This repository contains code for Fine Tuning DeepLabV3 ResNet101 in PyTorch. The model is from the torchvision module. The tutorial can be found here: https://towardsdatascience.com/transfer-learning-for-segmentation-using-deeplabv3-in-pytorch-f770863d6a42?sk=b403331a7b30c02bff165a93823a5524
I've fine tuned the model for the CrackForest data-set.
The model was fine tuned for 25 epochs and achieves an testing AUROC value of 0.842.
The segmentation output of the model on a sample image are shown below.
pip install -r requirements.txt
conda env create -f environment.yml
Usage: main.py [OPTIONS]
Options:
--data-directory TEXT Specify the data directory. [required]
--exp_directory TEXT Specify the experiment directory. [required]
--epochs INTEGER Specify the number of epochs you want to run the
experiment for. Default is 25.
--batch-size INTEGER Specify the batch size for the dataloader. Default is 4.
--help Show this message and exit.
To run the code with the CrackForest dataset and store the results in folder called CFExp use the following command.
python main.py --data-directory CrackForest --exp_directory CFExp
Inference Script
python inference.py --image_path=SmallObjectDataset/Test_Images/Images/0000000393.png --mask_path=SmallObjectDataset/Test_Images/Masks/0000000393.png --model_path=SmallObjectExp/0-25EpochsNoWeights/checkpoint_0025_DeepLabV3_SmallObject.pt --apply_transforms=True
The datahandler module has two functions for creating datasets fron single and different folders.
-
def get_dataloader_sep_folder(data_dir: str, image_folder: str = 'Image', mask_folder: str = 'Mask', batch_size: int = 4)
Create Train and Test dataloaders from two separate Train and Test folders. The directory structure should be as follows.
data_dir --Train ------Image ---------Image1 ---------ImageN ------Mask ---------Mask1 ---------MaskN --Test ------Image ---------Image1 ---------ImageM ------Mask ---------Mask1 ---------MaskM
-
def get_dataloader_single_folder(data_dir: str, image_folder: str = 'Images', mask_folder: str = 'Masks', fraction: float = 0.2, batch_size: int = 4)
Create from a single folder. The structure should be as follows.
--data_dir ------Image ---------Image1 ---------ImageN ------Mask ---------Mask1 ---------MaskN
The repository also contains a JupyterLab file with the loss and metric plots as well as the sample prediction code.
If you found this repository to be useful and use it in your work, please consider citing it:
Bibtex Entry:
@misc{minhas_2019, title={Transfer Learning for Semantic Segmentation using PyTorch DeepLab v3}, url={https://github.com/msminhas93/DeepLabv3FineTuning}, journal={GitHub.com/msminhas93}, author={Minhas, Manpreet Singh}, year={2019}, month={Sep}}
IEEE Format Citation:
M. S. Minhas, “Transfer Learning for Semantic Segmentation using PyTorch DeepLab v3,” GitHub.com/msminhas93, 12-Sep-2019. [Online]. Available: https://github.com/msminhas93/DeepLabv3FineTuning.