Subset-Selection Weight Post-Training Quantization Method for Learned Image Compression Task

Abstract

Post-training quantization(PTQ) has been widely studied in recent years because it does not require retraining the network or the entire training dataset. However, naively applying the PTQ method to quantize low-level image tasks such as learned image compression(LIC) usually incurs significant accuracy degradation. Existing solutions often aim to optimize the calibration process or minimize quantization loss in the context of uniform quantization, which makes it difficult to further reduce the quantization loss. This work achieves accurate weight quantization using a non-uniform quantization method called subset-selection within the PTQ method. Subset-selection method uses a clustering algorithm to select a proper batch of quantization points from a large uniformly constructed quantization points pool that matches the current layer’s weight distribution correctly. Compared to the state-of-the-art LIC PTQ method, experimental results show that our proposed method achieves an average 0.8% better BD-rate on several LIC models. Meanwhile, we extended our method to image classification models and achieved an average 0.64% better accuracy, further proving the generalization of our method.

Keywords

• Convolutional neural network
• post-training quantization
• non-uniform quantization
• learned image compression