Function graph_instance_norm_prelu

src/framework/audio/zipenhancer.cpp:211–250

}

core::TensorValue graph_instance_norm_prelu(
    core::ModuleBuildContext & ctx,
    const core::TensorValue & input,
    const std::unordered_map<std::string, Param> & params,
    const std::string & norm_weight_name,
    const std::string & norm_bias_name,
    const std::string & prelu_weight_name) {
    const auto & norm_weight = require_param(params, norm_weight_name);
    const auto & norm_bias = require_param(params, norm_bias_name);
    const auto & prelu_weight = require_param(params, prelu_weight_name);
    core::validate_shape(norm_weight.tensor, core::TensorShape::from_dims({input.shape.dims[1]}), "ZipEnhancer instance norm weight");
    core::validate_shape(norm_bias.tensor, core::TensorShape::from_dims({input.shape.dims[1]}), "ZipEnhancer instance norm bias");
    core::validate_shape(prelu_weight.tensor, core::TensorShape::from_dims({input.shape.dims[1]}), "ZipEnhancer PReLU weight");

    auto mean_f = modules::ReduceMeanModule({3}).build(ctx, input);
    auto mean = modules::ReduceMeanModule({2}).build(ctx, mean_f);
    auto mean_rep = core::wrap_tensor(ggml_repeat(ctx.ggml, mean.tensor, input.tensor), input.shape, GGML_TYPE_F32);
    auto centered = core::wrap_tensor(ggml_sub(ctx.ggml, input.tensor, mean_rep.tensor), input.shape, GGML_TYPE_F32);
    auto squared = core::wrap_tensor(ggml_sqr(ctx.ggml, centered.tensor), centered.shape, GGML_TYPE_F32);
    auto var_f = modules::ReduceMeanModule({3}).build(ctx, squared);
    auto var = modules::ReduceMeanModule({2}).build(ctx, var_f);
    auto stddev = core::wrap_tensor(
        ggml_sqrt(ctx.ggml, ggml_scale_bias(ctx.ggml, var.tensor, 1.0f, 1.0e-5f)),
        var.shape,
        GGML_TYPE_F32);
    auto stddev_rep = core::wrap_tensor(ggml_repeat(ctx.ggml, stddev.tensor, input.tensor), input.shape, GGML_TYPE_F32);
    auto normed = core::wrap_tensor(ggml_div(ctx.ggml, centered.tensor, stddev_rep.tensor), input.shape, GGML_TYPE_F32);
    auto weight = core::reshape_tensor(ctx, norm_weight.tensor, core::TensorShape::from_dims({1, input.shape.dims[1], 1, 1}));
    auto bias = core::reshape_tensor(ctx, norm_bias.tensor, core::TensorShape::from_dims({1, input.shape.dims[1], 1, 1}));
    auto weight_rep = core::wrap_tensor(ggml_repeat(ctx.ggml, weight.tensor, normed.tensor), normed.shape, GGML_TYPE_F32);
    auto bias_rep = core::wrap_tensor(ggml_repeat(ctx.ggml, bias.tensor, normed.tensor), normed.shape, GGML_TYPE_F32);
    auto affine = core::wrap_tensor(ggml_add(ctx.ggml, ggml_mul(ctx.ggml, normed.tensor, weight_rep.tensor), bias_rep.tensor), normed.shape, GGML_TYPE_F32);

    auto positive = modules::ReluModule().build(ctx, affine);
    auto negative = modules::ReluModule().build(ctx, core::wrap_tensor(ggml_scale(ctx.ggml, affine.tensor, -1.0f), affine.shape, GGML_TYPE_F32));
    auto slope = core::reshape_tensor(ctx, prelu_weight.tensor, core::TensorShape::from_dims({1, input.shape.dims[1], 1, 1}));
    auto slope_rep = core::wrap_tensor(ggml_repeat(ctx.ggml, slope.tensor, affine.tensor), affine.shape, GGML_TYPE_F32);
    auto scaled_negative = core::wrap_tensor(ggml_mul(ctx.ggml, slope_rep.tensor, negative.tensor), affine.shape, GGML_TYPE_F32);
    return core::wrap_tensor(ggml_sub(ctx.ggml, positive.tensor, scaled_negative.tensor), affine.shape, GGML_TYPE_F32);
}

core::TensorValue graph_dense_block(
    core::ModuleBuildContext & ctx,