travisergodic / kaggle-cat-and-dog

Kaggle Cat and Dog Competition

Introduction

Kaggle Dog v.s. Cat competition

Model & Performance

1. Model: ResNet50
2. Train & Validation ratio: 0.85: 0.15
3. Performance
   • accuracy: 99.25%
   • recall: 99.04%
   • precision: 99.46%
   • auc: 99.97%

Resource

1. Environment: Linux
2. Python version: 3.10.12
3. GPU: V100, 16G

Installation

1. Create virtualenv

   python -m venv cat_dog_env
   source cat_dog_env/bin/activate
   

2. Install dependencies

   pip install -r requirements.txt
   

Command Line Interface

1. Train: Train image classifier.

   python tools/train.py --config_file ${config_file} \
                         --exp_name ${exp_name} \
                         --img_dir ${img_dir} 
                         --test_ratio ${test_ratio}  \
                         --bs ${batch_size} \
                         --lr ${lr} \
                         --weight_decay ${weight_decay} \ 
                         --weight ${weight}
                         [--adam] \ 
                         [--adamw]
   

   • config_file: Training configuration file. File contains model confiiguration, training strategy, etc. Configuration files are in configs/ folder.
   • exp_name: The name of current experiemnt. Training aritfacts (model weights, ROC graph, confusion matrix graph, performance report, etc) will be save in checkpoints/{exp_name} folder.
   • img_dir: Folder containing all training images.
   • test_ratio: Validation set ratio.
   • bs: Training batch size.
   • lr: Learning rate.
   • adam: Use Adam optimizer during training.
   • adamw: Use AdamW optimizer during training.
   • weight_decay: Weight_decay regularization.
   • weights: Pretained model weight path.

2. Inference: Predict image categories in the folder.

   python tools/inference.py --config_file ${config_file} \
                             --img_dir ${img_dir} \
                             --bs ${batch_size} \ 
                             --weight ${weight} \
                             --save_path ${save_path}
   

   • config_file: Configuration file for inference.
   • img_dir: Folder containing all images to be predicted.
   • bs: Inference batch size.
   • weight: Pretained model weight path.
   • save_path: Prediction result will save to this file.

3. Evaluate: Evaluate the model performance.

   python tools/evaluate.py --config_file ${config_file} \
                             --csv_path ${csv_path} \ 
                             --img_dir ${img_dir} \
                             --bs ${batch_size} \ 
                             --weight ${weight} \
                             --save_path ${save_path}
   

   • config_file: Configuration file for evaluation.
   • csv_path: csv file containing testing image names and labels. csv_file should contain name(image name) & label(image category) columns. Example
   • img_dir: Folder containing all images to be predicted.
   • bs: Inference batch size.
   • weight: Pretained model weight path.
   • save_dir: Path to save evaluation artifacts, including ROC graph, confusion matrix graph.

Colab Practice

Experiments

1. Resnet50: link
2. Convnext
3. Swin Transformer
4. BEITv2