Hyperface Architecture and train pipeline for Hyperface paper in Keras with Tensorflow backend For now only AlexNet based architecture is implemented.
-
Change the sample data path in
config.py
. The data format should be numpy serializable and have the following format For example, the sample data will be a numpy file. It will be a shape of ( X, 6 ) where X is the number of data points. The content of each element will be- Image (numpy array)
- List which has [ 1, 1 ] if the given image is face or [ 0, 0 ] if the given image is non-face
- List of normalized facial coordinates [ 0.2536, 0.7890, …….. ]
- Visibility array [ 0, 0, 0, 1, 1, 1 …... ]
- Pose array [ 0.24304312, -0.42484489, -0.04113968 ]
- Gender array- [ 1, 0 ] for male and [ 0, 1 ] for female
-
Change other parameters for training in
config.py
. -
Start the training with
$ python train.py
- Python 2.7 or Python3.7 (tested on both)
- tensorflow
- keras
- numpy
- Opencv (for image resizing)