ChazzyG/Retrieval-based-Voice-Conversion-WebUI

1from infer_pack.models_onnx_moess import SynthesizerTrnMs256NSFsidM2import torch3 4person = "Shiroha/shiroha.pth"5exported_path = "model.onnx"6 7 8cpt = torch.load(person, map_location="cpu")9cpt["config"][-3] = cpt["weight"]["emb_g.weight"].shape[0]  # n_spk10print(*cpt["config"])11net_g = SynthesizerTrnMs256NSFsidM(*cpt["config"], is_half=False)12net_g.load_state_dict(cpt["weight"], strict=False)13 14test_phone = torch.rand(1, 200, 256)15test_phone_lengths = torch.tensor([200]).long()16test_pitch = torch.randint(size=(1, 200), low=5, high=255)17test_pitchf = torch.rand(1, 200)18test_ds = torch.LongTensor([0])19test_rnd = torch.rand(1, 192, 200)20input_names = ["phone", "phone_lengths", "pitch", "pitchf", "ds", "rnd"]21output_names = [22    "audio",23]24device = "cpu"25torch.onnx.export(26    net_g,27    (28        test_phone.to(device),29        test_phone_lengths.to(device),30        test_pitch.to(device),31        test_pitchf.to(device),32        test_ds.to(device),33        test_rnd.to(device),34    ),35    exported_path,36    dynamic_axes={37        "phone": [1],38        "pitch": [1],39        "pitchf": [1],40        "rnd": [2],41    },42    do_constant_folding=False,43    opset_version=16,44    verbose=False,45    input_names=input_names,46    output_names=output_names,47)48