DeepSchool.ai

Welcome to my website! I’m Dr. Sachin Abeywardana, an experienced Deep Learning Engineer specialized in NLP and Computer Vision. I offer my expertise as a consultant, having successfully delivered transformative solutions across various industries. Notably, as Lead ML Engineer at Canva, I innovated customer service using NLP, employing BERT for feedback classification and fine-tuning GPT2 for content titles. At Remi AI, I predicted demand and optimized prices using advanced DL models, whilst also mentoring other data scientists. In my role at Data Processors, I improved sports betting models and video analysis using Computer Vision. Holding a Ph.D. in Bayesian ML, my research includes quantile regression and market segmentation. I’m also a Udemy Instructor, contributing to the ML community. As a consultant, I aim to drive innovation and collaboration, offering custom AI solutions to push boundaries and achieve exceptional results. Feel free to get in touch with me on LinkedIn.

Featured Posts

DINO Self Supervised Vision Transformers

pytorch

Loss Function

Implementing DINO self-supervised learning for Vision Transformers in PyTorch to extract image embeddings without labeled data or negative samples

Sachin Abeywardana

Multilingual CLIP with Huggingface + PyTorch Lightning 🤗 ⚡

pytorch

Loss Function

Training a multilingual version of OpenAI’s CLIP model using HuggingFace and PyTorch Lightning on Google Colab

Sachin Abeywardana

All Posts

Stop Prompting, Start Building: A Real Guide to Coding Agents

AI

Agents

There is a lot of discussion about agents that sounds either magical or completely useless. My experience has been much less dramatic. They are not replacing engineering…

Calculating Loss on a Transformer

Deep Learning

LLM

Transformers

PyTorch

A deep dive into how transformer models calculate loss, from first principles to the HuggingFace implementation

Sachin Abeywardana

CLIP - JEPA: Converting LLMs to extract Embeddings

Deep Learning

Loss Functions

Notes + results from training a CLIP-style JEPA using Qwen3 (text) + MobileNetV4 (vision) with CyCLIP + SigLIP losses

Sachin Abeywardana

Image Segmentation with Deep Learning

Deep Learning

Loss Functions

A practical guide to image segmentation using UNet, covering data augmentation strategies and loss functions for training segmentation models in PyTorch

Sachin Abeywardana

Massively Parallel Experimentation: Ship Better Models, Faster 🚀

MLOps

This guide dives into setting up large-scale machine learning experiments with a clean code structure, local testing, and efficient parallelization using tools like Typer, Pydantic, and Argo Workflows. Learn how to move beyond notebooks, structure ML projects for scalability, and log results effectively to accelerate model development.

Sachin Abeywardana

Transformers Inference Optimizations ⏰🚀

LLM

pytorch

Optimization tricks used for speeding up transformer inference

Sachin Abeywardana

LLM Finetuning: Demystifying Huggingface Trainer 🚀

LLM

Tutorial on finetuning LLMs via HF transformers library with wandb logging

Sachin Abeywardana

Supercharging Structured Outputs with Open Source Models 🚀

LLM

This study finds NuExtract performs best for structured outputs, with KV caching improving speed and accuracy for larger models despite some hallucinations.

Sachin Abeywardana

Vison Language Models from Scratch

Deep Learning

LLM

huggingface

Multimodal

Building a vision-language model from scratch by combining small language models with lightweight vision encoders to generate image captions

Sachin Abeywardana

BuildKite + ArgoWF for ML Training Jobs

MLOps

MLOps: Leveraging ArgoWF and Buildkite to train models

Sachin Abeywardana

Prompt Caching: Poor man’s guide to zero shot vision-LLM classification

Deep Learning

LLM

huggingface

Using KV caching and logit ratios to speed up and control LLM/ VLM outputs.

Sachin Abeywardana

From DDPM to DDIM: A Mathematcal Deep Dive

Deep Learning

Diffusion Models

A mathematical deep dive in excruciating detail into the DDIM model

Why I am voting YES!

politics

Why voting YES in this referendum is the right thing to do.

Sachin Abeywardana

Dreambooth Tutorial

LLM

Multimodal

Diffusion Models

Dreambooth is a cool technique that lets you convert existing diffusion models to output personalized images. It’s one of the two main methods for fine-tuning diffusion…

Creating a Caption Model from Scratch

LLM

Multimodal

Building an image captioning model from scratch using ConvNext and Flan-T5 with cross-attention projection layers in PyTorch

Thoughts on CosineEmbeddingLoss

pytorch

How PyTorch’s CosineEmbeddingLoss works to push similar embeddings closer and dissimilar ones apart using cosine distance and a margin threshold

Annotated DDPM

Deep Learning

Diffusion Models

An annotated implementation of Denoising Diffusion Probabilistic Models (DDPM) with step-by-step code for training on MNIST in PyTorch

Sachin Abeywardana

Pydantic Post Init

python

Workaround for Pydantic v1 post-init logic using properties and custom init overrides

Timm choosing a model

pytorch

How to choose a small, high-accuracy timm backbone for transfer learning by filtering ImageNet results by accuracy and parameter count

Pytorch Lightning Learning rate finder

pytorch

Using PyTorch Lightning’s built-in learning rate finder to plot loss vs learning rate and pick an optimal value

Fine Tuning T5 for Grammar Correction

Deep Learning

LLM

Fine-tuning Google’s T5 model for grammar correction as a sequence-to-sequence task using HuggingFace Transformers

Sachin Abeywardana

Fine Tuning GPT2 for Grammar Correction

Deep Learning

LLM

Fine-tuning GPT-2 for grammar correction as a sequence-to-sequence task using HuggingFace Transformers

Sachin Abeywardana

Transformer Model Compression (Attempt)

Deep Learning

A failed attempt at model compression using student teacher learning

Sachin Abeywardana

Unit Testing for Data Science

python

Software Engineering

A practical guide to writing unit tests for machine learning pipelines in Python using pytest

Sachin Abeywardana

Vector Database from Scratch

pytorch

Implementing Approximate Nearest Neighbours Oh Yeah (ANNOY)

Sachin Abeywardana

KMeans in PyTorch with Cosine Distance🥧🔦

pytorch

Implementing kmeans with cosine distance

Sachin Abeywardana

PyTorch prefetch or rather the lack of it

pytorch

How prefetch_factor did not help in streaming data

Sachin Abeywardana

Generating captions with ViT and GPT2 using 🤗 Transformers - Part 2

pytorch

huggingface

Using Encoder Decoder models in HF to combine vision and text

Sachin Abeywardana

Generating captions with ViT and GPT2 using 🤗 Transformers

pytorch

huggingface

Using Encoder Decoder models in HF to combine vision and text

Sachin Abeywardana

HuggingFace Tokenizers as Collate Functions Timing 🤗 🤖

pytorch

huggingface

Timing comparison of tokenizer as collate function and after batching

Sachin Abeywardana

Zero Shot Classification with Huggingface + Sentence Transformers 🤗 🤖

pytorch

huggingface

How to perform fast zero-shot text classification using HuggingFace Sentence Transformers and cosine similarity without fine-tuning

Sachin Abeywardana

An Intuitive Loss for Imbalanced Classification

A simple weighted cross-entropy loss function as an intuitive alternative to focal loss for handling imbalanced classification

Coco Semantic Segmentation in PyTorch - Data Prep

pytorch

data

How to prepare and transform image data for segmentation

Sachin Abeywardana

DINO Self Supervised Vision Transformers

pytorch

Loss Function

Implementing DINO self-supervised learning for Vision Transformers in PyTorch to extract image embeddings without labeled data or negative samples

Sachin Abeywardana

PyTorch Image Patches

pytorch

How to split images into patches for Vision Transformers (ViT) using PyTorch unfold and einops rearrange

Collate function tutorial

pytorch

A practical guide to writing custom PyTorch collate functions for DataLoader to handle variable-length sequences and complex batching

The Annotated TabNet

Tabular Data

Deep Learning

Creating TabNet from Scratch in Tensorflow 2.0

Sachin Abeywardana

Multilingual CLIP with Huggingface + PyTorch Lightning 🤗 ⚡

pytorch

Loss Function

Training a multilingual version of OpenAI’s CLIP model using HuggingFace and PyTorch Lightning on Google Colab

Sachin Abeywardana

Tensorflow Learning Rate Finder

Deep Learning

Implementing a learning rate finder in TensorFlow using custom Keras callbacks to find the optimal learning rate for training

Focal Loss for Multi-class Classification

Loss Function

Extending focal loss to multi-class classification problems to handle class imbalance in deep learning models

Docker for Data Science

Docker

Docker is a tool that simplifies the installation process for software engineers. Coming from a statistics background I used to care very little about how to install…

DeepSchool.io

NodeSchool is one of the most inclusive software communities that I have come across. What I liked about it the most is its emphasis on writing code. There are so many…

Keras LSTMs

Deep Learning

Keras has been one of the really powerful Deep Learning libraries that allow you to have a Deep Net running in a few lines of codes. Best part, don’t worry about the math.…

Deep Learning Quantile Regression - Keras

Loss Function

The loss function is simple as doing the following. Which is simply the pin-ball loss function.

XgBoost - Machine Learning made EASY!

Machine Learning

One of the machine learning frameworks that has been exploding on the Kaggle scene has been Xgboost. In my personal experience it has been an extremely powerful machine…

Reversible jump MCMC

Bayesian

Unsupervised Learning

Reversible jump MCMC is a Bayesian algorithm to infer the number of components/ clusters from a set of data. For this illustration we shall consider a two component model at…

Chinese Restuarant Process

Bayesian

Unsupervised Learning

In this instance we generate the parameters \[\theta_k\] from \[\mathcal{N}(\mathbf{0},3\mathbf{I})\]. The data is generated from \[\mathcal{N}(\theta_k,0.1\mathbf{I})\].…

von Mises-Fisher Distribution

Statistics

The von Mises Fisher Distribution is a multivariate distribution on a hyper sphere. I have decided to share the expectation and covariance of the vMF distribution. The…

Sample Variance

Statistics

People often question why is there a “n-1” term when I calculate the variance. Why not divide through by “n”. Most stats courses dismiss this question by saying, “oh, that’s…

Normal Distribution

Statistics

No stats blog would be complete without a discussion of the Gaussian distribution. In the following video I discuss how to obtain the mean and variance of a Gaussian. You do…