app.py

import os
import gradio as gr
import time
import numpy as np
import onnx
import onnxruntime as ort
import torch
from onnxconverter_common.float16 import convert_float_to_float16
from onnxslim import slim
from spandrel import ModelLoader, ImageModelDescriptor
import spandrel_extra_arches
from rich.traceback import install

def get_out_path(out_dir: str, name: str, opset: int, fp16: bool = False, optimized: bool = False, static: bool = False) -> str:
    filename = f"{name}_fp{'16' if fp16 else '32'}{'_static' if static else ''}_op{opset}{'_onnxslim' if optimized else ''}.onnx"
    return os.path.normpath(os.path.join(out_dir, filename))

def convert_and_save_onnx(model, name: str, torch_input, out_dir: str, opset: int, use_static_shapes: bool) -> tuple[onnx.ModelProto, str]:
    if use_static_shapes:
        dynamic_axes = None
        input_names = None
        output_names = None
        #input_names = ["input"]
        #output_names = ["output"]
    else:
        dynamic_axes = {
            "input": {0: "batch_size", 2: "width", 3: "height"},
            "output": {0: "batch_size", 2: "width", 3: "height"},
        }
        input_names = ["input"]
        output_names = ["output"]

    out_path = get_out_path(out_dir, name, opset, False, False, use_static_shapes)

    #this class was taken from chainner. Running the model through this seems to fix some issues with various arches.
    class FakeModel(torch.nn.Module):
        def __init__(self, model: ImageModelDescriptor):
            super().__init__()
            self.model = model

        def forward(self, x: torch.Tensor):
            return self.model(x)

    model = FakeModel(model)

    torch.onnx.export(
        model,
        (torch_input,),
        out_path,
        dynamo=False,
        verbose=False,
        opset_version=opset,
        dynamic_axes=dynamic_axes,
        input_names=input_names,
        output_names=output_names,
    )
    model_proto = onnx.load(out_path)
    return model_proto, out_path

def verify_onnx(model, torch_input, onnx_path: str) -> None:
    with torch.inference_mode():
        torch_output_np = model(torch_input).cpu().numpy()

    onnx_model = onnx.load(onnx_path)
    onnx.checker.check_model(onnx_model)

    try:
        ort_session = ort.InferenceSession(onnx_path, providers=["CPUExecutionProvider"])
        ort_inputs = {ort_session.get_inputs()[0].name: torch_input.cpu().numpy()}
        onnx_output = ort_session.run(None, ort_inputs)
        np.testing.assert_allclose(
            torch_output_np,
            onnx_output[0],
            rtol=1e-02,
            atol=1e-03,
        )
        print("ONNX output verified against PyTorch output successfully.")
    except AssertionError as e:
        print(f"ONNX verification completed with warnings: {e}")
        gr.Warning("ONNX verification completed with warnings")
    except Exception as e:
        print(f"ONNX verification failed: {e}")
        gr.Warning("ONNX verification failed")


def convert_pipeline(model_path: str, opset: int = 17, verify: bool = True, optimize: bool = True, fp16: bool = False, static: bool = False) -> str:
    loader = ModelLoader()
    model_desc = loader.load_from_file(model_path)
    assert isinstance(model_desc, ImageModelDescriptor)

    model = model_desc.model.to("cpu").eval()
    model_name = os.path.splitext(os.path.basename(model_path))[0]

    # Generate dummy input
    if static:
        height, width = 256, 256
        torch_input = torch.randn(1, model_desc.input_channels, height, width, device="cpu")
    else:
        torch_input = torch.randn(1, model_desc.input_channels, 32, 32, device="cpu")

    out_dir = "./onnx"
    os.makedirs(out_dir, exist_ok=True)

    # Convert to ONNX
    start_time = time.time()
    model_proto, out_path_fp32 = convert_and_save_onnx(
        model, model_name, torch_input, out_dir, opset, static
    )
    out_path = out_path_fp32

    print(f"Saved to {out_path_fp32} in {time.time() - start_time:.2f} seconds.")

    # Verify
    if verify:
        verify_onnx(model, torch_input, out_path_fp32)

    # Optimize
    if optimize:
        #model_proto = slim(model_proto)
        session_opt = ort.SessionOptions()
        session_opt.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_EXTENDED
        session_opt.optimized_model_filepath = get_out_path(out_dir, model_name, opset, False, True, static)
        ort.InferenceSession(out_path_fp32, session_opt)
        if verify:
            verify_onnx(model, torch_input, session_opt.optimized_model_filepath)
        model_proto = onnx.load(session_opt.optimized_model_filepath)
        out_path = session_opt.optimized_model_filepath

    # Convert to FP16
    if fp16:
        start_time = time.time()
        out_path = get_out_path(out_dir, model_name, opset, True, optimize, static)
        model_proto_fp16 = convert_float_to_float16(model_proto)
        onnx.save(model_proto_fp16, out_path)
        print(f"Saved to {out_path_fp32} in {time.time() - start_time:.2f} seconds.")

    return out_path

def load_model(model_path: str):
    if not model_path:
        return "Ready"
    
    loader = ModelLoader()
    try:
        model = loader.load_from_file(model_path)
        assert isinstance(model, ImageModelDescriptor)

        architecture_info = {
            'architecture_name': getattr(model.architecture, 'name', str(model.architecture)),
            'input_channels': model.input_channels,
            'output_channels': model.output_channels,
            'scale': model.scale,
            'tags': model.tags,
            'supports_fp16': model.supports_half
            #'supports_bf16': model.supports_bfloat16,
            #'size_requirements': model.size_requirements
        }
        
        if model.supports_half:
            return [str(architecture_info), gr.Radio(choices=["True", "False"], interactive=True, label="FP16 - Export at half precision. Not supported by all models.")]
        else: 
            return [str(architecture_info), gr.Radio(choices=["True", "False"], value="False", interactive=False, label="FP16 - Export at half precision. Not supported by all models.")]
        
    except Exception as e:
        return [f"Error loading model: {e}", gr.Radio(choices=["True", "False"], interactive=True, label="FP16 - Export at half precision. Not supported by all models.")]

def process_choices(opset, fp16, static, slim, file):
    if not file:
        print("No file loaded.")
        gr.Warning("No file loaded.")
        yield [gr.Button("Convert", interactive=True), gr.DownloadButton(label="💾 Download Converted Model", visible=False)]
        return
    
    # Convert string choices to boolean
    fp16 = fp16 == "True"
    static = static == "True" 
    slim = slim == "True"
    
    yield [gr.Button("Processing", interactive=False), gr.DownloadButton(label="💾 Download Converted Model", visible=False)]
    
    try:
        result = convert_pipeline(file, opset, True, slim, fp16, static)
        short_name = os.path.basename(result)
        yield [gr.Button("Convert", interactive=True), gr.DownloadButton(label=f"💾 {short_name}", value=result, visible=True)]
        return
    except Exception as e:
        print(f"{e}")
        gr.Warning("Conversion error.")
        yield [gr.Button("Convert", interactive=True), gr.DownloadButton(label="💾 Download Converted Model", visible=False)]
        return

# Create Gradio interface
with gr.Blocks(title="PTH to ONNX Converter") as demo:
    install()
    spandrel_extra_arches.install()

    file_upload = gr.File(label="Upload a PyTorch model", file_types=['.pth', '.pt', '.safetensors'])
    metadata = gr.Textbox(value="Ready", label="File Information")
    
    dropdown_opset = gr.Dropdown(choices=[17, 18, 19, 20], value=20, label="Opset")
    radio_fp16 = gr.Radio(choices=["True", "False"], value="False", label="FP16 - Not supported by all models. Not very useful because FP16 TRT engines can still be built from FP32 ONNX models.")
    radio_static = gr.Radio(choices=["True", "False"], value="False", label="Static Shapes - Might be required by some models, but can cause slower performance.")
    radio_slim = gr.Radio(choices=["True", "False"], value="False", visible=False, label="OnnxSlim - Can cause issues in some models. I have not yet found any cases where it helps. May remove in the future.") #turn off for now
    
    gr.Markdown("After converting, click the logs button at the top to check for any errors or warnings.")
    process_button = gr.Button("Convert", interactive=True)
    file_output = gr.DownloadButton(label="💾 Download Converted Model", visible=False)

    gr.Markdown("""
    # Resources
    - [OpenModelDB](https://openmodeldb.info): Find upscaling models here
    - [VideoJaNai](https://github.com/the-database/VideoJaNai): For upscaling videos using ONNX models
    - [REAL Video Enhancer](https://github.com/TNTwise/REAL-Video-Enhancer): For upscaling videos using ONNX models
    """)
    
    process_button.click(fn=process_choices, 
                        inputs=[dropdown_opset, radio_fp16, radio_static, radio_slim, file_upload], 
                        outputs=[process_button, file_output])
    file_upload.upload(fn=load_model, inputs=file_upload, outputs=[metadata, radio_fp16])

if __name__ == "__main__":
    demo.launch(show_error=True, inbrowser=True, show_api=False, debug=False)