We try to quantize a h5 model for the use with the NXP i.mx8 Plus NPU
The model is the keras-VGGFace model : https://github.com/rcmalli/keras-vggface
FILES:
quantizationforNPU.py loads a saved h5 model and fully quantizes it to int8. The h5 model can be found here : https://mega.nz/file/ZN8nhYxC#N7iJ8LfACH2N34n3tPkz6a5we6Fog4utIgcHOLx53iQ
The 0002.jpeg is a sample image for the representative_dataset_gen(). The model was trained on 224x224 images. Actually the images needs to be normalized before, however to only get it running this should be sufficiant.