-NNCF provides an alternate mode, namely AutoQ, for mixed-precision automation. It is an AutoML-based technique that automatically learns the layer-wise bitwidth with explored experiences. Based on [HAQ](https://openaccess.thecvf.com/content_CVPR_2019/papers/Wang_HAQ_Hardware-Aware_Automated_Quantization_With_Mixed_Precision_CVPR_2019_paper.pdf), AutoQ utilizes an actor-critic algorithm, Deep Deterministic Policy Gradient (DDPG) for efficient search over the bitwidth space. DDPG is trained in an episodic fashion, converging to a deterministic mixed-precision policy after a number of episodes. An episode is constituted by stepping, the DDPG transitions from quantizer to quantizer sequentially to predict a precision of a layer. Each quantizer essentially denotes a state in RL framework and it is represented by attributes of the associated layers. For example, a quantizer for 2D Convolution is represented by its quantizer Id (integer), input and output channel size, feature map dimension, stride size, if it is depthwise, number of parameters etc. It is recommended to check out ```_get_layer_attr``` in [```quantization_env.py```](https://github.com/openvinotoolkit/nncf/blob/develop/nncf/automl/environment/quantization_env.py#L333) for the featurization of different network layer types.
0 commit comments