EnglishGarfield

first few files and a repo established

Im calling it a day for this project, Ive hit the limit for what my machine can do and its not feasible to let it sit for hours at a time anymore. Ive learned alot but its onto my next project https://summer.hackclub.com/projects/10107 a music prediction algorithm. ta ta and farewell (like if you get the reference)

300 rounds into the data chunking, this will take a while to process before i can parse it into the new network / model

Ive just finalised what is quite possibly the best bit of code ive ever written. CLI args, multimode operation, fully user (when building / training) adjustable and inbuilt error messages that point to usable lines of code instead of blank space, i am so incredibly proud of this code file :)

website created and pushed to demo link button along with a readme.md

first few code methods set up and its starting to come together. not much to show yet

the process gave a SIGKILL 9 so the program ran out of memory, im going to revmove these limits and try again

Ive started the third training round and im now through 92/382 data chuncks with aprox ~300k games loaded. making full use of gc.collect()

I have added 4 model architectures
1. ResNet with Attention (recommended)
2. Transformer-based
3. EfficientNet-style
4. Hybrid CNN-Transformer

I have also added 6 Learning Rate schedule options
1. None (constant learning rate)
2. ReduceLROnPlateau (automatic reduction when loss plateaus)
3. Exponential decay
4. Step decay
5. Cosine decay
6. Warmup + Cosine decay

which are explained at the bottom of this file in the log book in triple quotes
https://github.com/English-Garfield/CompSci_IK/blob/main/data_processing/training.py

Im starting the next round of training with ResNet with Attention architecture that can be seen bellow as the image

Ive updated the demo webpage to include a playable version of the game with my homegrown model as the AI player. I used Claude 4 to help debug some bugs as i am new to HTML + CSS

It works!! video attached of device functioning

Ive put shortcuts key inputs into the code and coded the key matrix for the commands to be held in. working on the colour assignment and toggles for the key stroke inputs on the pad

Ive set up the shortcuts and Im now working on implementing them into the code. still working on the customisation of key colours

Ive gotten the pages implemented and now moving onto the hotkeys, Im going to use keyboard shortcuts to trigger automations on the target computer. each key will have a shortcut key stroke input assigned dependent on the page

I've made the pad respond to inputs and made a pretty wave animation with the onboard LED's. time for gives as i have no tangible thing to put as the image

Ive created the website demo created for project concept and deployed using a workflow to a github page. Onto coding the actual device now

MacroPad for live event productions as a vision mixer / Audio engineer. will allow me to do my voluntary work better and can be customised for home use. cheeper alternative to Steamdeck

Getting closer to full completion, 25 epochs to go and accuracy is improving

Today the accuracy of my model is at 0.2524 with 35 epochs to go!!! its around 400-800 elo so lots to improve on in future iterations. Lets just see where this round of training finishes

nearly halfway through and at an accuracy of 0.2518 (its statistics) which is getting better. 58 epochs to go tho so stil loads of time to improve :)

just realised i have not put a devlog of the game working so here it is, the game works. Yay

Training is underway, I have to make this devlog to satisfy the SoM gods who believe progress comes every 10 hours (it doesn't)

sadly after 18 hours the model ram out of RAM and stopped, i have adjusted the parameters to allow for more efficient memory usage and the library gc has been implemented more effectively. I have made errors that have been corrected, see you in 68ish hours

epoch 28/100
accuracy - 0.2434
overall progressing quite nicely and it is approximately 18/68 hours into first training session

I am Still waiting for First round of training to finish. approximately 20/100 epochs through and 8 hours deep. however a basic demo page is now active!

Training in progress for new model with 2 million parameters that are trainable

main.exe created in /src/build/main

too see devlog for previous work look in the /decumentation file at the .docx file