update 2019-08-14
Add tensorflow 2.0 version, link here. Repo still need to update. Welcome to raise questions.
update 2019-04-01
paf loss
weight in total loss. Previous cpm_loss : paf_loss == 1:1
and now is '1:2'update 2019-03-20
Add evaluation part. The evaluation based on ai-challenger evaluation solution. More information please refer to this link.
Easy way for how to use it:
test&model/model_json.py
to generate a json file. The param
in this file contains everything you need.
Remember that param['img_path] && param['jason_file']
parameters is the groundtruth test files that you need to test.Then, run test&model/model_eval.py
, the command line would be like this:
python model_eval.py --submit predictions.json --ref groundtruth.json
this will give you a score that about your model performance, between 0~1
, 0 is worst and 1 is best.
Make sure that you need run python2
instead of python3
because some errors will occur in eval.py
file.
update 2019-03-18
Based on official pytorch implementation, re-write tf-model, see lightweight_openpose.py
for detailed information, corresponding net structure picture is named lightweight.png
. New pre_trained
model will be upload soon.
A tensorflow implementation about Arxiv Paper "Real-time 2D Multi-Person Pose Estimation on CPU: Lightweight OpenPose"
Thanks for the author provide PyTorch version
(This repo is more useful than mine i think, enjoy it!) Pytorch implementation.
trained model
model
folder, model.ckpt-1008540.*
.model
folder, named model.ckpt-61236
, on ai_test_A dataset, get score 0.0377, so bad.update
The original caffe prototxt(provided by paper author) has been upload, you can found in repo file named "lightweight_openpose.prototxt"
Requirement
Train
python3 train.py
all parameters has been set in src/train_config.py
.
Train Dataset
we use ai-challenger format dataset, which can found in this website.
Note
dataset.repeat(1)
. And use make_initializable_iterator()
instead of `make_one_shot_iterator().