Method build

src/framework/modules/conformer_modules.cpp:121–178

}

ConvSubsamplingOutputs ConvSubsamplingModule::build(
    core::ModuleBuildContext & ctx,
    const core::TensorValue & input,
    const core::TensorValue & lengths,
    const ConvSubsamplingWeights & weights) const {
    core::validate_rank_between(input, 3, 3, "input");
    core::validate_last_dim(input, config_.input_features, "input");

    const int64_t batch = input.shape.dims[0];
    const int64_t frames = input.shape.dims[1];
    const int64_t total_padding = config_.kernel_size - 1;
    const int64_t feat_after = calc_subsampled_dim(config_.input_features, total_padding, config_.kernel_size, 2, config_.stride);
    const int64_t frames_after = calc_subsampled_dim(frames, total_padding, config_.kernel_size, 2, config_.stride);
    auto current_lengths = lengths;

    auto x4 = core::reshape_tensor(ctx, input, core::TensorShape::from_dims({batch, 1, frames, config_.input_features}));
    x4 = apply_lengths_time_mask_4d(ctx, x4, current_lengths);
    x4 = Conv2dModule({
        1,
        config_.conv_channels,
        config_.kernel_size,
        config_.kernel_size,
        config_.stride,
        config_.stride,
        static_cast<int>(total_padding / 2),
        static_cast<int>(total_padding / 2),
        1,
        1,
        config_.use_bias,
    }).build(ctx, x4, weights.conv0);
    x4 = ReluModule().build(ctx, x4);
    current_lengths = calc_subsampled_lengths(ctx, current_lengths, total_padding, config_.kernel_size, 1, config_.stride);
    x4 = apply_lengths_time_mask_4d(ctx, x4, current_lengths);
    x4 = Conv2dModule({
        config_.conv_channels,
        config_.conv_channels,
        config_.kernel_size,
        config_.kernel_size,
        config_.stride,
        config_.stride,
        static_cast<int>(total_padding / 2),
        static_cast<int>(total_padding / 2),
        1,
        1,
        config_.use_bias,
    }).build(ctx, x4, weights.conv1);
    x4 = ReluModule().build(ctx, x4);
    current_lengths = calc_subsampled_lengths(ctx, current_lengths, total_padding, config_.kernel_size, 1, config_.stride);
    x4 = apply_lengths_time_mask_4d(ctx, x4, current_lengths);
    x4 = tensor_layout::swap_channel_time_axes_4d(ctx, x4);
    x4 = core::wrap_tensor(ggml_cont(ctx.ggml, x4.tensor), x4.shape, x4.type);
    auto flat = core::reshape_tensor(ctx, x4, core::TensorShape::from_dims({batch, frames_after, config_.conv_channels * feat_after}));
    auto output = LinearModule({config_.conv_channels * feat_after, config_.output_features, config_.use_bias}).build(ctx, flat, weights.linear);
    return {
        output,
        current_lengths,
    };
}