用tensorRT加速pytorch推理

使用的模型是比较简单的Unet，没有用pytorch那些复杂的函数

首先要把pytorch保存的pt模型转成onnx模型：

def save_onnx_illu():
    input_name = ['input']
    output_name = ['output']
    # 加载两个模型
    models_sdf = models.__dict__["KD_teacher"](input_channels=6)
    models_sdf = models_sdf.cuda()

    args = parse_args()
    with open('models_illu/%s/config.yml' % args.testname, 'r') as f:
        config = yaml.load(f, Loader=yaml.FullLoader)
    print('models_illu/%s/model.pth'%config['testname'])
    models_sdf.load_state_dict(torch.load('models_illu/%s/model.pth' %
                                          config['testname']))
    models_sdf.eval()
    input = Variable(torch.randn(1, 6, 1024, 1024)).cuda()
    torch.onnx.export(models_sdf, input, 'static_model/models_illu.onnx', input_names=input_name, output_names=output_name, verbose=False)