Build a dataset from scratch

This code example shows how you can use this library to prepare a dataset of images for training. You can perform data augmentation on your images and transform the dataset into tfrecords to be used to train a classification model. The data can come from any source and the only requirement is that the images are saved in separate folders for each class. For multilabel classification, however, all images need to be stored in a single folder while training and testing.

Splitting a set of images into a training and a validation set

If you have a single set of images, but would like to split it into a training and a validation set, first place your images in a train folder located in ‘data/image_dataset’ for this example. Using hockey and soccer players as examples of classes, your dataset should be organized in the directory as follows:

data/
  image_dataset/
    train/
      hockey_player/
        hockey_player_1.jpg
        hockey_player_2.jpg
        ...
      soccer_player/
        soccer_player_1.jpg
        soccer_player_2.jpg
        ...

However, for multilabel classification, all the images must be in a single folder and there must exist a JSON file with the keys being the filenames and the values being lists of labels. Your dataset should be organized in the directory as follows:

data/
  image_dataset/
    train/
        sun_aabeeufygtjcsego.jpg
        sun_aabmvnjgiecutfpx.jpg
        sun_aaclfhnpacadfwfc.jpg
        ...

and the json file as code:

{
"sun_aabeeufygtjcsego.jpg": ["grass", "dirt/soil", "natural_light", "man-made", "open_area", "far-away_horizon", "sky", "barn"],
"sun_aabmvnjgiecutfpx.jpg": ["trees", "grass", "asphalt", "natural_light", "man-made", "open_area", "no_horizon", "badminton_court"],
"sun_aaclfhnpacadfwfc.jpg": ["clouds", "natural_light", "man-made", "open_area", "no_horizon", "sky", "buildings"],
...
}

Then, you can run the following code:

decavision.utils.data_utils.split_train(path='data/image_dataset', split=0.2)

This will go through all the classes in the training set, randomly pick a fraction (split argument) of the images, and move them from the train/ directory to a newly created val/ directory.

Make data augmentation on training images

You can increase the training dataset by using data augmentation. You can choose between all these options : distortion, flip_horizontal, flip_vertical, random_crop, random_erasing, rotate, resize, skew, shear, brightness, contrast, color. All the images in the specified folder will have a combination of the desired augmentations. Then you need to specify a desired number of images for all the classes, and it will generate new images to get as close as possible to that number:

augmentor = decavision.dataset_preparation.data_augmentation.DataAugmentor(path='data/image_dataset/train', distortion=True, flip_horizontal=True, flip_vertical=True)
augmentor.generate_images(100)

If you don’t want to generate new images there is an option to do data augmentation during training by modifying the training data online.

Generate Tfrecord files

Once this is done you can generate Tfrecords, which is hardly recommended if you want to use TPU, or just to get better performance from your GPU. It is recommended that you create 100 Mb file using multiple shards. Also, you can save the generated files in a GCS Bucket (gs://) as the output folder (only if you are on colab and have authenticated your account):

generator = decavision_dataset_preparation.generate_tfrecords.TfrecordsGenerator()
generator.convert_image_folder(img_folder='data/image_dataset/train', output_folder='data/tfrecords/train')

You need to use this code once for your train, val (and test folders if you have one). This will delete the old records in the folders and create a csv file with the names of your classes.

The images are not resized by default because different models use different sizes during training. There is an option to specify a target size if desired.