kono-dada/4220final

4220 Final Project

The report is report.pdf.

Installation

Step 0. Install MMCV using MIM.

pip install -U openmim
mim install mmengine
mim install "mmcv>=2.0.0"

Step 1. Install MMSegmentation.

pip install -v -e .
# '-v' means verbose, or more output
# '-e' means installing a project in editable mode,
# thus any local modifications made to the code will take effect without reinstallation.

Training

In this project, We mainly conduct four experiments.

1. FCN+U-Net

   python /tools/train.py my_config/fcn_unet.py --work-dir fcn
   

2. Deeplabv3+

   python /tools/train.py my_config/deeplabv3plus.py --work-dir deeplabv3plus
   

3. Deeplabv3+ with pretraining

   python /tools/train.py my_config/deeplabv3plus_pretrain.py --work-dir deeplabv3plus_pretrain
   

4. Deeplabv3+ with pretraining and Lovasz loss

   python /tools/train.py my_config/deeplabv3plus_pretrain_lovasz.py --work-dir deeplabv3plus_pretrain_lovasz
   

Unzip the checkpoint

The best-performanced checkpoint is stored in checkpoints/iter_2000.zip. Since Github has a limited file size of 100mb, the checkpoint is zipped into 4 files.

Testing

Since we only upload the model with the best performance, only the third experiment has the checkpoint.

1. FCN+U-Net

   python /tools/test.py my_config/fcn_unet.py {checkpoint} --work-dir fcn
   

2. Deeplabv3+

   python /tools/train.py my_config/deeplabv3plus.py {checkpoint} --work-dir deeplabv3plus
   

3. Deeplabv3+ with pretraining

   python /tools/train.py my_config/deeplabv3plus_pretrain.py checkpoints/iter_2000.pth --work-dir deeplabv3plus_pretrain
   

4. Deeplabv3+ with pretraining and Lovasz loss

   python /tools/train.py my_config/deeplabv3plus_pretrain_lovasz.py {checkpoint} --work-dir deeplabv3plus_pretrain_lovasz
   

Inferencing

python inference.py

Please see details in inference.py.