This is an unofficial implement of LEDNet.
the official version：LEDNet-official
- Python 3.6
- PyTorch 1.1
- Base Size 1024, Crop Size 768, only fine. (old-version, without dropout)
| Model | Paper | OHEM | Epoch | val (crop) | val |
|---|---|---|---|---|---|
| LEDNet | / | ✗ | 240 | 44.67/91.85 | 49.79/91.31 |
| LEDNet | / | ✗ | 1000 | 53.77/93.45 | 59.04/93.27 |
- Base Size 1356, Crop Size 1024, only fine. (old-version, without dropout)
| Model | Paper | OHEM | Epoch | val (crop) | val |
|---|---|---|---|---|---|
| LEDNet | / | ✗ | 1000 | 56.30/93.90 |
The paper only provide the test results: 69.2/86.8 (class mIoU/category mIoU)
- reference the Fast-SCNN, we choose epoch=1000
- Height 1024, Width 512. (new-version)
| Model | Paper | OHEM | Epoch | val (crop) | val |
|---|---|---|---|---|---|
| LEDNet | / | ✗ | 300 | 39.03/88.60 | 21.17/72.79 |
| LEDNet | / | ✗ | 800 | 41.70/89.46 |
TODO
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]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 240] [--warmup-factor 0.1] [--warmup-iters 200] [--log-step 10] [--save-epoch 40] [--lr 0.0001]Your can reference gluon-cv-cityspaces to prepare the dataset