howiema / cpm_pytorch Goto Github PK

This is the Pytorch

There are 7 files in this folder

--handpose_data_cpm.py
data loader for Hand Pose dataset

--handpose_no_label.py
data loader for Hand Pose dataset without ground truth

--cpm.py
Pytorch cpm model

--train.py
--test.py --save.py
--predict.py
--conf.text

python cpm_train.py   

You may revise the variable in conf.text

train_data_dir =
train_label_dir =
learning_rate = 8e-6
batch_size = 16
epochs = 50
begin_epoch = 0

Thus change the path to your own datasets and train CPM on your own REMEMBER that you may implement new data loader for you own datasets.

After this, you will get models for several epoches. The models are saved in folder ckpt/ like

ckpt/model_epoch10.pth

python cpm_Test.py         

After running this, you will get PCK score for each epoch
You can select the best trained models

python cpm_save.py    

After step 2, you will know which is the best epoch, thus you should revise conf.text and change the value of best_model

python cpm_predict.py    

This step is for applying trained model on datasets without ground truth