Skip to content

Latest commit

 

History

History
62 lines (41 loc) · 1.86 KB

GETTING_STARTED.md

File metadata and controls

62 lines (41 loc) · 1.86 KB
Raw

Pretrained Model

We only provide pretrained models for 19 classes.

Target Cityscapes IDD Mapillary Model
Source only 37.2 36.1 37.9 model
Warm up (1) 44.3 40.5 41.9 model
Warm up (2) 46.3 43.9 47.6 model
BARS 53.5 49.8 52.8 model

Training

Set the hyperparameters in each script. The main ones are as follows: '--ckpt', '--date', and '--exp' represent the checkpoint folder, date, and experiment name, respectively. Folders for saving models and logs will be created using these parameters. '--tb_path' is the path where TensorBoard files will be saved. '--snapshot' is the path to the segmentation model to be loaded. '--DT_snapshot' is the path to the MTDT-Net to be loaded. For other parameters, refer to args.py.

If you do not have more than 32G of memory, you can reduce '--crop_size' to 512 or similar. However, high performance cannot be expected in this case.

1. Train source-only model

sh scripts/source_only_19.sh

2. Train MTDT-Net

sh scripts/train_MTDTNet_19.sh

3. Warm up (1) with MTDT-Net

sh scripts/da_MTDT_19.sh

4. Warm up (2) with AdaptSeg

sh scripts/da_AdaptSeg_MTDT_19.sh

5. Domain adaptation with BARS

sh scripts/da_BARS_19.sh

Test

Input the model to be evaluated in the script using the '--snapshot' option.

sh scripts/test_19.sh