picostitch
crafting (and) JavaScript

ML for VRT - Part 3: Hands On

Wolfram Kriesing - August 9, 2020 - tagged with: #learning #machine learning #automation #tidbit

I renamed the series to "ML for VRT" instead of "ML vs. VRT" since that is what I actually want to achieve (eventually), as you can read in Part 1 - Machine Learning vs. Screenshot Comparing?. Read on to figure out how I ended up next to learn about neural networks, because I figured out just applying some code examples learned from tutorials won't suffice my learning and they won't answer the questions I actually have to understand how to tune a neural network to do what I image it to do, detecting screenshots for visual regression testing.

Use Keras, an ML Library, for the First Time

As described in part 2 I started to learn Keras in order to write a first machine learning algorithm myself.

If you want to dive right in, find this repo "learning-ml" where I built a docker container which runs a jupyter notebook and where I played around with, first the example from a course and second exploring a simple task to use ML on. But let's go step by step.

The Example From the Keras Tutorial

As described in the earlier parts I followed Kamal's suggestions to learn using ML by reading Introduction to Keras for Engineers which turned out to raise many more questions than answers. My way of learning was to read the tutorial and figure out if I can understand what is being done. I failed.

Why did I fail Learning from "Keras for Engineers"

First, one should understand what is my background. I am a normal software developer, I have done a lot of low-level (without web), in 1999 I jumped into the web, I did a lot of PHP, Python, etc. and for the last 20 years I do mostly JavaScript extensively. I love to read the specs of technologies I use (if available) and I like to learn and understand the history and reasons why a certain technology was developed and is used for. And I know the applications technologies are used for change over time.

Back to learning from "Keras for Engineers". I thought I read through the code on the page, read the explanations and I will be set to go and make use of machine learning. Actually, my initial expectation was even simpler, I thought I download some library and throw right and wrong screenshots in some directories and tell the code to learn and then I have trained a model to detect right and wrong screenshots. It turned out to not be that easy.

You might succeed with that simple approach by going another path, but I am the kind of person, I like to understand what is happening under the hood. I started my career learning low-level languages like C and Assembler, I knew (more or less) well what was going on under the hood. So I expect the same here. I want to understand how does a ML code come up with the result that it comes up with.

Well, I already learned over the last years at HolidayCheck from my colleagues that worked with ML, that "understanding" and maybe even verifying why an ML model comes up with a certain result is not really how it works. Anyways, I am a bit stubborn :).

Switch to a TensorFlow Course via Video

I searched around and got stuck with a freeCodeCamp.org video Keras with TensorFlow Course - Python Deep Learning and Neural Networks for Beginners Tutorial. I did work through the first hour, because already learned in the first ten minutes to understand terminology which had raised question marks in the previously mentioned Keras tutorial. So I followed along.

I think I kinda got the hang of it. I learned to at least understand how the parts are connected, not that I really understand all the terms used yet, but the course gave me a good feeling of where things belong. So I did the first example from scratch, writing all the code by hand by fast forwarding the video for a second time to the places where the code was shown. Even though I did not totally understand the reason for building the first model I got it to run. See the repo and the source code online of the first "patients example". You can run it either in a jupyter notebook, which is handy, but not how I prefer it to do. Or you can run it on the command line after cloning the repo like so ./run.sh python 1-patients-example1.py. See the README for more detailed instructions.

My Own First Model Training

The actual fun part comes now. I want to train a model that I came up with, where I can also judge the predictions.

So I needed a problem, a simple problem to solve first. While I first thought I was in a dead end when the input is just simple numbers, I quickly realized I can come up with many examples that can be simple things for an AI to learn. I thought why not train an AI to learn to find even and odd numbers? Let me take away, the result first. I failed. I could not make the model learn that on my first try. But knowing and trying to find out why is a great learning too.

Even or Odd - Training my First Model

To be written about: