Adaptive quantization and mixed precision in a network
A method of adaptive quantization for a convolutional neural network, includes at least one of receiving an acceptable model accuracy, determining a float value multiply accumulate for the layer based on a float value weight and a float value input, quantizing the float value weight at multiple weight quantization precisions, quantizing the float value input at multiple input quantization precisions, determining a multiply accumulate at multiple multiply accumulate quantization precisions based on the weight quantization precisions and the input quantization precisions, determining multiple quantization errors based on differences between the float value multiply accumulate and the multiple multiply accumulate quantization precisions and selecting one of the multiple weight quantization precisions, one of the multiple input quantization precisions and one of the multiple multiply accumulate quantization precisions based on the predetermined acceptable model accuracy and the multiple quantization errors.
1. A method of adaptive quantization for a convolutional neural network, comprising:
receiving a predetermined acceptable model accuracy for a layer in the network;
determining a float value multiply accumulate for the layer based on a float value weight and a float value input;
quantizing the float value weight at a plurality of weight quantization precisions;
quantizing the float value input at a plurality of input quantization precisions;
determining a multiply accumulate at a plurality of multiply accumulate quantization precisions based on the plurality of weight quantization precisions and the plurality of input quantization precisions;
determining a plurality of quantization errors based on differences between the float value multiply accumulate and the plurality of multiply accumulate quantization precisions; and
selecting the one of the plurality of weight quantization precisions, one of the plurality of input quantization precisions and one of the plurality of multiply accumulate quantization precisions based on the predetermined acceptable model accuracy for the layer and the plurality of quantization errors.
2. The method of adaptive quantization of claim 1 , further comprising determining a weight quantization sensitivity based on the plurality of weight quantization precisions and the float value weight.
3. The method of adaptive quantization of claim 2 , further comprising determining an input quantization sensitivity based on the plurality of input quantization precisions and the float value input.
4. The method of adaptive quantization of claim 3 , further comprising determining a multiply accumulate quantization sensitivity based on the plurality of multiply accumulate quantization precisions and the float value multiply accumulate.
5. The method of adaptive quantization of claim 4 , further comprising determining an overall quantization sensitivity based on the weight quantization sensitivity, the input quantization sensitivity and the multiply accumulate quantization sensitivity.
6. The method of adaptive quantization of claim 5 , further comprising minimizing the quantization sensitivity based on the predetermined acceptable model accuracy.
7. The method of adaptive quantization of claim 1 , further comprising iterating the layer in the network.
8. The method of adaptive quantization of claim 1 , further comprising:
inputting the float value weight for the layer; and
inputting the float value input for the layer.
9. The method of adaptive quantization of claim 8 , wherein the float weight is signed.
10. The method of adaptive quantization of claim 8 , wherein float input is unsigned.
11. The method of adaptive quantization of claim 1 , further comprising selecting the number of quantization bits for the layer.
12. The method of adaptive quantization of claim 11 , wherein the number of quantization bits for the layer is determined based on its influence to an overall accuracy for the network.
13. The method of adaptive quantization of claim 11 , wherein a lower number of quantization bits for the layer is selected if the layer is less sensitive to quantization.