操作步骤：

1. 先验证官方模型的预测结果：

```

python tools/infer.py -c configs/centernet/centernet_dla34_140e_coco.yml -o use_gpu=false weights=https://bj.bcebos.com/v1/paddledet/models/centernet_dla34_140e_coco.pdparams --infer_img=demo/000000570688.jpg

```

结果正确。

2. 导出官方模型：

```

python tools/export_model.py -c configs/centernet/centernet_dla34_140e_coco.yml --output_dir=./inference_model -o use_gpu=false weights=https://bj.bcebos.com/v1/paddledet/models/centernet_dla34_140e_coco.pdparams

```

3. 用导出的模型来预测同一个图片：

```

python deploy/python/infer.py --model_dir inference_model/centernet_dla34_140e_coco/ --device CPU --output_dir infer_output/centernet_dla34_140e_coco/ --image_file demo/000000570688.jpg

```

结果错误，表现为预测框错位。

错误的预测结果图片如下：


预测时的控制台输出为：

```

----------- Running Arguments -----------

batch_size: 1

camera_id: -1

cpu_threads: 1

device: CPU

enable_mkldnn: False

image_dir: None

image_file: demo/000000570688.jpg

model_dir: inference_model/centernet_dla34_140e_coco/

output_dir: infer_output/centernet_dla34_140e_coco/

reid_batch_size: 50

reid_model_dir: None

run_benchmark: False

run_mode: paddle

save_images: False

save_mot_txt_per_img: False

save_mot_txts: False

scaled: False

threshold: 0.5

trt_calib_mode: False

trt_max_shape: 1280

trt_min_shape: 1

trt_opt_shape: 640

use_dark: True

use_gpu: False

video_file: None

------------------------------------------

----------- Model Configuration -----------

Model Arch: CenterNet

Transform Order:

--transform op: WarpAffine

--transform op: NormalizeImage

--transform op: Permute

--------------------------------------------

class_id:0, confidence:0.7541, left_top:[340.11,342.66],right_bottom:[383.31,463.95]

class_id:33, confidence:0.6722, left_top:[584.21,195.76],right_bottom:[603.37,214.80]

class_id:0, confidence:0.6665, left_top:[482.51,347.15],right_bottom:[529.86,479.07]

class_id:0, confidence:0.6513, left_top:[257.18,317.54],right_bottom:[337.76,470.28]

class_id:0, confidence:0.6493, left_top:[598.02,360.13],right_bottom:[662.09,477.77]

class_id:33, confidence:0.6432, left_top:[403.26,161.21],right_bottom:[544.81,218.00]

class_id:33, confidence:0.6079, left_top:[409.43,64.38],right_bottom:[468.53,86.07]

class_id:33, confidence:0.6043, left_top:[177.51,315.77],right_bottom:[242.79,360.59]

class_id:33, confidence:0.5843, left_top:[321.80,124.17],right_bottom:[343.96,135.97]

class_id:33, confidence:0.5681, left_top:[256.50,235.80],right_bottom:[334.02,258.06]

class_id:33, confidence:0.5640, left_top:[312.54,21.85],right_bottom:[323.61,29.44]

class_id:33, confidence:0.5600, left_top:[539.81,252.46],right_bottom:[558.75,268.81]

class_id:0, confidence:0.5552, left_top:[106.13,339.19],right_bottom:[137.33,424.96]

class_id:33, confidence:0.5445, left_top:[442.35,120.70],right_bottom:[455.25,130.22]

class_id:33, confidence:0.5360, left_top:[295.58,177.89],right_bottom:[532.20,242.15]

class_id:33, confidence:0.5320, left_top:[358.71,217.07],right_bottom:[501.16,266.67]

class_id:0, confidence:0.5315, left_top:[577.07,381.62],right_bottom:[610.70,479.78]

class_id:0, confidence:0.5211, left_top:[530.56,370.44],right_bottom:[585.49,480.41]

class_id:33, confidence:0.5056, left_top:[511.19,30.82],right_bottom:[530.04,44.08]

class_id:0, confidence:0.5018, left_top:[252.27,380.57],right_bottom:[288.42,417.50]

save result to: infer_output/centernet_dla34_140e_coco/000000570688.jpg

------------------ Inference Time Info ----------------------

total_time(ms): 9323.2, img_num: 1

average latency time(ms): 9323.20, QPS: 0.107259

preprocess_time(ms): 52.00, inference_time(ms): 9271.20, postprocess_time(ms): 0.00