…# Pytorch Implementation of Region Ensemble Network
Pull requests are welcome.
I did not replicate the error results as that was not the goal of my project.
Results during the demo are pretty good.
I also added another option to change the input size. The network will splice the features according. It will give an assertion error if the input size results in the features dimension being odd number.
pytorch 0.4.0
opencv 3.4.0
Right now the code only supports the MSRA Hand Gesture dataset. If I have time, i will add support for IVCL and NYU.
Put the MSRA dataset in data/
data/
data/P0/..
data/P1/1/000000.bin
... and several other similar folders...
It is assumed that CUDA is installed.
python main.py --name experiment_1 --save_dir experiments/ --batchSize 128
To resume training, add the checkpoint option accordingly.
python main.py --checkpoint checkpoint.pth.tar
To use a model as a pretrained checkpoint/finetuning,
python main.py --checkpoint checkpoint.pth.tar --finetune
If you only want to train on certain poses/subjects,
python main.py --poses 1 2 3 --persons 1 2 3
The train loss and validation loss will be saved in an out file in save_dir
Many thanks to the original authors of the paper who readily and kindly offered help