An AI for the web based game geoguessr. The model uses a ConVneXt-B backbone pretrained on ImageNet22k, which was then finetuned on a self collected dataset.
The conda environment required to train the model yourself can be created with the following cmd:
conda create -n <environment-name> --file req.txt
main thing is that it uses pytorch 1.12.0 and python 3.8.15. Probably works with most environments, there isnt any special packages used.
For inference, I think it should be able to run with ~1GB of RAM on CPU. I haven't done any performance benchmarks so feel free to try it out and let me know. Training was done on a RTX 3090 ~30 epochs, note that I was able to train with batch_size=7 with torch.amp which shortened training time from 1.5hrs per epoch to 50 minutes.
The data used to train the bot is pulled from Google street view static API. You will first need to create an account and fill in your API key and secret SHA256 hash in creds.py. Then run
create_data.py
The model then can be trained by running main.py with your own args.
The weights to the two models can be found here: drive link
- currently only works for USA. I'm too poor to download enough training images for the entire world. Paypal me $20k and I'll do it
Naive: just predict x,y
Bin Based: predict state then residual
- No pretrain is the bin based model without pretraining on Imagenet22k
random guess: guessing a random state
always center: always predicting the center of the USA [39.50,-98.35]
Create browser extension to interface with game UI