General code for pytorch model to libtorch and onnx format

依赖

• torch
• onnx
• onnx simplifer

Important parameters that need to be set by yourself

• model_path 模型权重路径
• model 网络实例
• inp 样例输入,就是一个shape合法的tensor,batchsize（第一维）设置为1就行

下面以torchvision自带的resnet101模型为例.The weights are using the official pretrained model,调用resnet101(pretrained=True)will be downloaded automatically%USERPROFILE%/.cache/torch/hub下面

import onnx
import torch
from torch.utils.mobile_optimizer import optimize_for_mobile
from torchvision.models.resnet import resnet101

from utils.func import file_size, colorstr

model_path = './weights/resnet101.pth'  # 模型权重路径
model = resnet101()  # 模型对象
height, width = 640, 640
inp = torch.zeros([1, 3, height, width])  # 样例输入,用于trace
# common
half = True  # fp16量化
# onnx profile
onnx_export = True  # 是否输出onnx格式
opset_version = 13  # 算子集版本
dynamic = False  # Whether to enter dynamicallybatchsize,The following two options need to be set
input_names = ['inputs']
dynamic_axes = {
    'inputs': {
    0: 'batch', 1: 'kp28'},  # 动态batchsize设置
                'output': {
    0: 'batch', 1: 'classes'}}
simplify = True  # whether to simplify
# libtorch profile
libtorch_export = True  # 是否输出libtorch格式
optimize = False  # 针对移动端优化,Not for mobile use
strict = False  # 严格模式,设置False就行

if __name__ == '__main__':
    model.load_state_dict(torch.load(model_path))
    model.cpu().eval()

    if half:
        inp, model = inp.half(), model.half()
    if onnx_export:
        prefix = colorstr('ONNX:')
        f = model_path.replace('.pth', '.onnx')  # filename

        torch.onnx.export(model, inp, f, verbose=False, opset_version=opset_version, input_names=input_names,
                          training=torch.onnx.TrainingMode.EVAL,
                          do_constant_folding=True,
                          dynamic_axes=dynamic_axes if dynamic else None)
        # Checks
        model_onnx = onnx.load(f)  # load onnx model
        onnx.checker.check_model(model_onnx)  # check onnx model
        # print(onnx.helper.printable_graph(model_onnx.graph)) # print

        # Simplify
        if simplify:
            try:
                import onnxsim

                print(f'simplifying with onnx-simplifier {
      onnxsim.__version__}...')
                model_onnx, check = onnxsim.simplify(
                    model_onnx,
                    dynamic_input_shape=dynamic,
                    input_shapes={
    'images': list(inp.shape)} if dynamic else None)
                assert check, 'assert check failed'
                onnx.save(model_onnx, f)
            except Exception as e:
                print(f'{
      prefix} simplifier failure: {
      e}')
        print(f'{
      prefix} export success, saved as {
      f} ({
      file_size(f):.1f} MB)')

    if libtorch_export:
        prefix = colorstr('TorchScript:')
        try:
            print(f'\n{
      prefix} starting export with torch {
      torch.__version__}...')
            f = model_path.replace('.pt', '.torchscript.pt')  # filename
            ts = torch.jit.trace(model, inp, strict=strict)
            (optimize_for_mobile(ts) if optimize else ts).save(f)
            print(f'{
      prefix} export success, saved as {
      f} ({
      file_size(f):.1f} MB)')
        except Exception as e:
            print(f'{
      prefix} export failure: {
      e}')