bug fix

游雁

2024-12-23 1e5ef6ed9a6f64ecca7b9ef9481519b271f793a3

 funasr/utils/export_utils.py

@@ -2,6 +2,10 @@
import torch
import functools

import warnings

warnings.filterwarnings("ignore")


def export(
    model, data_in=None, quantize: bool = False, opset_version: int = 14, type="onnx", **kwargs
@@ -35,8 +39,16 @@
            if hasattr(m, "encoder") and hasattr(m, "decoder"):
                _bladedisc_opt_for_encdec(m, path=export_dir, enable_fp16=True)
            else:
                print(f"export_dir: {export_dir}")
                _torchscripts(m, path=export_dir, device="cuda")
        print("output dir: {}".format(export_dir))

        elif type == "onnx_fp16":
            assert (
                torch.cuda.is_available()
            ), "Currently onnx_fp16 optimization for FunASR only supports GPU"

            if hasattr(m, "encoder") and hasattr(m, "decoder"):
                _onnx_opt_for_encdec(m, path=export_dir, enable_fp16=True)

    return export_dir

@@ -50,17 +62,27 @@
    **kwargs,
):

    device = kwargs.get("device", "cpu")
    dummy_input = model.export_dummy_inputs()

    if isinstance(dummy_input, torch.Tensor):
        dummy_input = dummy_input.to(device)
    else:
        dummy_input = tuple([input.to(device) for input in dummy_input])

    verbose = kwargs.get("verbose", False)

    export_name = model.export_name + ".onnx"
    if isinstance(model.export_name, str):
        export_name = model.export_name + ".onnx"
    else:
        export_name = model.export_name()
    model_path = os.path.join(export_dir, export_name)
    torch.onnx.export(
        model,
        dummy_input,
        model_path,
        verbose=verbose,
        do_constant_folding=True,
        opset_version=opset_version,
        input_names=model.export_input_names(),
        output_names=model.export_output_names(),
@@ -68,25 +90,30 @@
    )

    if quantize:
        from onnxruntime.quantization import QuantType, quantize_dynamic
        import onnx
        try:
            from onnxruntime.quantization import QuantType, quantize_dynamic
            import onnx
        except:
            raise RuntimeError(
                "You are quantizing the onnx model, please install onnxruntime first. via \n`pip install onnx`\n`pip install onnxruntime`."
            )

        quant_model_path = model_path.replace(".onnx", "_quant.onnx")
        if not os.path.exists(quant_model_path):
            onnx_model = onnx.load(model_path)
            nodes = [n.name for n in onnx_model.graph.node]
            nodes_to_exclude = [
                m for m in nodes if "output" in m or "bias_encoder" in m or "bias_decoder" in m
            ]
            quantize_dynamic(
                model_input=model_path,
                model_output=quant_model_path,
                op_types_to_quantize=["MatMul"],
                per_channel=True,
                reduce_range=False,
                weight_type=QuantType.QUInt8,
                nodes_to_exclude=nodes_to_exclude,
            )
        onnx_model = onnx.load(model_path)
        nodes = [n.name for n in onnx_model.graph.node]
        nodes_to_exclude = [
            m for m in nodes if "output" in m or "bias_encoder" in m or "bias_decoder" in m
        ]
        print("Quantizing model from {} to {}".format(model_path, quant_model_path))
        quantize_dynamic(
            model_input=model_path,
            model_output=quant_model_path,
            op_types_to_quantize=["MatMul"],
            per_channel=True,
            reduce_range=False,
            weight_type=QuantType.QUInt8,
            nodes_to_exclude=nodes_to_exclude,
        )


def _torchscripts(model, path, device="cuda"):
@@ -100,7 +127,12 @@
            dummy_input = tuple([i.cuda() for i in dummy_input])

    model_script = torch.jit.trace(model, dummy_input)
    model_script.save(os.path.join(path, f"{model.export_name}.torchscript"))
    if isinstance(model.export_name, str):
        model_script.save(os.path.join(path, f"{model.export_name}".replace("onnx", "torchscript")))
    else:
        model_script.save(
            os.path.join(path, f"{model.export_name()}".replace("onnx", "torchscript"))
        )


def _bladedisc_opt(model, model_inputs, enable_fp16=True):
@@ -153,7 +185,7 @@

    # Rescale encoder modules
    fp16_scale = int(2 * absmax // 65536)
    print(f"rescale encoder modules with factor={fp16_scale}")
    print(f"rescale encoder modules with factor={fp16_scale}\n\n")
    model.encoder.model.encoders0.register_forward_pre_hook(
        functools.partial(_rescale_input_hook, scale=fp16_scale),
    )
@@ -194,3 +226,62 @@
    model.decoder = _bladedisc_opt(model.decoder, tuple(decoder_inputs))
    model_script = torch.jit.trace(model, input_data)
    model_script.save(os.path.join(path, f"{model.export_name}_blade.torchscript"))


def _onnx_opt_for_encdec(model, path, enable_fp16):

    # Get input data
    # TODO: better to use real data
    input_data = model.export_dummy_inputs()

    if isinstance(input_data, torch.Tensor):
        input_data = input_data.cuda()
    else:
        input_data = tuple([i.cuda() for i in input_data])

    # Get input data for decoder module
    decoder_inputs = list()

    def get_input_hook(m, x):
        decoder_inputs.extend(list(x))

    hook = model.decoder.register_forward_pre_hook(get_input_hook)
    model = model.cuda()
    model(*input_data)
    hook.remove()

    # Prevent FP16 overflow
    if enable_fp16:
        _rescale_encoder_model(model, input_data)

    fp32_model_path = f"{path}/{model.export_name}_hook.onnx"
    print("*" * 50)
    print(f"[_onnx_opt_for_encdec(fp32)]: {fp32_model_path}\n\n")
    if not os.path.exists(fp32_model_path):

        torch.onnx.export(
            model,
            input_data,
            fp32_model_path,
            verbose=False,
            do_constant_folding=True,
            opset_version=13,
            input_names=model.export_input_names(),
            output_names=model.export_output_names(),
            dynamic_axes=model.export_dynamic_axes(),
        )

    # fp32 to fp16
    fp16_model_path = f"{path}/{model.export_name}_hook_fp16.onnx"
    print("*" * 50)
    print(f"[_onnx_opt_for_encdec(fp16)]: {fp16_model_path}\n\n")
    if os.path.exists(fp32_model_path) and not os.path.exists(fp16_model_path):
        try:
            from onnxconverter_common import float16
        except:
            raise RuntimeError(
                "You are converting the onnx model to fp16, please install onnxconverter-common first. via `pip install onnxconverter-common`."
            )
        fp32_onnx_model = onnx.load(fp32_model_path)
        fp16_onnx_model = float16.convert_float_to_float16(fp32_onnx_model, keep_io_types=True)
        onnx.save(fp16_onnx_model, fp16_model_path)