cifar

Installation Guideline

git clone git@github.com:Anuj040/cifar.git [-b <branch_name>]
cd cifar (Work Directory)

# local environment settings
pyenv local 3.6.9
python -m pip install poetry
poetry config virtualenvs.create true --local
poetry config virtualenvs.in-project true --local

# In case older version of pip throws installation errors
poetry run python -m pip install --upgrade pip 

# local environment preparation
poetry install

Working with the code

All the code executions shall happen from work directory.

Training mode

There are three different training modes for the classifier training, plus one for training of auto-encoder (AE) only

1. For auto-encoder only training, please execute the following

poetry run python cifar/start.py --train_mode=pretrain --epochs=400

2. For classifier's training

• classifier only training

poetry run python cifar/start.py --train_mode=classifier --epochs=400

• AE only training followed by classifier only training

poetry run python cifar/start.py --train_mode=both --epochs=400

• simultaneous AE-classifier training in multitask setting

poetry run python cifar/start.py --train_mode=combined --epochs=400

Saving the model files

1. For AE only or "both" training, at the end of AE training, a model file will be saved in the following directory.

    cifar {work_directory}
    ├── cifar
    ├── ae_model                   
    │   ├──ae_model.h5         # Saved model
    └── ...

2. For classifier only or "both" training, during classifier training, progressively better models will be saved in the following directory.

    cifar {work_directory}
    ├── cifar
    ├── class_model                   
    │   ├──class_model_{epoch:04d}_{val_acc:.4f}.h5         # Saved model # epoch & val_acc = epoch at which model was saved and corresponding validation accuracy
    └── ...

3. For "combined" training, during the training, progressively better models will be saved in the following directory.

    cifar {work_directory}
    ├── cifar
    ├── com_model                   
    │   ├──com_model_{epoch:04d}_{val_acc:.4f}.h5         # Saved model # epoch & val_acc = epoch at which model was saved and corresponding validation accuracy
    └── ...

**Note: By default the current implementation, splits the original training set of cifar-10 into train/val (80:20) split for classifier training.

Evaluation

To run the evaluation on test dataset, please execute the following from the work_directory

poetry run python cifar/start.py --train_mode=combined --model_path=com_model/com_model_0500_0.0000.h5 --mode=eval

Inference

To run inference on a single image eg: test0.png, please execute the following from the work_directory

poetry run python cifar/start.py --train_mode=combined --model_path=com_model/com_model_0500_0.0000.h5 --mode=infer --img_path=./test0.png

** Note: please make sure to use the right combination of train_mode and model_path flags, eg: as shown above or (classifier, class_model/class_model_0500_0.0000.h5).