I’m Nikita Melkozerov. I work as a software engineer and write about software development and machine learning.

Image Classification from scratch without a PhD

Hello folks! Recently I trained a handwritten digit classifier using the MNIST dataset from scratch, and it was an eye-opening experience for me. What looked like magic to me before now is just a few mathematical concepts applied together. I was very excited about how simple the whole process was, and I want to share it with everyone who still wonders what kind of magic is happening inside. All you need to know is a bit of python and a few concepts from high-school math....

February 24, 2021 · 26 min

Transfer Learning and ResNet: In search of a perfect batch size

TL;DR: batch size 32 is probably going to be a good default candidate for many cases. In this post, we will observe how different batch sizes change learning metrics when we train a model using Transfer Learning and the fast.ai library. We will try to find out which batch sizes are good and which are better to be avoided. ...

January 6, 2021 · 5 min

Building a FoundationDB Cluster: Roles, Classes, and Processes

How to prepare your FoundationDB cluster to run it in production and set up process classes to achieve maximum performance. ...

June 15, 2019 · 8 min