lednet's Introduction

LEDNet

This is an unofficial implement of LEDNet.

the official version：LEDNet-official

Environment

Python 3.6
PyTorch 1.1

Performance

Base Size 1024, Crop Size 768, only fine. (new-version, with dropout)

Model	Paper	OHEM	Drop-rate	lr	Epoch	val (crop)	val
LEDNet	/	✗	0.1	0.0005	800	60.32/94.51	66.29/94.40
LEDNet	/	✗	0.1	0.005	600	61.29/94.75	66.56/94.72
LEDNet	/	✗	0.3	0.01	800	63.84/94.83	69.09/94.75

Note:

The paper only provide the test results: 69.2/86.8 (class mIoU/category mIoU).

And the training setting is a little different with original paper (original paper use 1024x512)

Some things you can use to improve the performance:

use larger learning rate (like 0.01)
use more epochs (like 1000)
use larger training input size (like Base Size 1344, Crop Size 1024)

Demo

Please download pretrained model first

$ python demo.py [--input-pic png/demo.png] [--pretrained your-root-of-pretrained] [--cuda true]

Evaluation

The default data root is ~/.torch/datasets (You can download dataset and build a soft-link to it)

$ python eval.py [--mode testval] [--pretrained root-of-pretrained-model] [--cuda true]

Training

Recommend to using distributed training.

$ export NGPUS=4
$ python -m torch.distributed.launch --nproc_per_node=$NGPUS train.py [--dataset citys] [--batch-size 8] [--base-size 1024] [--crop-size 768] [--epochs 800] [--warmup-factor 0.1] [--warmup-iters 200] [--log-step 10] [--save-epoch 40] [--lr 0.005]