Level 11: Data Generation#

Pre-training of deep learning models for vision tasks can increase model accuracy. Training model with the synthetic dataset is one of famous pre-training approach since the manual annotations is quite expensive work.

Base on the [FractalDB], Datumaro provides a fractal image dataset (FractalDB) generator that can be utilized to pre-train the vision models. Learning visual features of FractalDB is known to increase the performance of Vision Transformer (ViT) models. Note that a fractal patterns in FractalDB is calculated mathmatically using the interated function system (IFS) with random parameters. We thus don’t need to concern about any privacy issues.

We can generate the synthetic images by the following CLI command:

datum generate -o <path/to/data> --count GEN_IMG_COUNT --shape GEN_IMG_SHAPE

GEN_IMG_COUNT is an integer that indicates the number of images to be generated. (e.g. --count 300) GEN_IMG_SHAPE is the shape (width height) of generated images (e.g. --shape 240 180)

With Pthon API, we can generate the synthetic images as below.

from datumaro.plugins.synthetic_data import FractalImageGenerator

FractalImageGenerator(output_dir=<path/to/data>, count=GEN_IMG_COUNT, shape=GEN_IMG_SHAPE).generate_dataset()

GEN_IMG_COUNT is an integer that indicates the number of images to be generated. (e.g. count=300) GEN_IMG_SHAPE is a tuple representing the shape of generated images as (width, height) (e.g. shape=(240, 180))

Congratulations! You completed reading all Datumaro level-up documents for the intermediate skills.


Nakashima, Kodai, et al. “Can vision transformers learn without natural images?.” Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 36. No. 2. 2022.