wanglixilinx / arxiv_daily

Tutorial for 4-bit FPN_ResNet18 for Semantic Segmentation on Cityscapes dataset

Contents

1. Installation
2. Usage

Installation

• Install the torch, torchvision and some other requirements:

pip install torch torchvision tqdm Pillow scipy

Usage

.{CURRET_PATH}
├── Dataset
│   ├── demo_images
│   └── cityscapes
├── encoding
│   ├── models
│   └── datasets
├── train_and_eval
│   ├── test.py
│   ├── demo.py
│   └── color_q4_results

Support

low-bit Files

• [QFPN(ResNet18)] (encoding/models/fpnQ.py)

• [quantize files] (encoding/models/quantize.py & encoding/models/sim_bn_foldto_conv.py)

Weights for QFPN(ResNet18)

• [float] (encoding/weight/float.pth.tar)
• [4bit] (encoding/weight/q4/q6240.pth.tar)

Demo

cd train_and_eval 
sh run_demo_quantize.sh

Evaluation on Cityscapes Validation

cd train_and_eval
sh run_eval_quantize.sh

Input size	A/W bit	mIoU(val)	FLOPs
256 * 512	float	62.9%	10G
256 * 512	4/4	62.0%	10G