Awesome

Classification of flower images starting from a small dataset

Installation

Python3 is required alogn with cudnna and CUDA Toolkit.

Use the package manager pip to install the required dependencies.

pip install -r requirements.txt

Download the dataset from https://www.robots.ox.ac.uk/~vgg/data/flowers/102/

Create a folder called data and move the downloaded archive. Then execute

matlab -nodisplay -nosplash -nodesktop -r "run('split_dataset_paper.m');exit;"

in the same folder.

Usage

Training

trainer.py [-h] [--batch [BATCH]] [--arch [ARCHITECTURE]] [--opt [OPTIMIZER]] [--clr [CLR]] [--step [STEP]] [--dropout [DROPOUT]] [--config [CONFIG]] [--mp] [--da] [--epoch [EPOCH]]

optional arguments:
  -h, --help            show this help message and exit
  --batch [BATCH]       Batch size used during training
  --arch [{efficientnetb4,frozenefficientnetb4,inceptionv3,resnet18}]
                        Architecture
  --opt [{Adam,SGD}]    Optimizer
  --clr [{triangular,triangular2,exp}]
                        Cyclical learning rate
  --step [STEP]         Step size
  --dropout [DROPOUT]   Dropout rate (when used with FrozenEfficientNetB4 it's used for the freeze rate)
  --config [CONFIG]     Configuration file
  --mp                  Enable mixed precision operations (16bit-32bit)
  --da                  Enable Data Augmentation
  --epoch [EPOCH]       Set the number of epochs

The script will produce plots and checkpoints in ./output/plots and ./output/checkpoints

Learning Rate Finder

python learningratefinder.py [-h] [--batch [BATCH]] [--arch [ARCHITECTURE]] [--opt [OPTIMIZER]] [--dropout [DROPOUT]] [--config [CONFIG]] [--da] [--freeze [FREEZE]] [--epoch [EPOCH]]

optional arguments:
  -h, --help            show this help message and exit
  --batch [BATCH]       Batch size used during training
  --arch [{efficientnetb4,frozenefficientnetb4,inceptionv3,resnet18}]
                        Architecture
  --opt [{Adam,SGD}]    Optimizer
  --dropout [DROPOUT]   Dropout rate
  --config [CONFIG]     Configuration file
  --da                  Enable Data Augmentation
  --freeze [FREEZE]     Frozen layers
  --epoch [EPOCH]       Set the number of epochs

Visualization

To visualize Saliency Map and Grad-CAM run

python3 visualize.py

The image must be store with the following pattern ./images/{class}/{image}.jpg

Contributing

The repository is hosted in Github https://github.com/firaja/aml

License

MIT