Generative Zero-shot Network Quantization
arXiv:2101.08430
Abstract
Convolutional neural networks are able to learn realistic image priors from numerous training samples in low-level image generation and restoration. We show that, for high-level image recognition tasks, we can further reconstruct "realistic" images of each category by leveraging intrinsic Batch Normalization (BN) statistics without any training data. Inspired by the popular VAE/GAN methods, we regard the zero-shot optimization process of synthetic images as generative modeling to match the distribution of BN statistics. The generated images serve as a calibration set for the following zero-shot network quantizations. Our method meets the needs for quantizing models based on sensitive information, \textit{e.g.,} due to privacy concerns, no data is available. Extensive experiments on benchmark datasets show that, with the help of generated data, our approach consistently outperforms existing data-free quantization methods.
Technical report
References in corpus (26)
- Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift
- Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks
- PyTorch: An Imperative Style, High-Performance Deep Learning Library
- Distilling the Knowledge in a Neural Network
- Deep Learning with Differential Privacy
- Large Scale GAN Training for High Fidelity Natural Image Synthesis
- Self-Attention Generative Adversarial Networks
- DoReFa-Net: Training Low Bitwidth Convolutional Neural Networks with Low Bitwidth Gradients
- Quantized Neural Networks: Training Neural Networks with Low Precision Weights and Activations
- Tutorial on Variational Autoencoders
- DARTS: Differentiable Architecture Search
- The Concrete Distribution: A Continuous Relaxation of Discrete Random Variables
- Quantizing deep convolutional networks for efficient inference: A whitepaper
- Ternary Weight Networks
- PACT: Parameterized Clipping Activation for Quantized Neural Networks
- Data-Free Knowledge Distillation for Deep Neural Networks
- Resiliency of Deep Neural Networks under Quantization
- Zero-Shot Knowledge Distillation in Deep Networks
- Training Binary Neural Networks with Real-to-Binary Convolutions
- Generative Multi-Adversarial Networks
- Batch-Shaping for Learning Conditional Channel Gated Networks
- Characterising Bias in Compressed Models
- Probabilistic Binary Neural Networks
- BinaryDuo: Reducing Gradient Mismatch in Binary Activation Network by Coupling Binary Activations
- Dream Distillation: A Data-Independent Model Compression Framework
- Fighting Quantization Bias With Bias