mit-han-lab / litepose

[CVPR'22] Lite Pose: Efficient Architecture Design for 2D Human Pose Estimation

Home Page:https://hanlab.mit.edu

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

How to get LitePose-Auto-L model with a coco dataset?

xddlj opened this issue · comments

commented

get the LitePose-Auto-L model coco dataset (map 62.5) is through Normal Training or Super-net Training?

To train network with search-L architecture with coco dataset, I use the following script:
python dist_train.py --cfg experiments/coco/mobilenet/mobile.yaml --superconfig mobile_configs/search-L.json

compare the provided LitePose-Auto-L coco model (map 62.5) with my model after training,the effect of my model is poor.

Have you adjusted the resolution in your config file? If not, just add it to your command:
python dist_train.py --cfg experiments/coco/mobilenet/mobile.yaml --superconfig mobile_configs/search-L.json DATASET.INPUT_SIZE 512 DATASET.OUTPUT_SIZE "[128, 256]"