1.日志分析 analyze_logs.py
https://blog.csdn.net/jy1023408440/article/details/105701705
2.可视化数据集 browse_dataset.py
python tools/browse_dataset.py m1/faster_rcnn_r50_fpn_1x_coco.py
![](https://img-blog.csdnimg.cn/20210111112404238.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2p5MTAyMzQwODQ0MA==,size_16,color_FFFFFF,t_70)
ConfigDict' object has no attribute 'pipeline'
报错,不过我用的voc格式的个人数据集
3.模型复杂度 get_flops.py
python tools/get_flops.py m1/faster_rcnn_r50_fpn_1x_coco.py
![](https://img-blog.csdnimg.cn/20210111113048728.png)
4.发布模型 publish_model.py
python tools/publish_model.py m1/f16/epoch_12.pth m1/f16/ys.pth (FP16)
![](https://img-blog.csdnimg.cn/20210111113614751.png)
模型从236.47MB--->79.38MB
79.38/236.47=0.335687 只有原来的1/3
下面是没有经过FP16训练的。模型本身较大,经过清除还是较大。
![](https://img-blog.csdnimg.cn/2021011111370238.png)
模型从315.32MB--->158.23MB
158.23/315.32=0.5018 原来的1/2
5.评估指标eval_metric.py
![](https://img-blog.csdnimg.cn/20210111134611263.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2p5MTAyMzQwODQ0MA==,size_16,color_FFFFFF,t_70)
python tools/eval_metric.py m1/faster_rcnn_r50_fpn_1x_coco.py m1/faster/ep10.pkl --eval mAP
![](https://img-blog.csdnimg.cn/20210111134704660.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2p5MTAyMzQwODQ0MA==,size_16,color_FFFFFF,t_70)
6.打印config文件 tools/print_config.py
![](https://img-blog.csdnimg.cn/20210111135411309.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2p5MTAyMzQwODQ0MA==,size_16,color_FFFFFF,t_70)
会将改动的,多层嵌套的config一次性打印出来