LEDNet

This is an unofficial implement of LEDNet.

the official version：LEDNet-official

<div align="center"> <img src="png/demo.png" width="250px"><img src="png/gt.png" width="250px"><img src="png/output.png" width="250px"> </div>

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:

1. use larger learning rate (like 0.01)
2. use more epochs (like 1000)
3. 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]

Prepare data

Your can reference gluon-cv-cityspaces to prepare the dataset