vall-e/vall_e.cpp/include/decoder.h

114 lines
2.9 KiB
C++

#pragma once
#include <vector>
#include "ggml.h"
#include "ggml-alloc.h"
#include "ggml-backend.h"
#include "lstm.h"
#include "utils.h"
struct encodec_decoder_block {
// upsampling layers
struct ggml_tensor *us_conv_w;
struct ggml_tensor *us_conv_b;
// conv1
struct ggml_tensor *conv_1_w;
struct ggml_tensor *conv_1_b;
// conv2
struct ggml_tensor *conv_2_w;
struct ggml_tensor *conv_2_b;
// shortcut
struct ggml_tensor *conv_sc_w;
struct ggml_tensor *conv_sc_b;
};
struct encodec_decoder {
struct ggml_tensor *init_conv_w;
struct ggml_tensor *init_conv_b;
encodec_lstm lstm;
struct ggml_tensor *final_conv_w;
struct ggml_tensor *final_conv_b;
std::vector<encodec_decoder_block> blocks;
};
struct ggml_tensor *encodec_forward_decoder(
const struct encodec_decoder *decoder, struct ggml_context *ctx0,
struct ggml_tensor *quantized_out, const int *ratios, const int kernel_size, const int res_kernel_size,
const int stride) {
if (!quantized_out) {
fprintf(stderr, "%s: null input tensor\n", __func__);
return NULL;
}
struct ggml_tensor *inpL = strided_conv_1d(
ctx0, quantized_out, decoder->init_conv_w, decoder->init_conv_b, stride);
// lstm
{
struct ggml_tensor *cur = inpL;
const encodec_lstm lstm = decoder->lstm;
// first lstm layer
char l0_prefix[7] = "dec_l0";
struct ggml_tensor *hs1 = forward_pass_lstm_unilayer(
ctx0, cur, lstm.l0_ih_w, lstm.l0_hh_w, lstm.l0_ih_b, lstm.l0_hh_b, l0_prefix);
// second lstm layer
char l1_prefix[7] = "dec_l1";
struct ggml_tensor *out = forward_pass_lstm_unilayer(
ctx0, hs1, lstm.l1_ih_w, lstm.l1_hh_w, lstm.l1_ih_b, lstm.l1_hh_b, l1_prefix);
inpL = ggml_add(ctx0, inpL, out);
}
for (int layer_ix = 0; layer_ix < 4; layer_ix++) {
encodec_decoder_block block = decoder->blocks[layer_ix];
// upsampling layers
inpL = ggml_elu(ctx0, inpL);
inpL = strided_conv_transpose_1d(
ctx0, inpL, block.us_conv_w, block.us_conv_b, ratios[layer_ix]);
struct ggml_tensor *current = inpL;
// shortcut
struct ggml_tensor *shortcut = strided_conv_1d(
ctx0, inpL, block.conv_sc_w, block.conv_sc_b, stride);
// conv1
current = ggml_elu(ctx0, current);
current = strided_conv_1d(
ctx0, current, block.conv_1_w, block.conv_1_b, stride);
// conv2
current = ggml_elu(ctx0, current);
current = strided_conv_1d(
ctx0, current, block.conv_2_w, block.conv_2_b, stride);
// residual connection
inpL = ggml_add(ctx0, current, shortcut);
}
// final conv
inpL = ggml_elu(ctx0, inpL);
struct ggml_tensor *decoded_inp = strided_conv_1d(
ctx0, inpL, decoder->final_conv_w, decoder->final_conv_b, stride);
return decoded_inp;
}