PyTorch 模型 onnx 文件导出及调用详情

import torch
from torchvision.models import resnet50
 
file = 'resnet.onnx'
# 声明模型
resnet = resnet50(pretrained=False).eval()
image = torch.rand([1, 3, 224, 224])
# 导出为 onnx 文件
torch.onnx.export(resnet, (image,), file)

# 选其一即可
pip install onnxruntime        # CPU 版本
pip install onnxruntime-gpu    # GPU 版本

import onnxruntime as ort
import numpy as np
file = 'resnet.onnx'
# 找到 GPU / CPU
provider = ort.get_available_providers()[
    1 if ort.get_device() == 'GPU' else 0]
print('设备:', provider)
# 声明 onnx 模型
model = ort.InferenceSession(file, providers=[provider])
# 参考: ort.NodeArg
for node_list in model.get_inputs(), model.get_outputs():
    for node in node_list:
        attr = {'name': node.name,
                'shape': node.shape,
                'type': node.type}
        print(attr)
    print('-' * 60)
 
# 得到输入、输出结点的名称
input_node_name = model.get_inputs()[0].name
ouput_node_name = [node.name for node in model.get_outputs()]
image = np.random.random([1, 3, 224, 224]).astype(np.float32)
print(model.run(output_names=ouput_node_name,
                input_feed={input_node_name: image}))

class Onnx_Module(ort.InferenceSession):
    ''' onnx 推理模型
        provider: 优先使用 GPU'''
    provider = ort.get_available_providers()[
        1 if ort.get_device() == 'GPU' else 0]
 
    def __init__(self, file):
        super(Onnx_Module, self).__init__(file, providers=[self.provider])
        # 参考: ort.NodeArg
        self.inputs = [node_arg.name for node_arg in self.get_inputs()]
        self.outputs = [node_arg.name for node_arg in self.get_outputs()]
 
    def __call__(self, *arrays):
        input_feed = {name: x for name, x in zip(self.inputs, arrays)}
        return self.run(self.outputs, input_feed)

import numpy as np
file = 'resnet.onnx'
model = Onnx_Module(file)
image = np.random.random([1, 3, 224, 224]).astype(np.float32)
print(model(image))

class Timer:
    repeat = 3
 
    def __new__(cls, fun, *args, **kwargs):
        import time
        start = time.time()
        for _ in range(cls.repeat): fun(*args, **kwargs)
        cost = (time.time() - start) / cls.repeat
        return cost * 1e3  # ms
 
 
class Onnx_Module(ort.InferenceSession):
    ''' onnx 推理模型
        provider: 优先使用 GPU'''
    provider = ort.get_available_providers()[
        1 if ort.get_device() == 'GPU' else 0]
 
    def __init__(self, file):
        super(Onnx_Module, self).__init__(file, providers=[self.provider])
        # 参考: ort.NodeArg
        self.inputs = [node_arg.name for node_arg in self.get_inputs()]
        self.outputs = [node_arg.name for node_arg in self.get_outputs()]
    def __call__(self, *arrays):
        input_feed = {name: x for name, x in zip(self.inputs, arrays)}
        return self.run(self.outputs, input_feed)
 
    @classmethod
    def test(cls, model, args, file, **export_kwargs):
        # 测试 Torch 的运行时间
        torch_output = model(*args).data.numpy()
        print(f'Torch: {Timer(model, *args):.2f} ms')
        # model: Torch -> onnx
        torch.onnx.export(model, args, file, **export_kwargs)
        # data: tensor -> array
        args = tuple(map(lambda tensor: tensor.data.numpy(), args))
        onnx_model = cls(file)
        # 测试 onnx 的运行时间
        onnx_output = onnx_model(*args)
        print(f'Onnx: {Timer(onnx_model, *args):.2f} ms')
        # 计算 Torch 模型与 onnx 模型输出的绝对误差
        abs_error = np.abs(torch_output - onnx_output).mean()
        print(f'Mean Error: {abs_error:.2f}')
        return onnx_model