IVS-Caffe

IVS-Caffe simulates bit-accurate inference of convolutional neural networks and enables training under user-defined quantization to evaluate and improve CNN hardware accelerator accuracy.

Key Features:

• Hardware-oriented bit-accurate simulation: Simulates CNN inference with bit-accurate emulation of arithmetic units to reflect hardware accelerator behavior.
• Quantization of weights, inputs, and outputs: Applies fixed-point quantization to weights, layer inputs, and outputs for bit-accurate evaluation.
• Emulation of multipliers and accumulators: Precisely emulates multipliers and accumulators to produce bit-accurate results during inference calculations.
• Gradient backpropagation correction algorithm: Implements an algorithm that mitigates deviations encountered during gradient backpropagation within bit-accurate quantized multipliers and accumulators.
• Training with user-defined bit widths: Supports retraining of models under user-specified bit-width constraints for multipliers and accumulators.
• Evaluation on detection models and architectures: Evaluates one-stage and two-stage object detection frameworks such as Faster R-CNN + ZF-Net, SSD + VGG, SSD + MobileNet, and Tiny YOLO v2, and architectures including inception modules and depthwise separable convolutions.
• Quantitative mAP assessment under bit constraints: Reports measured effects on mean average precision (mAP), including ≈2% mAP drop with accumulators ≤14 bits and multipliers ≤12 bits and mAP loss reduced below 1% after retraining even with 11-bit multipliers.

Scientific Applications:

• Hardware accelerator validation: Validates accuracy of CNN hardware accelerators by comparing bit-accurate simulated inference to intended behavior.
• Quantization-aware model training: Enables training strategies to retain model accuracy under stringent fixed-point quantization constraints.
• Hardware design trade-off analysis: Analyzes model performance across accelerator bit widths to inform hardware customization for reduced power or area with minimal accuracy loss.
• Benchmarking object detection frameworks: Provides empirical evaluation of object detection frameworks (Faster R-CNN + ZF-Net, SSD + VGG, SSD + MobileNet, Tiny YOLO v2) under quantization.

Methodology:

Simulates inference with fixed-point quantization of weights, inputs, and outputs; emulates multipliers and accumulators bit-accurately; applies a correction algorithm for gradient backpropagation deviations in quantized arithmetic; and retrains models with user-defined bit widths.