main/library/algorithm/onnx_export.py

import os
import io
import sys
import onnx
import json
import torch
import onnxsim
import warnings

sys.path.append(os.getcwd())

from main.library.algorithm.synthesizers import SynthesizerONNX

warnings.filterwarnings("ignore")

def onnx_exporter(input_path, output_path, is_half=False, device="cpu"):
    cpt = (torch.load(input_path, map_location="cpu", weights_only=True) if os.path.isfile(input_path) else None)
    cpt["config"][-3] = cpt["weight"]["emb_g.weight"].shape[0]

    model_name, model_author, epochs, steps, version, f0, model_hash, vocoder, creation_date, energy_use = cpt.get("model_name", None), cpt.get("author", None), cpt.get("epoch", None), cpt.get("step", None), cpt.get("version", "v1"), cpt.get("f0", 1), cpt.get("model_hash", None), cpt.get("vocoder", "Default"), cpt.get("creation_date", None), cpt.get("energy", False)
    text_enc_hidden_dim = 768 if version == "v2" else 256
    tgt_sr = cpt["config"][-1]

    net_g = SynthesizerONNX(*cpt["config"], use_f0=f0, text_enc_hidden_dim=text_enc_hidden_dim, vocoder=vocoder, checkpointing=False, energy=energy_use)
    net_g.load_state_dict(cpt["weight"], strict=False)
    net_g.eval().to(device)
    net_g = (net_g.half() if is_half else net_g.float())

    phone = torch.rand(1, 200, text_enc_hidden_dim).to(device)
    phone_length = torch.tensor([200]).long().to(device)
    ds = torch.LongTensor([0]).to(device)
    rnd = torch.rand(1, 192, 200).to(device)

    if f0:
        pitch = torch.randint(size=(1, 200), low=5, high=255).to(device)
        pitchf = torch.rand(1, 200).to(device)

    if energy_use: 
        energy = torch.rand(1, 200).to(device)

    args = [phone, phone_length, ds, rnd]
    input_names = ["phone", "phone_lengths", "ds", "rnd"]
    dynamic_axes = {"phone": [1], "rnd": [2]}

    if f0:
        args += [pitch, pitchf]
        input_names += ["pitch", "pitchf"]
        dynamic_axes.update({"pitch": [1], "pitchf": [1]})

    if energy_use:
        args.append(energy)
        input_names.append("energy")
        dynamic_axes.update({"energy": [1]})

    try:
        with io.BytesIO() as model:
            torch.onnx.export(
                net_g, 
                tuple(args), 
                model, 
                do_constant_folding=True, 
                opset_version=17, 
                verbose=False, 
                input_names=input_names, 
                output_names=["audio"], 
                dynamic_axes=dynamic_axes
            )

            model, _ = onnxsim.simplify(onnx.load_model_from_string(model.getvalue()))
            model.metadata_props.append(
                onnx.StringStringEntryProto(
                    key="model_info", 
                    value=json.dumps(
                        {
                            "model_name": model_name, 
                            "author": model_author, 
                            "epoch": epochs, 
                            "step": steps, 
                            "version": version, 
                            "sr": tgt_sr, 
                            "f0": f0, 
                            "model_hash": model_hash, 
                            "creation_date": creation_date, 
                            "vocoder": vocoder, 
                            "text_enc_hidden_dim": text_enc_hidden_dim,
                            "energy": energy_use
                        }
                    )
                )
            )

        if is_half:
            try:
                import onnxconverter_common
            except:
                os.system(f"{sys.executable} -m pip install onnxconverter_common")
                import onnxconverter_common

            model = onnxconverter_common.convert_float_to_float16(model, keep_io_types=True)

        onnx.save(model, output_path)
        return output_path
    except:
        import traceback
        print(traceback.print_exc())
        return None