Neural network training environment (including various MLOps tools) designed to compare the explainability of CNNs (using Class Activation Maps) and ViTs (using attention rollout). Project is based on DeepTrainer.
# clone project
git clone https://github.com/lus105/DeepVisionXplain.git
# change directory
cd DeepVisionXplain
# update conda
conda update -n base conda
# create conda environment and install dependencies
conda env create -f environment.yaml -n DeepVisionXplain
# activate conda environment
conda activate DeepVisionXplainTrain model with default configuration (check if environment is properly set up):
# train on CPU (mnist dataset)
python src/train.py trainer=cpu
# train on GPU (mnist dataset)
python src/train.py trainer=gpu
Two cnn models were trained for experimentation.
full size: efficientnet_v2_s. features.7 -> [1, 1280, 7, 7]
downscaled: efficientnet_v2_s. features.6.0.block.0 -> [1, 960, 14, 14]
full size: mobilenet_v3_large. features.16 -> [1, 960, 7, 7]
downscaled: mobilenet_v3_large. features.13.block.0 -> [1, 672, 14, 14]Train cnn/vit model:
# train cnn
python src/train.py experiment=train_cnn
# train vit
python src/train.py experiment=train_vitTrain cnn/vit model with hparams search:
# train cnn
python src/train.py hparams_search=cnn_optuna experiment=train_cnn
# train vit
python src/train.py hparams_search=vit_optuna experiment=train_vitRun explainability segmentation evaluation for all models:
scripts\eval_segmentation.bat- Research paper
- Defect detection datasets.
- Experiment logs.
- Trained models.