Function graph_zipformer_layer

src/framework/audio/zipenhancer.cpp:692–711

}

core::TensorValue graph_zipformer_layer(
    core::ModuleBuildContext & ctx,
    std::vector<GraphConstant> & constants,
    const core::TensorValue & input,
    const std::unordered_map<std::string, Param> & params,
    const std::string & prefix) {
    const auto orig = input;
    auto attn = graph_attention_weights(ctx, constants, input, params, prefix);
    auto x = graph_add(ctx, input, graph_feed_forward(ctx, input, params, prefix, "feed_forward1", 192));
    x = graph_add(ctx, x, graph_nonlin_attention(ctx, x, attn.front(), params, prefix));
    x = graph_add(ctx, x, graph_self_attention(ctx, x, attn, params, prefix, "self_attn1"));
    x = graph_add(ctx, x, graph_conv_module(ctx, x, params, prefix, "conv_module1"));
    x = graph_add(ctx, x, graph_feed_forward(ctx, x, params, prefix, "feed_forward2", 256));
    x = graph_bypass(ctx, orig, x, require_param(params, prefix + ".bypass_mid.bypass_scale"));
    x = graph_add(ctx, x, graph_self_attention(ctx, x, attn, params, prefix, "self_attn2"));
    x = graph_add(ctx, x, graph_conv_module(ctx, x, params, prefix, "conv_module2"));
    x = graph_add(ctx, x, graph_feed_forward(ctx, x, params, prefix, "feed_forward3", 320));
    x = graph_bias_norm(ctx, x, params, prefix);
    return graph_bypass(ctx, orig, x, require_param(params, prefix + ".bypass.bypass_scale"));
}

std::vector<float> softmax_values(const Param & bias) {
    const float max_value = *std::max_element(bias.values.begin(), bias.values.end());