chenyilun95/tf-cpn

using dataset COCO2017

vacing opened this issue · 1 comments

vacing commented

使用COCO2017数据集的修改点：

  1. data/COCO/COCOAllJoints.py 读取数据未使用config.py的配置，而是内部硬编码的。
    1. 将cfg传入COCOJoints(cfg)的构造函数(cfg is a global variable, can be used any where)
  2. data/COCO/COCOAllJoints.py-> COCOJoints() 生成image路径时，强依赖2014年数据集，需要修改成：train2017/xxxx.jpg模式
  3. models/COCO.res101.384x288.CPN/config.py，多进程会导致GPU内存分配错误，关闭：dpflow_enable = False
  1. data/COCO/COCOAllJoints.py dosen't use configuration in config.py but uses inner hardcoded path
  2. data/COCO/COCOAllJoints.py-> COCOJoints(), when generate image paths from annotation, the way it uses can't match file name pattern in coco2017 any more
  3. models/COCO.res101.384x288.CPN/config.py : dpflow_enable enble may cause OOM, if so disable this configuration
mathpopo commented

dpflow_enable = False can use