May 2021 Summaries

This article provides a comprehensive guide on how to efficiently start a multiclass classification project using Activeloop Hub's dataset management tool for deep learning applications, with a focus on computer vision. It demonstrates the process of preparing and manipulating datasets, including creating subsets, version controlling datasets, and training a simple CNN model. The example used is the Kaggle Simpsons Characters Dataset, which contains jpg images of every character directly taken and labeled from TV show episodes.

The text explores the challenges of using less popular datasets for machine and deep learning tasks and introduces Activeloop's Hub as a solution. Hub allows for efficient storage and access to datasets as cloud-native multidimensional arrays, simplifying the data wrangling process and enabling seamless operations across different machines. The tutorial showcases how to use Hub to manage the "Hot Dog - Not Hot Dog" dataset, employing the platform's features like version control and simplified dataset loading. It further demonstrates the process of building a binary image classifier using ResNet18, emphasizing data preprocessing and transfer learning. The tutorial highlights the ease of converting datasets into PyTorch-compatible formats and training a model with minimal code compared to traditional methods. Overall, the text underscores the benefits of Data 2.0, where data management is streamlined, allowing machine learning engineers to focus more on model training and less on data handling.

Activeloop Hub is a new dataset format for AI that allows storing any type of dataset as cloud-native NumPy-like arrays, making it easier and faster to access big datasets like Google Objectron. Google Objectron is a large dataset containing nearly four million annotated images of various objects commonly found in modern settings. With Hub, users can now access the data in just 28.6 seconds or less per category. The dataset has been divided into nine separate categories and includes approximately 150,000 - 580,000 video frames that can be used for training machine learning models.