This is a demonstration of video classification performed using Inception 3D module.
Download the data from 20BN-Jester V1 dataset. Follow the instructions in their website and extract them. Place the resulting video folder in Data folder if you would like to run on default parameter settings.
Inception 3D module has a necessity of fixed size input of 224 x 224. Hence you will have to resize the downloaded videos into that fixed size. Also, we will be fixing the number of frames in each video which is currently set at 36. In Preprocess.py, set the parameter of SRC_PATH to the folder path where the downloaded videos are present and DST_PATH to the folder path where the resized videos should be stored in.
Run the following code to resize the videos.
python3 Preprocess.py
After resizing, set the parameters in Main.py which are explained next.
EPOCHS: Number of epochs you would like to run the code
BATCH_SIZE: Batch size during training
LEARNING_RATE: Starting Learning rate
DIVIDE_LEARNING_RATE_AT: At which epochs, learning rate should be divided by 10. Epoch count starts from 0.
DATA_PATH: Folder path of the resized videos folder
TRAIN_LABELS: File path of the downloaded train file
VAL_LABELS: File path of the downloaded validation file
TEST_LABELS: File path of the downloaded test file
LABEL_TEXT: File path of the downloaded labels file
SEQ_LEN: Number of frames in each video clip
IMG_WIDTH: Width of video
IMG_HEIGHT: Height of video
REQD_LABELS: The set of labels you wish to train on.
To start the training, run the following code:
python3 Main.py
After the training, to test the model directly using your camera, configure the parameters in Infer.py which is explained below:
SEQ_LEN: Number of frames in each video clip
IMG_WIDTH: Width of video
IMG_HEIGHT: Height of video
LABELS: List all the labels you have trained your model on. Do not change the order from that used during training
CHECKPOINT_PATH: Folder path where your saved_model is present
To start the real-time testing application, run the following code
python3 Infer.py
This is the sample output result of performing inference on the trained module.
