ubuntu 20.04 gpu: GTX 1060 6GB
- main environment
- conda 4.8.3
- python 3.6.12
- harness gpu
- cuda 10.0
- cudnn 7.6.5
- tensorflow 1.15
- keras 2.3.1
- visualization
- matplotlib 3.3.1
- seaborn 0.11.1
- process data
- pandas 1.1.1
- sklearn 0.23.2
$ python3 video2pickle.py --video [video_name] --savefile [file_name_to_save]
We used ildoonet/tf-pose-estimation to extract each body part informations save informations to pickle
$ python3 preprocessing.py --rawroot [raw_file_name]
In preprocessing.py...
- top: Nos, Lea, Ley, Rea, Rey
- mid: Nec, Lel, Lsh, Rel, Rsh
- Calculate the variations of the top and mid part per every 50 frames and add labels.
$ python3 build_trainset.py --name [person_initial] --index [index_number]
merge dataset and shuffle to prevent biased labeled value
$ python3 run_dnn.py --file [name of pickle] --plot [graph_idx] --size [dataset_size] --epoch [number of epoch]
- 1st hidden layer: dimension:8, activation: relu
- 2nd hidden layer: dimension:8, activation: relu
- output layer: sigmoid
compared between size of data