【pytorch】onnx

pytorch任意形式的model（.t7、.pth等等）转.onnx全都可以采用固定格式。

完整实现：

def pth2onnx(self, simplify_onnx_sw=True):
    import torch
    os.environ['KMP_DUPLICATE_LIB_OK'] = 'True'

    model = torch.nn.DataParallel(self.model)
    _state_dict = torch.load(pth_path, map_location=torch.device('cpu'))
    model.load_state_dict(_state_dict, strict=True)
    model.eval()
    torch.onnx.export(model.module,
                      torch.randn(batch_size, *C.input_shape),
                      pure_onnx_path,
                      input_names=["input"],
                      output_names=["output"]
                      )

    if simplify_onnx_sw:
        os.system('python -m onnxsim {} {}'.format(pure_onnx_path, simplified_onnx_path))
        print('n Simplified onnx has been save to {}n'.format(simplified_onnx_path))
        os.remove(pure_onnx_path)
    else:
        print('n Pure onnx has been save to {}n'.format(pure_onnx_path))

实验举例：

model_dir = './'
pth_path = model_dir + 'A.pth'
onnx_path = model_dir + 'A.onnx'
batch_size = 1
input_shape = (3, 112, 112)

cfg = Config()
cfg.load_from_file(args.model_cfg_file)

model = PFLD_SE3_eval(cfg.model_conf.layer_cfg, cfg.model_conf.num_points)

model.load(pth_path)
model.eval()
torch.onnx.export(model,
                  torch.randn(batch_size, *input_shape),
                  onnx_path,
                  input_names=["input"],
                  output_names=["output_0", "output_1"],
                  )

print('nn onnx has been save to {}nn'.format(onnx_path))
如在mac下执行，还需要加上这行环境配置： 
os.environ['KMP_DUPLICATE_LIB_OK']='True'

可能的报错：

importError: cannot import name 'get_all_providers' from 'onnxruntime.capi._pybind_state' 

mac下的通用解决方法：

brew install libomp

如果还是报相同错误，则可能是版本问题。换版本即可。例如我是执行：

pip install onnxruntime==1.2.0