Implementing Darknet19 from scratch

amqdn · March 3, 2019, 12:23am

Hi, all –

Here’s a gist of my implementation of Darknet19 that classifies the MNIST dataset. I modeled the experiment off of course-v3’s Lesson 7 notebook because I found a need to better understand pytorch, fast.ai, and the underlying theory. I ran a whole bunch of experiments and left only the most relevant ones in the notebook to keep the information straight-forward.

You’ll see that my implementation achieves 99.6% accuracy on MNIST in 12 epochs, clocking in at under 2mins training time on an NVIDIA Tesla P100.

Let me know what you think!

gist.github.com

https://gist.github.com/amqdn/cf9106e73d2999f593febf19946b3061

fastai-darknet19-mnist.ipynb

{
  "cells": [
    {
      "metadata": {},
      "cell_type": "markdown",
      "source": "## MNIST CNN"
    },
    {
      "metadata": {},
      "cell_type": "markdown",

This file has been truncated. show original

julclu · March 5, 2019, 10:03pm

Hey @amqdn - just wanted to say thanks for this, this helped me a lot with my own work! I didn’t really understand how to create a model on my own until this.

amqdn · March 5, 2019, 10:23pm

Awesome!