For Quantization-aware training, we model the effect of quantization using simulated quantization operations, which consist of a quantizer followed by a de-quantizer.
In our implementation of post-training quantization, we utilize the ImageNet-1k dataset for training and evaluation. We set the bitwidth of activation to 8 and allow the bitwidth of weight to vary within the set ${2, 3, 4}$. We randomly sample 425 bit configurations and record their quantization accuracy to build the MQ-Bench-101 (version 1.0). Moreover, we are actively training all bit configurations with a batch size of 64. All of the experiment of MQ-Bench-101 is computed by running each bit configuration on a single GPU (NVIDIA RTX 3090). Our chosen convolutional neural network is ResNet18, and we apply layer-wise quantization for weights. We perform weight rounding calibration with 20,000 iterations, where the temperature during calibration ranges from 20 to 2. We adopt asymmetry quantization and use mean squared error (MSE) as the scale method. The learning rate is set to 4e-4, and we use a calibration data size of 1,024. We would like to acknowledge OMPQ for the implementation of our post-training quantization, which is based on their official repository.
