2)
        var = torch.sum(alpha * (x ** 2), dim=2) - mean ** 2
        var = torch.sum(alpha * (x**2), dim=2) - mean**2
        std = torch.sqrt(var.clamp(min=1e-10))
        return torch.cat([mean, std], dim=1)