New·8 min read
Small Vision Transformers: Fine-Tune or Train from Scratch on 1,020 Images
ViT-B/16 on Flowers-102 with 10 images per class: 39.0% from scratch, 98.4% with a linear probe and 99.2% with LoRA, plus a check for leaked test images.
Lead AI Engineer · Writing since 2016
Hi, I'm Sina. I build and train AI systems.
I write about reinforcement learning, transformers, multi-agent systems, and the tools I build along the way.
Follow along
Latest writing
Full archive →
·6 min·Small Vision Transformers 2/3
Small Vision Transformers: What the Attention Heads Look At
A small ViT trained on CIFAR-10 learns local attention in layer 1 and global attention in the last layers. The fixes that raise accuracy make it more local.
·10 min·Small Vision Transformers 1/3
Small Vision Transformers: A ViT from Scratch on CIFAR-10, with the Failures
A small Vision Transformer starts at 57% on CIFAR-10 and reaches 92.9% after eight fixes. Each fix has a measured gain, and ResNet-18 still wins.
·9 min·Open Source
Sabattier: A Batch Photo Editor with On-Device AI
Sabattier is a free batch photo editor for Windows and macOS with auto enhance, masks, portrait retouching and background removal, all on your computer.
·8 min·Open Source
Mullion FileMan: A Dual-Pane File Manager and a Set of Folder Tools
Mullion FileMan is a C++17 dual-pane terminal file manager. It has twenty folder_* CLI tools that compare and sync folders, and an MCP server for AI agents.
·6 min·Open Source
Redthread: Portable, Git-Backed Memory for AI Coding Agents
Redthread keeps your AI coding agent's memory in a git repo and serves it over MCP, so Claude Code and other agents remember on every machine.
·3 min·Deep RL with Unity 3/3
DDPG on Reacher: Twenty Arms Chasing a Balloon
DDPG solves Unity's Reacher in 101 episodes with twenty arms sharing one replay buffer, and why the noise that failed on Tennis worked here.
Read in order
Series
Multi-part guides that build on each other, start to finish.
Small Vision Transformers
3 parts- 01A ViT from Scratch on CIFAR-10, with the Failures
- 02What the Attention Heads Look At
- 03Fine-Tune or Train from Scratch on 1,020 Images
Deep RL with Unity
3 parts- 01Training a DQN Agent to Collect Bananas in Unity
- 02Two DDPG Agents Learning to Play Tennis
- 03DDPG on Reacher: Twenty Arms Chasing a Balloon
Learn SSH
3 partsLAPACK on macOS
2 partsMS SQL on macOS
2 partsFrom the archive
All 19 posts by year →- Two DDPG Agents Learning to Play Tennis3 min
- Training a DQN Agent to Collect Bananas in Unity3 min
- Learn SSH: Config File2 min
- Learn SSH: CLI Commands2 min
- Install the Latest Python on CentOS, Rocky Linux and RHEL4 min
- Learn SSH: Introduction to SSH on Unix-Based Systems2 min
- How to Install LAPACK on macOS1 min
- Write a Solver for Systems of Equations Using LAPACK on macOS4 min
New posts, no noise.
Subscribe to the RSS feed, or follow along on GitHub, LinkedIn and the RL Factory YouTube channel.