APE

APE is A simPle image classification framEwork for quick and easy training/evaluation/deployment.

File structure

Codes are in the ape folder.

train.py is for training, test.py is for testing accuracy using test dataset, predict.py is for predicting the label of given images.

Configuration files are in configs folder. Configs are python files.

Requirements

• python>=3.6
• pytorch>=1.1.0
• tensorboard>=1.14.0 (better with tensorflow installed, but not required)
• joblib>=0.14.1

Quick start

Dataset

Datasets are loaded using torchvision.datasets.ImageFolder. Simply arrange your dataset by categories, in this way:

$DATASET_PATH/dog/xxx.png
$DATASET_PATH/dog/xxy.png
$DATASET_PATH/dog/xxz.png

$DATASET_PATH/cat/123.png
$DATASET_PATH/cat/nsdf3.png
$DATASET_PATH/cat/asd932_.png

And it will be loaded automatically.

Training

To train the model, first define a model config, configs/model.py for example. Then run:

python -u ape/train.py --config /path/to/your/config
# Example:
python -u ape/train.py --config configs/model.py

Note that config files should always be placed within the configs folder due to import mechanisms of python.

You can also resume a training process using a model checkpoint. Set resume = True and set the checkpoint path in the config, and it will be resumed from the checkpoint.

Evaluation

To test the model using test dataset, run:

python -u ape/test.py --config /path/to/your/config
# Example:
python -u ape/test.py --config configs/model.py

Inference

You can use predict.py to inference images using trained model. Run:

python -u ape/predict.py --config /path/to/your/config --path /path/to/your/images --ext <Extension-Name> --output /path/to/output/result
# Example:
python -u ape/predict.py --config configs/model.py --path /data/dataset/ --ext png --output infer_output/