How to quickly reproduce your Computer Vision Models

Nowadays, Machine Learning is extensively used to solve complex Computer Vision problems like Image Classification, Object Detection, Object Segmentation and so on, achieving state-of-the-art results during the latest years. Typically, an image-based training workflow comprises two main processes: dataset creation and training.

Supposing a data scientist team is building a face recognition model, a dataset creation process will accept a set of input files and produce a set of output files. For example, one might take selfies as input and extract tightly-cropped faces from them. The training processes have no input and accept outputs from dataset creation processes. For example, we might train several models from the tightly-cropped selfie dataset. The number of images the team needs to perform both operations properly could be massive and varying over time following the requirements of the experiments. Making all the experiments reproducible introduces some technical challenges to tackle.

We need to design a reproducible workflow that allows the data scientist to:

• Create new datasets and explore an index of existing datasets.

• Explore data.

• Track the data a model was trained with.

• Quickly create new datasets from both original data or other datasets.

• Explore the contents of a dataset and remove specific items from it.

Inspired by this well written tutorial about Data Versioning With Python and DVC, we will see how to deal with all of them using different tools and reducing costs by not duplicating original media or other large data needlessly.