A general pipeline

1. Create a new extension for your project.

bash scripts/add_extension.sh my_extension

This creates a following folder structure:

extensions/my_extension/
├── dataloaders
├── models
├── runners
├── scripts
└── config.yaml

2. Define your dataloaders, models, runners, scripts in corresponding directories.

   • Each file in models dir should implement get_args, get_net, get_native_transform functions.
   • ...

   Use classification.py dataloader/runner as an example and resnet_classification.py as an example of a model.

Classification

Classification works out of the box, runner, dataloader are already implemented there. However, for clarity, you need to create an extension first.

bash scripts/add_extension.sh my_classification

Here is an easy way to solve a classification problem.

1. Prepare your dataset to have this structure:

dataset
├── class1
├── class2
└── class3

Images for each class in a corresponding folder.

2. Run this script to gather images and split into train and validation.

python scripts/train_test_split.py --classes_dir path/to/dataset --save_dir extensions/my_extension/data/splits --test_size 0.

3. Learn something

python train.py --extension my_classification --config_name classification