instruction.md

instruction of the code

1. Environment Setup

   • comp0197-cw1-pt base environment
   • cython ~pip install Cython
   • pydensecrf ~pip install pydensecrf
   • cv2 ~pip install opencv-python

2. Data Preparation

   • Download the dataset from the link provided in the assignment
   • Unzip the dataset and put it in the same directory as the code
   • For the background enhancing image, download from: https://www.kaggle.com/datasets/balraj98/stanford-background-dataset
   • The dataset should be in the following format:
     • data/
       • oxford-iiit-pet/
         • annotation/
         • images/
         • annotations.tar.gz
         • images.tar.gz
       • background
         • images/

3. models:

   • The Resnet50 backbone is modified with stride sequence [1,2,2,1] with last 2 convolutions outputs as the features for future cam. details in src.model.ResNet50
   • The pretrained Resnet50 can be found https://download.pytorch.org/models/resnet50-19c8e357.pth

4. Training and evaluation:

   • train backbone classifier: Run train_classifier/train_*.py file for your specific classifier finetuning
     • in the same folder, there is a evaluation_backbone.py file for the base CAM/ECS_CAM evaluation
     • for evaluation, please load the trained model path for evaluation_backbone.py
   • train CCAM: Run train_ccam/train_ccam.py file with specific parameters.
     • in the same folder, there is a evaluation_backbone.py file for the base CCAM/ECS_CCAM evaluation
     • this training may take quite a long time, you can go to https://drive.google.com/drive/folders/10sFREPCyJv_EqRCBarTwA5PabusN4_x1 for the existing model
   • train class specific CCAM:
     • run trani_class_specific_ccam/data_class_split.py first
     • run train_class_specific_ccam/train_ccam_on_cat_dog.py with specific parameters to get cat_dog specific CCAM
     • run train_class_specific_ccam/train_ccam_on_specific_37_class.py with specific parameters to get 37-class specific CCAM
     • pseudomask generating: run train_class_specific_ccam/generating_mask.py with pseudo-mask generating
     • for result evaluation: run train_class_specific_ccam/evaluate_class_specific_CCAM.py
   • Train Unet:
     • for Unet trained on ground truth run: train_unet/train_on_gt.py
     • for Unet trained on 37 specific class CCAM pseudo-mask run: train_unet/train_on_37CCAM.py
     • for Unet trained on 37 specific class CCAM pseudo-mask + CRF run: train_unet/train_on_37CCAM_CRF.py
     • for evaluating the performance run: train_unet/ccam_unet_eval.py