Convert XTTS-v2 to ONNX

[Dong-mel]

I am trying to convert the xtts-v2 model into onnx format. The specific code is based on the conversion method of vits. However, I encountered the following problems during the conversion. Can anyone give me some suggestions?

Code

import os

import torch
from TTS.tts.models.xtts import Xtts
from TTS.tts.configs.xtts_config import XttsConfig



def export_onnx(
        xtts_model,
        xtts_config,
        output_path: str = "xtts_model.onnx",
        verbose: bool = True
    ):
    """Export XTTS model to ONNX format for inference.

    Args:
        output_path (str): Path to save the exported model.
        verbose (bool): Print verbose information. Defaults to True.
    """

    _forward = xtts_model.forward
    training = xtts_model.training

    xtts_model.eval()

    def onnx_inference(text, language, gpt_cond_latent, speaker_embedding):

        ret_data = xtts_model.inference(text, language, gpt_cond_latent, speaker_embedding)
        
        wav = torch.tensor(ret_data['wav'], dtype=torch.float32, device=xtts_model.device)
        print(wav, wav.shape)
        return wav

    xtts_model.forward = onnx_inference

    text = "sévigné de qui les attraits servent aux grâces de modèle et qui naquîtes toute belle à votre indifférence près pourriez-vous être favorable aux jeux innocents d'une fable et voir sans vous épouvanter" 
    language = "fr"  
    ref_audio_path = "feamle1.wav"

    (gpt_cond_latent, speaker_embedding) = xtts_model.get_conditioning_latents(
        audio_path=ref_audio_path,
        gpt_cond_len=xtts_config.gpt_cond_len,
        gpt_cond_chunk_len=xtts_config.gpt_cond_chunk_len,
        max_ref_length=xtts_config.max_ref_len,
        sound_norm_refs=xtts_config.sound_norm_refs,
    )

    gpt_cond_latent = gpt_cond_latent.to(xtts_model.device)
    speaker_embedding = speaker_embedding.to(xtts_model.device)

    # temperature = xtts_config.temperature
    # length_penalty = xtts_config.length_penalty
    # repetition_penalty = xtts_config.repetition_penalty
    # top_k = xtts_config.top_k
    # top_p = xtts_config.top_p

    dummy_input = (
        text, 
        language, 
        gpt_cond_latent,
        speaker_embedding,
        # temperature, 
        # length_penalty, 
        # repetition_penalty, 
        # top_k, 
        # top_p
    )
    input_names = [
        "text", 
        "language", 
        "gpt_cond_latent",
        "speaker_embedding",
        # "temperature", 
        # "length_penalty", 
        # "repetition_penalty", 
        # "top_k", 
        # "top_p"
    ]

    torch.onnx.export(
        model=xtts_model,
        args=dummy_input,
        opset_version=15,
        f=output_path,
        verbose=verbose,
        input_names=input_names,
        output_names=["wav"],
        dynamic_axes={
            "text": {0: "batch_size"},
            "config": {0: "batch_size"},
            "speaker_wav": {0: "batch_size"},
            "language": {0: "batch_size"},
            "wav": {0: "batch_size", 1: "time"}
        },
    )

    xtts_model.forward = _forward
    if training:
        xtts_model.train()


if __name__ == "__main__":

    model_dir_path = "coqui-ai-TTS/model_finetuning/run/training/xtts-v2-September-29-2024_08+42PM-0a184182"
    yaml_path = os.path.join(model_dir_path, "config.json")
    model_path = os.path.join(model_dir_path, "best_model_reduce.pth")

    omodel = model_path.replace('.pth', '_onnx.pth')

    config = XttsConfig()
    config.load_json(yaml_path)
    xtts = Xtts.init_from_config(config)
    xtts.load_checkpoint(config, checkpoint_path=model_path)

    for param in xtts.parameters():
        param.data = param.data.detach()

    xtts.requires_grad_(False)

    xtts.eval()

    export_onnx(xtts, config, omodel)

Error

/users/anaconda3/envs/env/lib/python3.9/site-packages/torch/onnx/utils.py:2135: UserWarning: Provided key config for dynamic axes is not a valid input/output name
  warnings.warn(
/users/anaconda3/envs/env/lib/python3.9/site-packages/torch/onnx/utils.py:2135: UserWarning: Provided key speaker_wav for dynamic axes is not a valid input/output name
  warnings.warn(
/users/tts/Projects/coqui-ai-TTS/TTS/tts/models/xtts.py:542: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  assert (
/users/anaconda3/envs/env/lib/python3.9/site-packages/transformers/generation/utils.py:1479: TracerWarning: torch.tensor results are registered as constants in the trace. You can safely ignore this warning if you use this function to create tensors out of constant variables that would be the same every time you call this function. In any other case, this might cause the trace to be incorrect.
  return torch.tensor(token, device=device, dtype=torch.long)
/users/anaconda3/envs/env/lib/python3.9/site-packages/transformers/generation/utils.py:1513: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  if eos_token_id is not None and torch.isin(elements=eos_token_id, test_elements=pad_token_id).any():
The attention mask is not set and cannot be inferred from input because pad token is same as eos token.As a consequence, you may observe unexpected behavior. Please pass your input's `attention_mask` to obtain reliable results.
/users/anaconda3/envs/env/lib/python3.9/site-packages/transformers/generation/utils.py:1526: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  if eos_token_id is not None and (torch.is_floating_point(eos_token_id) or (eos_token_id < 0).any()):
/users/anaconda3/envs/env/lib/python3.9/site-packages/transformers/generation/utils.py:1670: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  if (
/users/anaconda3/envs/env/lib/python3.9/site-packages/transformers/generation/utils.py:1255: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  if input_ids_length >= generation_config.max_length:
/users/anaconda3/envs/env/lib/python3.9/site-packages/transformers/generation/utils.py:1272: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  if generation_config.min_length is not None and generation_config.min_length > generation_config.max_length:
/users/tts/Projects/coqui-ai-TTS/TTS/tts/layers/xtts/gpt_inference.py:79: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  if input_ids.shape[1] != 1:
/users/tts/Projects/coqui-ai-TTS/TTS/tts/layers/xtts/gpt_inference.py:83: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  if self.cached_prefix_emb.shape[0] != gen_emb.shape[0]:
/users/anaconda3/envs/env/lib/python3.9/site-packages/transformers/modeling_attn_mask_utils.py:114: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  if (input_shape[-1] > 1 or self.sliding_window is not None) and self.is_causal:
/users/anaconda3/envs/env/lib/python3.9/site-packages/transformers/modeling_attn_mask_utils.py:162: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  if past_key_values_length > 0:
/users/anaconda3/envs/env/lib/python3.9/site-packages/transformers/generation/logits_process.py:524: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  top_k = min(self.top_k, scores.size(-1))  # Safety check
/users/anaconda3/envs/env/lib/python3.9/site-packages/transformers/generation/stopping_criteria.py:77: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  if self.max_position_embeddings is not None and not is_done and cur_len >= self.max_position_embeddings:
/users/anaconda3/envs/env/lib/python3.9/site-packages/transformers/generation/utils.py:2085: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  elif this_peer_finished:
/users/tts/Projects/coqui-ai-TTS/TTS/tts/layers/xtts/gpt.py:39: TracerWarning: torch.tensor results are registered as constants in the trace. You can safely ignore this warning if you use this function to create tensors out of constant variables that would be the same every time you call this function. In any other case, this might cause the trace to be incorrect.
  return self.emb(torch.tensor([ind], device=dev)).unsqueeze(0)
/users/tts/Projects/coqui-ai-TTS/TTS/tts/models/xtts.py:562: TracerWarning: torch.tensor results are registered as constants in the trace. You can safely ignore this warning if you use this function to create tensors out of constant variables that would be the same every time you call this function. In any other case, this might cause the trace to be incorrect.
  expected_output_len = torch.tensor(
/users/tts/Projects/coqui-ai-TTS/TTS/tts/models/xtts.py:566: TracerWarning: torch.tensor results are registered as constants in the trace. You can safely ignore this warning if you use this function to create tensors out of constant variables that would be the same every time you call this function. In any other case, this might cause the trace to be incorrect.
  text_len = torch.tensor([text_tokens.shape[-1]], device=self.device)
/users/tts/Projects/coqui-ai-TTS/TTS/tts/layers/xtts/gpt.py:425: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  if max_mel_len > audio_codes.shape[-1]:
/users/tts/Projects/coqui-ai-TTS/TTS/tts/layers/xtts/gpt.py:429: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  assert (
/users/tts/Projects/coqui-ai-TTS/TTS/tts/layers/xtts/gpt.py:432: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  assert (
/users/tts/Projects/coqui-ai-TTS/TTS/tts/layers/xtts/gpt.py:244: TracerWarning: Using len to get tensor shape might cause the trace to be incorrect. Recommended usage would be tensor.shape[0]. Passing a tensor of different shape might lead to errors or silently give incorrect results.
  for b in range(len(code_lengths)):
/users/tts/Projects/coqui-ai-TTS/TTS/tts/layers/xtts/gpt.py:246: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  if actual_end < mel_input_tokens.shape[-1]:
/users/anaconda3/envs/env/lib/python3.9/site-packages/transformers/models/gpt2/modeling_gpt2.py:648: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  is_causal = True if attention_mask is None and q_len > 1 and not is_cross_attention else False
/users/tts/Projects/coqui-ai-TTS/TTS/tts/models/xtts.py:586: TracerWarning: Converting a tensor to a NumPy array might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  "wav": torch.cat(wavs, dim=0).numpy(),
/users/tts/Projects/coqui-ai-TTS/TTS/tts/models/xtts.py:587: TracerWarning: Converting a tensor to a NumPy array might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  "gpt_latents": torch.cat(gpt_latents_list, dim=1).numpy(),
/users/tts/code_Dong/TTS/xtts_convert_onnx_try.py:35: TracerWarning: torch.tensor results are registered as constants in the trace. You can safely ignore this warning if you use this function to create tensors out of constant variables that would be the same every time you call this function. In any other case, this might cause the trace to be incorrect.
  wav = torch.tensor(ret_data['wav'], dtype=torch.float32, device=xtts_model.device)
tensor([1.5133e-04, 8.5510e-05, 4.9809e-05,  ..., 4.1050e-05, 8.2305e-05,
        1.1775e-04]) torch.Size([356608])
Traceback (most recent call last):
  File "/users/tts/code_Dong/TTS/xtts_convert_onnx_try.py", line 130, in <module>
    export_onnx(xtts, config, omodel)
  File "/users/tts/code_Dong/TTS/xtts_convert_onnx_try.py", line 87, in export_onnx
    torch.onnx.export(
  File "/users/anaconda3/envs/env/lib/python3.9/site-packages/torch/onnx/utils.py", line 551, in export
    _export(
  File "/users/anaconda3/envs/env/lib/python3.9/site-packages/torch/onnx/utils.py", line 1648, in _export
    graph, params_dict, torch_out = _model_to_graph(
  File "/users/anaconda3/envs/env/lib/python3.9/site-packages/torch/onnx/utils.py", line 1170, in _model_to_graph
    graph, params, torch_out, module = _create_jit_graph(model, args)
  File "/users/anaconda3/envs/env/lib/python3.9/site-packages/torch/onnx/utils.py", line 1046, in _create_jit_graph
    graph, torch_out = _trace_and_get_graph_from_model(model, args)
  File "/users/anaconda3/envs/env/lib/python3.9/site-packages/torch/onnx/utils.py", line 950, in _trace_and_get_graph_from_model
    trace_graph, torch_out, inputs_states = torch.jit._get_trace_graph(
  File "/users/anaconda3/envs/env/lib/python3.9/site-packages/torch/jit/_trace.py", line 1497, in _get_trace_graph
    outs = ONNXTracedModule(
  File "/users/anaconda3/envs/env/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1553, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
  File "/users/anaconda3/envs/env/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1562, in _call_impl
    return forward_call(*args, **kwargs)
  File "/users/anaconda3/envs/env/lib/python3.9/site-packages/torch/jit/_trace.py", line 141, in forward
    graph, out = torch._C._create_graph_by_tracing(
RuntimeError: 0 INTERNAL ASSERT FAILED at "../torch/csrc/jit/ir/alias_analysis.cpp":615, please report a bug to PyTorch. We don't have an op for aten::full but it isn't a special case.  Argument types: int[], bool, int, NoneType, Device, bool,

Candidates:
        aten::full.names(int[] size, Scalar fill_value, *, str[]? names, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
        aten::full(SymInt[] size, Scalar fill_value, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
        aten::full.names_out(int[] size, Scalar fill_value, *, str[]? names, Tensor(a!) out) -> Tensor(a!)
        aten::full.out(SymInt[] size, Scalar fill_value, *, Tensor(a!) out) -> Tensor(a!)

[VivekGangurde14]

Did you find any solution as i m trying the same?

[Dong-mel]

not yet

[RoundTwoTwoTwo]

Hey, do you find any solution now? I meet the same question and have totally no idea about it