Smart Poi

Shout out to ChatGPT for making code documentation easy

Just a quick shout out to ChatGPT for helping me with some drudgery in my job. I think this is the best use for it – basically summarising text.

I do have reservations about the skill of ChatGPT and others when it comes to actual coding, having tried ChatGPT, HuggingChat and Claude, as well as Bard and Bing.

That’s it really. Docstrings for Python, JSDoc for JavaScript, etc. Just type something like “Please add c++ documentation for the following function in a class”. Or just “Add documentation here” works fine. If you are in a hurry you can type “docs” and it will most likely get the idea. You follow this up by a paste of your code of course.

Here is a sample of some cool looking documentation I generated recently for a Python helper class – with help from my new coding buddy, ChatGPT.

Don’t forget, always check the output from ChatGPT – it is often horribly wrong, trust me!

The free ChatGPT course was a lot of fun

About a month ago I completed the free course – direct from OpenAI – called “ChatGPT Prompt Engineering for Developers”

This course uses the OpenAI api and helps you create interactive chatbots with a purpose. I promptly started working on a chatbot to take orders for my entertainment company – Big Top Entertainment.

The instructions I gave to the bot included a price list for services and motivation (“you are a customer service agent, make sure to get payment”) and just for fun I included the option to accept chickens as payment. Here is an example conversation with the bot I coded (in Python):

Welcome to Big Top Entertainment Chat Bot, how can I help you?
Enter text: I need some Balloons pls
Assistant: Hi there! Great to hear that you’re interested in our Balloon Twisting service. How many hours of Balloon Twisting would you like to order?
Enter text: 4 hours
Assistant: Awesome! We can definitely arrange that for you. Just to confirm, you’d like to order 4 hours of Balloon Twisting, correct? Also, when would you like to schedule this service?
Enter text: monday
Assistant: Sure thing! Just to clarify, would you like to schedule the Balloon Twisting service for this coming Monday? And what time would you like the service to start?
Enter text: 2pm
Assistant: Great! So, to confirm, you’d like to schedule 4 hours of Balloon Twisting starting at 2pm this coming Monday. Is that correct?
Enter text: yes
Assistant: Perfect! I have you down for 4 hours of Balloon Twisting starting at 2pm this coming Monday. Is there anything else you’d like to add to your order? Maybe some Face Painting or Stilt Walking?
Enter text: No thanks
Assistant: No problem at all. Just one last thing, could you please provide me with the address where you’d like the Balloon Twisting service to take place?
Enter text: 4 Bigtop Lane
Assistant: Thank you! We have your order for 4 hours of Balloon Twisting at 4 Bigtop Lane on Monday at 2pm. The total cost for this service will be R1600.00. We accept payment in cash or chickens, with each chicken valued at R100. How would you like to proceed with payment?
Enter text: chickens
Please try again in a second (rate limiting)chickens
Enter text: chickens
Assistant: Great! We accept chickens as payment. Please let us know how many chickens you would like to use to pay for the service.
Enter text: chickens
Assistant: I apologize for the confusion. Could you please let me know how many chickens you would like to use to pay for the service?
Enter text: how many does it cost?
Please try again in a second (rate limiting)
Enter text: how many chickens??
Assistant: I apologize for the confusion. The cost of the Balloon Twisting service is R1600.00. Since each chicken is valued at R100, you would need to provide 16 chickens to cover the cost of the service.
Enter text: OK great here you go 16 chickens
Assistant: Thank you for your payment of 16 chickens. We have received your order for 4 hours of Balloon Twisting at 4 Bigtop Lane on Monday at 2pm. We look forward to providing you with our service. If you have any further questions or concerns, please don’t hesitate to contact us. Have a great day!
Enter text: Thank you!
Assistant: You’re welcome! It was a pleasure doing business with you. Have a great day!
Enter text:

Conclusion

Chat bots are fun to code. Clearly I could have expanded on this with a full web interface, handing off to a payment gateway (chicken farm?) etc but it was enough to get the idea.

I don’t think I will be going further with the paid api after my $5 free trial expires though (actually it’s my second one so far – tip: duckduckgo email aliases are great if two factor isn’t enforced). In further testing I managed to make a booking of 4 months straight (7 days per week) for 20 face painters – which we would never do – and in addition the bot refused to send a truck to collect the chickens (“As an AI Language Model, I am unable to do that…”). It did, however, refuse to take bookings of more than 8 hours per day. I didn’t tell it that, but we wouldn’t take anything longer for sure.

I have already attempted to do the same using huggingchat api but they changed something halfway and I gave up on it for now. Next time I am bored I will have another look, since it’s actually open source and self-hostable.

Image Classification for Magic Poi images using Machine Learning

The title says it all. Here are a couple of video’s to demonstrate the process:

Testing the version 5 ML model accuracy:

Testing the image recognition model

Deployed using temporary database:

I added a filter for different categories found by the machine learing model

Todo:

  • New images uploaded are not categorized – because I did the classification at home and just uploaded the database. I still need to get the image recognition to work on the server.
  • Scrolling is really not great, gets stuck sometimes and also new images loaded aren’t filtered.
  • Re-categorizing support – and support for more than one category per image
  • Remove bad images (or just hide them).
  • Refine the model to do better (“text-img” recognition isn’t great, for example).
  • Add more categories.

The categories are live! Check out http://magicpoi.circusscientist.com (and sign up for the good stuff).

Sign up for updates:

UPDATES:

Sign up for our update alerts:

bin files and Magic Poi

I recently received a question via email which I thought might need a bit of answering – “Hello, may I ask how to convert the image into a bin file? I couldn’t find the source code on the website and look forward to your reply. Thank you very much.” Here is my response:

(tl:dr I didn’t share source yet but there are two ways to do it)

Great question! I haven’t been working on the file access much yet for the new MagicPoi. The hardware and getting the binary files is coming second to having an easy way to manage images – which I am spending most of my time on right now.

The Android app had a bug with creating .bin files which I never managed to fix, so I did update the “Smartpoi Android App” page to link to two PC programs which you can download – see the page here: https://www.circusscientist.com/smart-poi-android-app/ – just scroll to the bottom to see the (Linux Only) versions. These are easy-to-use drag-and-drop apps which will convert the image files to .bin format and also upload wirelessly to any connected poi. The .bin files are stored temporarily in a folder in the program files, I am sure you can find them there once you try it. I didn’t share the source because frankly the code is a terrible mess which even I am embarrassed about #messyprogrammer – I am getting better at this by the way, look forward to some documentation coming soon!

Currently the magicpoi website api is limited to returning 10 image names. You can see this by typing this into the browser url bar:
http://magicpoi.circusscientist.com/api/get-filenames

To download one of the .bin files listed, you can use the filename(xxx.jpg.bin), also in the browser bar – eg:
http://magicpoi.circusscientist.com/api/output/ajzs2w1aczeq.jpg.bin

Incidentally the example PlatformIO sketch linked from the “api” section here https://www.circusscientist.com/magic-poi-api/ does exactly this, getting the list of 10 files and then using the /api/output endpoint to download .bin files – to the D1 Mini in this case – and then displaying them in 72px. I would have to check but I think maybe all .bin files are compressed to 72px.. yes I checked – currently that’s the default, with option to support any size (thank goodness).

All files which are uploaded using the magicpoi.circusscientist.com site are automatically converted to .bin files. The .bin files are simply compressed image files – with one byte representing one pixel in R3G3B2 format (in the case of MagicPoi I think I rotate them 90 degrees also for the line by line scanning).

To get any image .bin file from the MagicPoi site at the moment you can use a quick hack – if it is an image you uploaded yourself you have to “share” it for this to work. Then scroll down to where the original image is visible (on the “profile” page) – not the rotated version! Right click and select “Open image in new tab” or just “copy url” – then go to the new tab and copy the image name. Now use the name with “.bin” added on for the download api:
http://magicpoi.circusscientist.com/api/output/ajzs2w1aczeq.jpg.bin

I’m sorry I haven’t really been working on the poi hardware and that side of the project. Like I said, I’m focusing right now on front end usability for the web interface. Soon there will be categories for the images, for example, and sorting.

Once the interface is working I will return to hardware and firmware – and also documentation and new features as well as proper api accessibility.

I hope my response answered your question. Everything is a work in progress, but there is progress.

Regards

Tom

Computer generated Magic Poi patterns with P5.js

Computer generated patterns – let me count the ways.

  • C++ code on-device

    – a simple for-loop using FastLED’s “Palletes”
  • Processing code – on Android, streaming to poi.


    – created using Processing “Mask”, the coloured parts of the image change colour to create an ever-changing image sequence. I also implemented a sound activated version which shows a pattern according to volume on different frequencies (“Beats”) and a static “Cylon” effect (“Zap Game”)
  • Similarly C can be used to stream generated patterns to the poi (see here).
  • I have experimented with AI generators using different prompts – for example:

    – These are from “Midjourney”
  • Finally, my current mission, using P5.js within the magicpoi website to generate pics with code. Examples:

Example:

The above example creates an infinite succession of 72x72px images – perfect for display on my 72px poi. I adapted the code from this showcase example on the P5js website.

My plan is to add this – and many more – options to the magicpoi website. They will complement the user generated images and be downloadable directly to the internet connected Magic Poi.

PS: Currently, if you want to see what any of the images generated above will look like on the poi, you can right click on the changing image, select “open image in new tab”, download the image and then on the magicpoi profile page (you need to be signed in for this to work – sign up, it’s free!) select upload and drag the image onto the page, or choose it from the file menu. Then you can turn this: into this:

As you can see, the possibilities are endless. I plan on making the P5.js image generation into a plug-in system for Magic Poi.

Solving a simple problem with ML

ChatGPT has been helping me out

A fair amount if I am going to be honest here. The problem for me always comes down to syntax. Currently focused on Full Stack IOT development my day consists of switching between coding in Python (Flask api and web back-end), JavaScript (with JQuery for front-end) and C++ (embedded).

I find that one of the more interesting parts of any Full Stack IOT code base comes down to dealing with requests and responses. A website has a login page, sends a request to the server. An IOT device has a key to access the online api, sends a request to the server. The server needs to respond but first check the input is valid (sanitize it).

Python/JavaScript/C++ all have their own ways of working with text input. In JavaScript for example you need to trim() but in Python it’s strip() – and in C++ (embedded) it’s a bit more complicated operation involving char arrays. The point is that when switching around between languages I sometimes find myself going blank on the specifics of each approach. That’s when ChatGPT comes in handy for me – as a helper, faster than a google search.

I recently came across an issue which was helped by ML in a different way – namely parsing unstructured text input.

The Problem

I had a series of spreadsheets which I was extracting data from (using Pandas) but the date format was unfortunately unstructured. For example: 

12th, 13th Dec 2022 and 5th-8th Jan 2023

In order to continue with the project I needed the date in Pandas date format 

2022-12-12

If you give it “12th December 2022” as an input, Pandas can do this for you, but it couldn’t cope with the format in the spreadsheets. Regular expressions weren’t going to do the job either, considering the range of formats. So I asked ChatGPT. The prompt: 

Please can you parse the following dates as separate dates in list format, suitable for python pandas usage:
12th, 13th Dec 2022 and 5th-8th Jan 2023

The reply:

['2022-12-12', '2022-12-13', '2023-01-05', '2023-01-06', '2023-01-07', '2023-01-08']

As you can see, this is exactly what I was looking for! Unfortunately my ChatGPT free trial has ended – not that I am opposed to paying for software but that’s not the only issue stopping me from using the api for this. I have been burned in the past by proprietary software solutions which have arbitrarily changed and/or cancelled a product which I relied on. This is a simple use case but I can imagine myself taking advantage of it more and more, then OpenAI pulling the rug (or just raising prices) and leaving me with a whole bunch of useless code.

The Solution:

After looking around I found NLPCloud (no affiliation, just it looks cool and works for me*). They use Open Source Generative AI models which in theory I could replicate and host anywhere. I decided to give it a try.

It works! The free plan api key is now plugged into my program and when my program encounters a date format it can’t deal with, a query is sent. 

response = client.chatbot(f"Please can you parse the following dates as separate dates in list format, suitable for python pandas usage: {dates_unformatted}", "Human asking AI for help with a text summarization task. ", [])  #text, context, history

The response is exactly the same quality as the ChatGPT one, using the “finetuned-gpt-neox-20b” model.

What simple things has ML been helping you out with?

*Update: A couple of months later, tried out NLPCloud again and it’s broken. Luckily the same code works fine using hugging-chat-api. I really need to get onto self-hosting, every api breaks sooner or later it seems.

Upgrading the Magic Poi site

Currently the Magic Poi site relies on some Processing (Java) code which is called from python – as well as other bash scripts to do certain things. I wanted to make it faster and cleaner by using Python for everything, so I did some test scripts. (Github). Also, the Processing code required a virtual screen buffer to run on the server constantly!

The main functionality involves compressing images into the .bin files which are saved to the poi memory for display. Since there is limited memory, and to save bandwidth, I am using the R3G3B2 algorithm to compress images. The poi only accepts images which are rotated 90 degrees clockwise, so that needs to happen too, as well as dynamically re-sizing any images which are wider than the 72px available.

I am still conflicted about whether to keep the Android app as the control for the poi, or to do something else, but making everything available as an api on the server keeps it flexible for now. The RotateImageVisualPoiStyle api can be accessed right now. Upload your image, and see what it will look like on the poi! 

# change the /path/to/image.jpg to your own image file name
curl --location --request POST 'http://magicpoi.circusscientist.com/api/rotate_visual_poi_style' --form 'file=@"/path/to/image.jpg"' --form 'fixed_width="72"' --output 'rotated_image.jpg'

MagicPoi – slow progress

Announcement:

The factory where MagicPoi were being developed had a fire last month! All of the prototype poi melted, so all I have left is my breadboard version. This is going to cause some delays..

Updates:

  • The Magic Poi website now saves .bin files for every image uploaded. This is accessible via an api which I am working on accessing from the ESP32 and ESP8266. The .bin files are byte compressed images for display on the poi.

    Example – list all shared .bin files on the server:
    curl http://magicpoi.circusscientist.com/api/get-filenames

    One thing to note is that for testing purposes, all new .bin files are shared now. Please be aware that your images could be accessed (temporarily). This is because authentication is an extra step, so I’m working on the basic functionality first, with a public folder. The authentication works using jwt and I am having to access it from Java and C++. I already have a working Arduino test code which implements jwt correctly – security is important, so I want to do it right. One problem I already had to overcome is the fact that LittleFS file system has a limit of 31 characters for filenames (including path). The randomly generated filenames on the server are longer than this.
  • Not sure if I mentioned this before, the SmartPoi “Upload for offline use” option in the Android app doesn’t upload the whole image at the moment. I am not sure why, perhaps it’s something to do with the timing loop moving to the next image too fast.
    This option is going to be replaced by the web api anyway, so it’s not a priority, but I have compiled a desktop app which I use for converting and uploading (Linux only). You can get that for 36px here or 72px here – it can upload 20 images, just drag and drop your pics (need to be 90 degrees rotated, unlike the website) onto the window, connect to Poi AP and press “Upload”. You need to extract it and run from the command line but it should work on any Linux distro (tested on Ubuntu and Arch).
    At some point I will probably get annoyed at having to use my Laptop to upload images and fix the app though. Do email me if you need this for Windows, I might be persuaded to compile it for you using a VM if you ask nicely.
  • With the new APA102 2020 version being more available now, when Magic Poi is eventually launched it will have at least 100px, I’m thinking 128 is a good number but don’t quote me.
  • You may have noticed all of the hype around AI lately. A few years ago I used GAN with PyTorch to generate images for the poi (after training with around 100 poi pics, probably not enough). It was a big fail, but maybe it’s time to take another look. Would it be cool if you could just tell your poi what type of image you want to display, and it did it instantly?
    I do already have loads of Processing sketches I made to try and computer generate nice pics (some are on the Android App).

Feedback is always welcome.