P-I3D
Where to focus on for Human Action Recognition?
REQUIRED PACKAGES AND DEPENDENCIES
- python 3.6.8
- PyTorch 1.0.1
- Torchvision 0.2.0
- Tensorflow 1.13.0 (GPU compatible)
- keras 2.3.1
- scikit-image 0.16.2
- Pillow 6.2.1
- OpenCV 4.1.2
- Cuda 10.0
- CuDNN 7.4
- tqdm 4.41.1
INSTALLATION INSTRUCTIONS
Ensure that Cuda 10.0 and CuDNN 7.4 are installed to use GPU capabilities.
Ensure Anaconda 4.7 or above is installed using conda info, else refer to the Anaconda documentation
The following commands can then be used to install the dependencies:
conda create --name pi3d_env tensorflow-gpu==1.13.1 keras scikit-image opencvDATA
RGB frames for each dataset are in data/[DATASET]/[PART] folder. Skeleton files are present in data/[DATASET]/skeleton/ folder.
The model weights while training will be stored in weights/ folder according to the model name and the corresponding logs can be found in the logs/ folder.
TRAINING P-I3D
Training is done in multiple stages:
- Pretraining of I3D models (for full_body/left_hand/right_hand) can be done through the given scripts
conda activate pi3d_env
sh i3d_train.sh [DATASET] [PROTOCOL] [PART] [NUM_CLASSES] [BATCH_SIZE] [EPOCHS]-
The three layered stacked LSTM network is pretrained by following the following code.
-
The network is jointly trained using
conda activate pi3d_env
python lstm_train_attention.py (optional arguments)The testing script can be found in test.py and requires the model weights of the best epoch to be passed as an argument.