Mastering Computer Vision with TensorFlow 2.x

Model optimization

Model optimization, also known as quantization, can be performed with post-training quantization to improve CPU/GPU performance, without sacrificing accuracy. The optimization process can be performed using the following:

• Floating point to 8-bit precision (optimize for size)
• Full integer quantization using integer input and output for a microcontroller
• A bit of both – dynamically quantize with 8 bits but any outputs are stored in floating point form
• Pruning is another dynamic optimization method to eliminate the low value weights from the neural network during training. It can be initiated by following lines of code:

from tensorflow_model_optimization.sparsity import keras as sparsitypruning_params = { 'pruning_schedule': ...