onnx export tool is missing from this fork.

[SkybuckFlying]

G:\Tools\Ultimate-RVC\infer\modules\onnx\export.py

I recommend adding it so the obs-rvc plugin can be used.

This script is needed to convert rvc models/pth to onnx.

[SkybuckFlying]

Unfortunately the onnx model files produced crash the obs-rvc plugin.

I made some youtube videos, documenting how I build this plugin and also trying to convert the model.

At the end of this script was added to actually run it when it runs alone.
    
if __name__ == "__main__":
    export_onnx(
        r"G:\Tools\Ultimate-RVC (JackismyShephard)\models\rvc\voice_models\Transformers - Optimus Prime\OptimusPrime_550e_15950s.pth",
        r"G:\Tools\Ultimate-RVC (JackismyShephard)\models\rvc\voice_models\Transformers - Optimus Prime\OptimusPrime.onnx"
    )
    print("✅ Export complete.")    


To make it run alone a batch file was used:

@echo on
set PYTHONPATH="G:\Tools\Ultimate-RVC (Chinese)"
python -m infer.modules.onnx.export --model "G:\Tools\Ultimate-RVC (JackismyShephard)\models\rvc\voice_models\Transformers - Optimus Prime\OptimusPrime_550e_15950s.pth" --output "G:\Tools\Ultimate-RVC (JackismyShephard)\models\rvc\voice_models\Transformers - Optimus Prime\OptimusPrime.onnx" > output.log 2>&1
echo Exit code: %ERRORLEVEL%
pause

The paths are not yet retrieved by this scripts but shows how to do it, this batch file must be run from the root/base folder of the chinese version of this ultimate rvc...

Anyway this is as far as I got, would be cool if Jack could take a look at it and see what might be wrong and how to fix it...

The AI tried to fix it, some versions, but for now I give up, the exporter does seem to produce a valid onnx file, though not entirely sure, how to disable cmake from perl strawberry to use visual studio cmaka to install onnxsim properly and such to make this exporter run properly.

Some versions of this export by copilot:

import sys
sys.path.insert(0, r"G:\Tools\Ultimate-RVC (Chinese)")

import torch
import onnx
from infer.lib.infer_pack.models_onnx import SynthesizerTrnMsNSFsidM

def export_onnx(ModelPath, ExportedPath):
    cpt = torch.load(ModelPath, map_location="cpu")
    cpt["config"][-3] = cpt["weight"]["emb_g.weight"].shape[0]
    vec_channels = 256 if cpt.get("version", "v1") == "v1" else 768

    test_phone = torch.rand(1, 200, vec_channels)
    test_phone_lengths = torch.tensor([200]).long()
    test_pitch = torch.randint(size=(1, 200), low=5, high=255)
    test_pitchf = torch.rand(1, 200)
    test_ds = torch.LongTensor([0])
    test_rnd = torch.rand(1, 192, 200)

    device = "cpu"

    net_g = SynthesizerTrnMsNSFsidM(
        *cpt["config"], is_half=False, version=cpt.get("version", "v1")
    )
    net_g.load_state_dict(cpt["weight"], strict=False)

    input_names = ["phone", "phone_lengths", "pitch", "pitchf", "ds", "rnd"]
    output_names = ["audio"]

    torch.onnx.export(
        net_g,
        (
            test_phone.to(device),
            test_phone_lengths.to(device),
            test_pitch.to(device),
            test_pitchf.to(device),
            test_ds.to(device),
            test_rnd.to(device),
        ),
        ExportedPath,
        do_constant_folding=False,
        opset_version=13,
        verbose=False,
        input_names=input_names,
        output_names=output_names
    )

    try:
        import onnxsim
        model, _ = onnxsim.simplify(ExportedPath)
        onnx.save(model, ExportedPath)
        print("v ONNX model simplified.")
    except ImportError:
        print("x onnxsim not installed. Exported raw ONNX model.")

    print("v Export complete:", ExportedPath)

if __name__ == "__main__":
    import argparse
    parser = argparse.ArgumentParser()
    parser.add_argument("--model", required=True)
    parser.add_argument("--output", required=True)
    args = parser.parse_args()
    export_onnx(args.model, args.output)

import sys
sys.path.insert(0, r"G:\Tools\Ultimate-RVC (Chinese)")

import torch
import onnx
from infer.lib.infer_pack.models_onnx import SynthesizerTrnMsNSFsidM

def export_onnx(ModelPath, ExportedPath):
    cpt = torch.load(ModelPath, map_location="cpu")
    cpt["config"][-3] = cpt["weight"]["emb_g.weight"].shape[0]
    vec_channels = 256 if cpt.get("version", "v1") == "v1" else 768

    test_phone = torch.rand(1, 200, vec_channels)
    test_phone_lengths = torch.tensor([200]).long()
    test_pitch = torch.randint(size=(1, 200), low=5, high=255)
    test_pitchf = torch.rand(1, 200)
    test_ds = torch.LongTensor([0])
    test_rnd = torch.rand(1, 192, 200)

    device = "cpu"

    net_g = SynthesizerTrnMsNSFsidM(
        *cpt["config"], is_half=False, version=cpt.get("version", "v1")
    )
    net_g.load_state_dict(cpt["weight"], strict=False)

    input_names = ["phone", "phone_lengths", "pitch", "pitchf", "ds", "rnd"]
    output_names = ["audio"]

    torch.onnx.export(
        net_g,
        (
            test_phone.to(device),
            test_phone_lengths.to(device),
            test_pitch.to(device),
            test_pitchf.to(device),
            test_ds.to(device),
            test_rnd.to(device),
        ),
        ExportedPath,
        do_constant_folding=False,
        opset_version=13,
        verbose=False,
        input_names=input_names,
        output_names=output_names
    )

    try:
        import onnxsim
        model, _ = onnxsim.simplify(ExportedPath)
        onnx.save(model, ExportedPath)
        print("v ONNX model simplified.")
    except ImportError:
        print("x onnxsim not installed. Exported raw ONNX model.")

    print("v Export complete:", ExportedPath)

if __name__ == "__main__":
    import argparse
    parser = argparse.ArgumentParser()
    parser.add_argument("--model", required=True)
    parser.add_argument("--output", required=True)
    args = parser.parse_args()
    export_onnx(args.model, args.output)

import sys
sys.path.insert(0, r"G:\Tools\Ultimate-RVC (Chinese)")

import torch
import onnx
from infer.lib.infer_pack.models_onnx import SynthesizerTrnMsNSFsidM

def export_onnx(ModelPath, ExportedPath):
    cpt = torch.load(ModelPath, map_location="cpu")
    cpt["config"][-3] = cpt["weight"]["emb_g.weight"].shape[0]
    vec_channels = 256 if cpt.get("version", "v1") == "v1" else 768

    test_phone = torch.rand(1, 200, vec_channels)
    test_phone_lengths = torch.tensor([200]).long()
    test_pitch = torch.randint(size=(1, 200), low=5, high=255)
    test_pitchf = torch.rand(1, 200)
    test_ds = torch.LongTensor([0])
    test_rnd = torch.rand(1, 192, 200)

    device = "cpu"

    net_g = SynthesizerTrnMsNSFsidM(
        *cpt["config"], is_half=False, version=cpt.get("version", "v1")
    )
    net_g.load_state_dict(cpt["weight"], strict=False)

    input_names = ["phone", "phone_lengths", "pitch", "pitchf", "ds", "rnd"]
    output_names = ["audio"]

    torch.onnx.export(
        net_g,
        (
            test_phone.to(device),
            test_phone_lengths.to(device),
            test_pitch.to(device),
            test_pitchf.to(device),
            test_ds.to(device),
            test_rnd.to(device),
        ),
        ExportedPath,
        dynamic_axes={
            "phone": [1],
            "pitch": [1],
            "pitchf": [1],
            "rnd": [2],
        },
        do_constant_folding=False,
        opset_version=18,
        verbose=False,
        input_names=input_names,
        output_names=output_names,
    )

    try:
        import onnxsim
        model, _ = onnxsim.simplify(ExportedPath)
        onnx.save(model, ExportedPath)
        print("? ONNX model simplified.")
    except ImportError:
        print("?? onnxsim not installed. Exported raw ONNX model.")

    print("? Export complete:", ExportedPath)

# Run directly
if __name__ == "__main__":
    import argparse
    parser = argparse.ArgumentParser()
    parser.add_argument("--model", required=True)
    parser.add_argument("--output", required=True)
    args = parser.parse_args()
    export_onnx(args.model, args.output)


one of the versions changed opset... anyway these export routines create either crashing or hanging models when loaded into the obs-rvc plugin.

Maybe the obs-rvc plugin itself needs an update...

Some test code was written to test onnx models:

import onnxruntime as ort
session = ort.InferenceSession("OptimusPrime.onnx")
print("v Model loaded successfully.")

[SkybuckFlying]

Grok feedback to try:

The issue of "nothing being produced" when running your batch file to export an ONNX model could stem from several potential causes. Below, I’ll analyze the provided Python script and batch file, identify possible reasons for the failure, and suggest steps to diagnose and resolve the issue.

---

Key Components of the Problem

1. Python Script: The script (export_onnx) is designed to load a PyTorch model (.pth file), convert it to ONNX format, and save the resulting .onnx file using torch.onnx.export. It also applies ONNX simplification using onnxsim.
2. Batch File: The batch file sets the PYTHONPATH and runs the Python script with specific arguments for the model path and output path.
3. Symptom: The script runs without producing the expected .onnx file, and no clear error messages are provided (or they are not being captured).

---

Potential Reasons for "Nothing Being Produced"

Here are the likely reasons why the script is not producing the expected .onnx file:

1. Error in Python Script Execution

• Issue: The Python script might be failing silently due to an exception (e.g., file not found, module import error, or invalid model configuration).
• Why: The batch file does not redirect Python errors to a log file, and the echo Exit code: %ERRORLEVEL% might not provide enough context if the script crashes silently.
• Diagnosis:
  • Check if the Python script is running at all. Add print statements or logging to the export_onnx function to confirm it starts and progresses.
  • Modify the batch file to capture Python output and errors. For example:

    @echo on
    set PYTHONPATH=G:\Tools\Ultimate-RVC (Chinese)
    python -m infer.modules.onnx.export --model "G:\Tools\Ultimate-RVC (JackismyShephard)\models\rvc\voice_models\Transformers - Optimus Prime\OptimusPrime_550e_15950s.pth" --output "G:\Tools\Ultimate-RVC (JackismyShephard)\models\rvc\voice_models\Transformers - Optimus Prime\OptimusPrime.onnx" > output.log 2>&1
    echo Exit code: %ERRORLEVEL%
    pause

    This redirects both standard output and errors to output.log. Check this file for error messages.
• Common Errors:
  • FileNotFoundError: Ensure the model file (OptimusPrime_550e_15950s.pth) exists at the specified path.
  • ModuleNotFoundError: Verify that the infer.lib.infer_pack.models_onnx module and its dependencies (torch, onnx, onnxsim) are installed in the Python environment.
  • Invalid Model: The .pth file might be corrupted or incompatible with the SynthesizerTrnMsNSFsidM model architecture.

2. Incorrect PYTHONPATH or Environment Issues

• Issue: The PYTHONPATH is set to G:\Tools\Ultimate-RVC (Chinese), which may not include the necessary modules or may point to an incorrect directory.
• Why: If the infer module or its dependencies are not in the PYTHONPATH, Python will fail to import infer.lib.infer_pack.models_onnx.
• Diagnosis:
  • Verify that the directory G:\Tools\Ultimate-RVC (Chinese) contains the infer module or is the correct root for your project.
  • Run the Python command manually in a command prompt to see if the module imports correctly:

    set PYTHONPATH=G:\Tools\Ultimate-RVC (Chinese)
    python -c "from infer.lib.infer_pack.models_onnx import SynthesizerTrnMsNSFsidM"

    If this fails, you’ll see a ModuleNotFoundError.
• Fix:
  • Ensure all required packages (torch, onnx, onnxsim) are installed. Run:

    pip install torch onnx onnxsim

  • If the infer module is part of a custom project, ensure the PYTHONPATH points to the correct directory containing the infer module.
  • Consider using a virtual environment to isolate dependencies and avoid conflicts.

3. Command-Line Argument Parsing Issue

• Issue: The script is invoked with python -m infer.modules.onnx.export --model ... --output ..., but the script itself doesn’t show how it handles these arguments.
• Why: If the export_onnx function is not called correctly or if the command-line arguments are not parsed, the function might not execute.
• Diagnosis:
  • Check if the infer.modules.onnx.export module has a main function or argument parser (e.g., using argparse) that calls export_onnx(ModelPath, ExportedPath).
  • Add a minimal test script to verify argument passing:

    import sys
    print("Arguments:", sys.argv)

    Run it with:

    python -m infer.modules.onnx.export --model test.pth --output test.onnx

    Check if the arguments are received correctly.
• Fix:
  • Ensure the infer.modules.onnx.export module includes code to parse --model and --output arguments and pass them to export_onnx. For example:

    import argparse
    
    def main():
        parser = argparse.ArgumentParser()
        parser.add_argument("--model", required=True)
        parser.add_argument("--output", required=True)
        args = parser.parse_args()
        export_onnx(args.model, args.output)
    
    if __name__ == "__main__":
        main()

4. ONNX Export Failure

• Issue: The torch.onnx.export call might fail due to incompatible model operations, incorrect input shapes, or issues with the model architecture.
• Why: Not all PyTorch operations are supported by ONNX, especially for complex models like SynthesizerTrnMsNSFsidM. Additionally, the input tensors (test_phone, test_phone_lengths, etc.) might not match the model’s expectations.
• Diagnosis:
  • Add error handling around the torch.onnx.export call:

    try:
        torch.onnx.export(
            net_g,
            (
                test_phone.to(device),
                test_phone_lengths.to(device),
                test_pitch.to(device),
                test_pitchf.to(device),
                test_ds.to(device),
                test_rnd.to(device),
            ),
            ExportedPath,
            dynamic_axes={
                "phone": [1],
                "pitch": [1],
                "pitchf": [1],
                "rnd": [2],
            },
            do_constant_folding=False,
            opset_version=18,
            verbose=True,  # Enable verbose to debug
            input_names=input_names,
            output_names=output_names,
        )
        print("ONNX export successful")
    except Exception as e:
        print(f"ONNX export failed: {e}")

  • Set verbose=True in torch.onnx.export to get detailed logs about the export process.
  • Verify that the input tensor shapes (test_phone, test_pitch, etc.) match the model’s expected input shapes. For example, ensure vec_channels (256 or 768) aligns with the model’s configuration.
• Fix:
  • If specific operations are unsupported, simplify the model or use a different opset_version (e.g., try opset_version=16 or 17).
  • Ensure the model is in evaluation mode before export:

    net_g.eval()

  • Check the model’s forward pass with the test inputs to ensure it runs without errors:

    with torch.no_grad():
        output = net_g(
            test_phone.to(device),
            test_phone_lengths.to(device),
            test_pitch.to(device),
            test_pitchf.to(device),
            test_ds.to(device),
            test_rnd.to(device),
        )
        print("Model output shape:", output.shape)

5. ONNX Simplification Failure

• Issue: The onnxsim.simplify step might fail, causing the script to exit without saving the final .onnx file.
• Why: The ONNX model might be too complex, or onnxsim might encounter unsupported operations.
• Diagnosis:
  • Add error handling around the simplification step:

    try:
        model, _ = onnxsim.simplify(ExportedPath)
        onnx.save(model, ExportedPath)
        print("ONNX simplification successful")
    except Exception as e:
        print(f"ONNX simplification failed: {e}")

  • Try running the script without simplification to isolate the issue:

    # Comment out simplification
    # model, _ = onnxsim.simplify(ExportedPath)
    # onnx.save(model, ExportedPath)
    onnx.save(onnx.load(ExportedPath), ExportedPath)  # Save without simplification

• Fix:
  • If simplification fails, skip it and use the unsimplified .onnx file for testing.
  • Update onnxsim to the latest version:

    pip install --upgrade onnxsim

6. File Path Issues

• Issue: The input model path or output path might be incorrect or inaccessible.
• Why: Windows paths with spaces (e.g., Transformers - Optimus Prime) can cause issues if not handled properly, or the script might lack write permissions for the output directory.
• Diagnosis:
  • Verify that the input file exists:

    dir "G:\Tools\Ultimate-RVC (JackismyShephard)\models\rvc\voice_models\Transformers - Optimus Prime\OptimusPrime_550e_15950s.pth"

  • Verify that the output directory is writable by creating a test file:

    echo test > "G:\Tools\Ultimate-RVC (JackismyShephard)\models\rvc\voice_models\Transformers - Optimus Prime\test.txt"

  • Check if the paths are correctly passed to the script by printing them:

    print(f"ModelPath: {ModelPath}")
    print(f"ExportedPath: {ExportedPath}")

• Fix:
  • Use raw strings or double backslashes for Windows paths in the batch file:

    python -m infer.modules.onnx.export --model "G:\Tools\Ultimate-RVC (JackismyShephard)\models\rvc\voice_models\Transformers - Optimus Prime\OptimusPrime_550e_15950s.pth" --output "G:\Tools\Ultimate-RVC (JackismyShephard)\models\rvc\voice_models\Transformers - Optimus Prime\OptimusPrime.onnx"

  • Ensure the script has write permissions for the output directory.

7. Version or Dependency Mismatch

• Issue: The versions of torch, onnx, or onnxsim might be incompatible with the model or the ONNX export process.
• Why: The script uses opset_version=18, which requires a recent version of PyTorch and ONNX. Older versions might not support this opset or the model’s operations.
• Diagnosis:
  • Check the installed versions:

    pip show torch onnx onnxsim

  • Ensure PyTorch supports ONNX opset 18 (requires PyTorch 1.11 or later).
• Fix:
  • Update dependencies:

    pip install --upgrade torch onnx onnxsim

  • If issues persist, try a lower opset version (e.g., opset_version=16).

---

Debugging Steps

To systematically diagnose the issue, follow these steps:

1. Capture Output:

   • Modify the batch file to log output and errors (see modified batch file above).
   • Check output.log for any error messages.

2. Test Python Environment:

   • Verify that all required modules are installed and accessible:

     python -c "import torch, onnx, onnxsim, infer.lib.infer_pack.models_onnx"

   • If this fails, fix the PYTHONPATH or install missing dependencies.

3. Run Script Interactively:

   • Open a Python interpreter and run the export_onnx function manually:

     from infer.modules.onnx.export import export_onnx
     export_onnx(
         "G:\Tools\Ultimate-RVC (JackismyShephard)\models\rvc\voice_models\Transformers - Optimus Prime\OptimusPrime_550e_15950s.pth",
         "G:\Tools\Ultimate-RVC (JackismyShephard)\models\rvc\voice_models\Transformers - Optimus Prime\OptimusPrime.onnx"
     )

   • Add print statements to track progress.

4. Isolate ONNX Export:

   • Comment out the onnxsim.simplify step and check if the .onnx file is created before simplification.
   • If the file is created, the issue is with simplification.

5. Check Model and Inputs:

   • Verify that the .pth file loads correctly:

     import torch
     cpt = torch.load("G:\Tools\Ultimate-RVC (JackismyShephard)\models\rvc\voice_models\Transformers - Optimus Prime\OptimusPrime_550e_15950s.pth", map_location="cpu")
     print(cpt.keys())

   • Ensure the input tensors match the model’s expectations.

---

Suggested Fixes

Based on the likely issues, here are immediate fixes to try:

1. Add Error Handling to the Script:

   import torch
   import onnxsim
   import onnx
   from infer.lib.infer_pack.models_onnx import SynthesizerTrnMsNSFsidM
   
   def export_onnx(ModelPath, ExportedPath):
       try:
           print(f"Loading model from {ModelPath}")
           cpt = torch.load(ModelPath, map_location="cpu")
           print("Model loaded successfully")
           cpt["config"][-3] = cpt["weight"]["emb_g.weight"].shape[0]
           vec_channels = 256 if cpt.get("version", "v1") == "v1" else 768
   
           test_phone = torch.rand(1, 200, vec_channels)
           test_phone_lengths = torch.tensor([200]).long()
           test_pitch = torch.randint(size=(1, 200), low=5, high=255)
           test_pitchf = torch.rand(1, 200)
           test_ds = torch.LongTensor([0])
           test_rnd = torch.rand(1, 192, 200)
   
           device = "cpu"
           net_g = SynthesizerTrnMsNSFsidM(
               *cpt["config"], is_half=False, version=cpt.get("version", "v1")
           )
           net_g.load_state_dict(cpt["weight"], strict=False)
           net_g.eval()
           print("Model initialized successfully")
   
           input_names = ["phone", "phone_lengths", "pitch", "pitchf", "ds", "rnd"]
           output_names = ["audio"]
   
           print(f"Exporting to {ExportedPath}")
           torch.onnx.export(
               net_g,
               (
                   test_phone.to(device),
                   test_phone_lengths.to(device),
                   test_pitch.to(device),
                   test_pitchf.to(device),
                   test_ds.to(device),
                   test_rnd.to(device),
               ),
               ExportedPath,
               dynamic_axes={
                   "phone": [1],
                   "pitch": [1],
                   "pitchf": [1],
                   "rnd": [2],
               },
               do_constant_folding=False,
               opset_version=18,
               verbose=True,
               input_names=input_names,
               output_names=output_names,
           )
           print("ONNX export completed")
   
           print("Simplifying ONNX model")
           model, _ = onnxsim.simplify(ExportedPath)
           onnx.save(model, ExportedPath)
           print("ONNX simplification completed")
           return "Finished"
       except Exception as e:
           print(f"Error: {e}")
           raise

2. Update the Batch File:

   @echo on
   set PYTHONPATH=G:\Tools\Ultimate-RVC (Chinese)
   python -m infer.modules.onnx.export --model "G:\Tools\Ultimate-RVC (JackismyShephard)\models\rvc\voice_models\Transformers - Optimus Prime\OptimusPrime_550e_15950s.pth" --output "G:\Tools\Ultimate-RVC (JackismyShephard)\models\rvc\voice_models\Transformers - Optimus Prime\OptimusPrime.onnx" > output.log 2>&1
   echo Exit code: %ERRORLEVEL%
   type output.log
   pause

3. Verify Dependencies:

   pip install --upgrade torch onnx onnxsim

4. Test Without Simplification:

   • Temporarily comment out the onnxsim.simplify and onnx.save lines to see if the raw ONNX file is created.

---

Expected Outcome

After applying these fixes, you should see:

• The output.log file containing detailed logs or error messages.
• If successful, the OptimusPrime.onnx file in the specified output directory.
• If an error occurs, the exact exception message will help pinpoint the issue (e.g., file not found, unsupported operation, or import error).

If the issue persists after trying these steps, please share the contents of output.log or any error messages, and I can provide further assistance.

[SkybuckFlying]

Last attempt from me and grok for now, fixed script:

import torch
import onnxsim
import onnx
from infer.lib.infer_pack.models_onnx import SynthesizerTrnMsNSFsidM

def export_onnx(ModelPath, ExportedPath):
    print("grok")  # Ensure this is indented with 4 spaces if inside the function
    cpt = torch.load(ModelPath, map_location="cpu")
    cpt["config"][-3] = cpt["weight"]["emb_g.weight"].shape[0]
    vec_channels = 256 if cpt.get("version", "v1") == "v1" else 768

    test_phone = torch.rand(1, 200, vec_channels)
    test_phone_lengths = torch.tensor([200]).long()
    test_pitch = torch.randint(size=(1, 200), low=5, high=255)
    test_pitchf = torch.rand(1, 200)
    test_ds = torch.LongTensor([0])
    test_rnd = torch.rand(1, 192, 200)

    device = "cpu"
    net_g = SynthesizerTrnMsNSFsidM(
        *cpt["config"], is_half=False, version=cpt.get("version", "v1")
    )
    net_g.load_state_dict(cpt["weight"], strict=False)
    input_names = ["phone", "phone_lengths", "pitch", "pitchf", "ds", "rnd"]
    output_names = ["audio"]
    torch.onnx.export(
        net_g,
        (
            test_phone.to(device),
            test_phone_lengths.to(device),
            test_pitch.to(device),
            test_pitchf.to(device),
            test_ds.to(device),
            test_rnd.to(device),
        ),
        ExportedPath,
        dynamic_axes={
            "phone": [1],
            "pitch": [1],
            "pitchf": [1],
            "rnd": [2],
        },
        do_constant_folding=False,
        opset_version=18,
        verbose=False,
        input_names=input_names,
        output_names=output_names,
    )
    model, _ = onnxsim.simplify(ExportedPath)
    onnx.save(model, ExportedPath)
    return "Finished"

if __name__ == "__main__":
    import argparse
    parser = argparse.ArgumentParser()
    parser.add_argument("--model", required=True)
    parser.add_argument("--output", required=True)
    args = parser.parse_args()
    export_onnx(args.model, args.output)

[SkybuckFlying]

G:\Tools\Ultimate-RVC (Chinese)>type output.log
G:\Tools\Ultimate-RVC (Chinese)\infer\modules\onnx\export.py:8: FutureWarning: You are using torch.load with weights_only=False (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for weights_only will be flipped to True. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via torch.serialization.add_safe_globals. We recommend you start setting weights_only=True for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
cpt = torch.load(ModelPath, map_location="cpu")
C:\Python312\Lib\site-packages\torch\nn\utils\weight_norm.py:134: FutureWarning: torch.nn.utils.weight_norm is deprecated in favor of torch.nn.utils.parametrizations.weight_norm.
WeightNorm.apply(module, name, dim)
C:\Python312\Lib\site-packages\torch\onnx\utils.py:2139: UserWarning: No names were found for specified dynamic axes of provided input.Automatically generated names will be applied to each dynamic axes of input phone
warnings.warn(
C:\Python312\Lib\site-packages\torch\onnx\utils.py:2139: UserWarning: No names were found for specified dynamic axes of provided input.Automatically generated names will be applied to each dynamic axes of input pitch
warnings.warn(
C:\Python312\Lib\site-packages\torch\onnx\utils.py:2139: UserWarning: No names were found for specified dynamic axes of provided input.Automatically generated names will be applied to each dynamic axes of input pitchf
warnings.warn(
C:\Python312\Lib\site-packages\torch\onnx\utils.py:2139: UserWarning: No names were found for specified dynamic axes of provided input.Automatically generated names will be applied to each dynamic axes of input rnd
warnings.warn(
C:\Python312\Lib\site-packages\torch\onnx_internal\jit_utils.py:314: UserWarning: Constant folding - Only steps=1 can be constant folded for opset >= 10 onnx::Slice op. Constant folding not applied. (Triggered internally at C:\actions-runner_work\pytorch\pytorch\builder\windows\pytorch\torch\csrc\jit\passes\onnx\constant_fold.cpp:181.)
_C._jit_pass_onnx_node_shape_type_inference(node, params_dict, opset_version)
C:\Python312\Lib\site-packages\torch\onnx\utils.py:739: UserWarning: Constant folding - Only steps=1 can be constant folded for opset >= 10 onnx::Slice op. Constant folding not applied. (Triggered internally at C:\actions-runner_work\pytorch\pytorch\builder\windows\pytorch\torch\csrc\jit\passes\onnx\constant_fold.cpp:181.)
_C._jit_pass_onnx_graph_shape_type_inference(
C:\Python312\Lib\site-packages\torch\onnx\utils.py:1244: UserWarning: Constant folding - Only steps=1 can be constant folded for opset >= 10 onnx::Slice op. Constant folding not applied. (Triggered internally at C:\actions-runner_work\pytorch\pytorch\builder\windows\pytorch\torch\csrc\jit\passes\onnx\constant_fold.cpp:181.)
_C._jit_pass_onnx_graph_shape_type_inference(
grok

G:\Tools\Ultimate-RVC (Chinese)>pause
Press any key to continue . . .

[SkybuckFlying]

More adanced script to catch errors:

import torch
import onnxsim
import onnx
from infer.lib.infer_pack.models_onnx import SynthesizerTrnMsNSFsidM

def export_onnx(ModelPath, ExportedPath):
    try:
        print("grok")
        print(f"Loading model from {ModelPath}")
        cpt = torch.load(ModelPath, map_location="cpu", weights_only=False)
        print("Model loaded successfully")
        cpt["config"][-3] = cpt["weight"]["emb_g.weight"].shape[0]
        vec_channels = 256 if cpt.get("version", "v1") == "v1" else 768
        print(f"Vector channels: {vec_channels}")

        test_phone = torch.rand(1, 200, vec_channels)
        test_phone_lengths = torch.tensor([200]).long()
        test_pitch = torch.randint(size=(1, 200), low=5, high=255)
        test_pitchf = torch.rand(1, 200)
        test_ds = torch.LongTensor([0])
        test_rnd = torch.rand(1, 192, 200)

        device = "cpu"
        print("Initializing model")
        net_g = SynthesizerTrnMsNSFsidM(
            *cpt["config"], is_half=False, version=cpt.get("version", "v1")
        )
        net_g.load_state_dict(cpt["weight"], strict=False)
        net_g.eval()
        print("Model initialized successfully")

        input_names = ["phone", "phone_lengths", "pitch", "pitchf", "ds", "rnd"]
        output_names = ["audio"]
        print(f"Exporting ONNX to {ExportedPath}")
        torch.onnx.export(
            net_g,
            (
                test_phone.to(device),
                test_phone_lengths.to(device),
                test_pitch.to(device),
                test_pitchf.to(device),
                test_ds.to(device),
                test_rnd.to(device),
            ),
            ExportedPath,
            dynamic_axes={
                "phone": [1],
                "pitch": [1],
                "pitchf": [1],
                "rnd": [2],
            },
            do_constant_folding=False,
            opset_version=18,
            verbose=True,
            input_names=input_names,
            output_names=output_names,
        )
        print("ONNX export completed")

        print("Simplifying ONNX model")
        model, _ = onnxsim.simplify(ExportedPath)
        onnx.save(model, ExportedPath)
        print("ONNX simplification completed")
        return "Finished"
    except Exception as e:
        print(f"Error: {e}")
        raise

if __name__ == "__main__":
    import argparse
    parser = argparse.ArgumentParser()
    parser.add_argument("--model", required=True)
    parser.add_argument("--output", required=True)
    args = parser.parse_args()
    result = export_onnx(args.model, args.output)
    print(result)

[SkybuckFlying]

Some huge output generated but it seems to have worked, not yet sure if model is good:

Div_7"](%/dec/Add_12_output_0, %/dec/Constant_7_output_0), scope: infer.lib.infer_pack.models_onnx.SynthesizerTrnMsNSFsidM::/infer.lib.infer_pack.models_onnx.GeneratorNSF::dec # G:\Tools\Ultimate-RVC (Chinese)\infer\lib\infer_pack\models_onnx.py:509:0
%/dec/LeakyRelu_4_output_0 : Float(*, 32, , strides=[2048000, 64000, 1], requires_grad=1, device=cpu) = onnx::LeakyRelualpha=0.01, onnx_name="/dec/LeakyRelu_4", scope: infer.lib.infer_pack.models_onnx.SynthesizerTrnMsNSFsidM::/infer.lib.infer_pack.models_onnx.GeneratorNSF::dec # C:\Python312\Lib\site-packages\torch\nn\functional.py:1677:0
%/dec/conv_post/Conv_output_0 : Float(, 1, , strides=[64000, 64000, 1], requires_grad=0, device=cpu) = onnx::Conv[dilations=[1], group=1, kernel_shape=[7], pads=[3, 3], strides=[1], onnx_name="/dec/conv_post/Conv"](%/dec/LeakyRelu_4_output_0, %dec.conv_post.weight), scope: infer.lib.infer_pack.models_onnx.SynthesizerTrnMsNSFsidM::/infer.lib.infer_pack.models_onnx.GeneratorNSF::dec/torch.nn.modules.conv.Conv1d::conv_post # C:\Python312\Lib\site-packages\torch\nn\modules\conv.py:304:0
%audio : Float(, 1, *, strides=[64000, 64000, 1], requires_grad=1, device=cpu) = onnx::Tanhonnx_name="/dec/Tanh", scope: infer.lib.infer_pack.models_onnx.SynthesizerTrnMsNSFsidM::/infer.lib.infer_pack.models_onnx.GeneratorNSF::dec # G:\Tools\Ultimate-RVC (Chinese)\infer\lib\infer_pack\models_onnx.py:512:0
return (%audio)

grok
Loading model from G:\Tools\Ultimate-RVC (JackismyShephard)\models\rvc\voice_models\Transformers - Optimus Prime\OptimusPrime_550e_15950s.pth
Model loaded successfully
Vector channels: 768
Initializing model
Model initialized successfully
Exporting ONNX to G:\Tools\Ultimate-RVC (JackismyShephard)\models\rvc\voice_models\Transformers - Optimus Prime\OptimusPrime.onnx
ONNX export completed
Simplifying ONNX model
ONNX simplification completed
Finished

G:\Tools\Ultimate-RVC (Chinese)>pause
Press any key to continue . . .

[SkybuckFlying]

Check python script:

import onnx
model = onnx.load("G:\Tools\Ultimate-RVC (JackismyShephard)\models\rvc\voice_models\Transformers - Optimus Prime\OptimusPrime.onnx")
onnx.checker.check_model(model)
print("ONNX model is valid")

output:

G:\Tools\Ultimate-RVC (JackismyShephard)\models\rvc\voice_models\Transformers - Optimus Prime>python check.py
ONNX model is valid

G:\Tools\Ultimate-RVC (JackismyShephard)\models\rvc\voice_models\Transformers - Optimus Prime>

[SkybuckFlying]

import onnxruntime as ort
import numpy as np

Load the ONNX model

session = ort.InferenceSession("G:\Tools\Ultimate-RVC (JackismyShephard)\models\rvc\voice_models\Transformers - Optimus Prime\OptimusPrime.onnx")

Prepare dummy inputs (match the shapes used in export)

vec_channels = 768 # From the log: Vector channels: 768
test_phone = np.random.randn(1, 200, vec_channels).astype(np.float32)
test_phone_lengths = np.array([200], dtype=np.int64)
test_pitch = np.random.randint(low=5, high=255, size=(1, 200)).astype(np.int64)
test_pitchf = np.random.randn(1, 200).astype(np.float32)
test_ds = np.array([0], dtype=np.int64)
test_rnd = np.random.randn(1, 192, 200).astype(np.float32)

Run inference

inputs = {
"phone": test_phone,
"phone_lengths": test_phone_lengths,
"pitch": test_pitch,
"pitchf": test_pitchf,
"ds": test_ds,
"rnd": test_rnd
}
outputs = session.run(None, inputs)
print("Output shape:", outputs[0].shape)

output:

G:\Tools\Ultimate-RVC (JackismyShephard)\models\rvc\voice_models\Transformers - Optimus Prime>python interference.py
Output shape: (1, 1, 64000)

G:\Tools\Ultimate-RVC (JackismyShephard)\models\rvc\voice_models\Transformers - Optimus Prime>

if needed install:

pip install onnxruntime

So far so good, now for the ultimate testing, loading it in obs studio obs-rvc plugin...

[SkybuckFlying]

Nope, still crashes...

Adding crash file, maybe it will help somebody or the AI, yes AI can read it... going to link this in chinese issue, maybe they can follow along, don't know.

Crash 2025-10-20 13-50-19.txt

[JackismyShephard]

@SkybuckFlying Why do you need this exactly?