ClarityCoders
can you provide a colab version
You can use the code provided in the video. This code works.
[1] !pip install pandas --upgrade
[2] !pip install fastai --upgrade -q from fastai.vision.all import * from google.colab import drive
[3] from google.colab import drive drive.mount('/content/gdrive', force_remount=True)
[4] !ls '/content/' !tar -xvf "/content/gdrive/MyDrive/GateCrash.tar.gz"
[5] !ls '/content/GateCrash'
[6] path = Path('/content/GateCrash') fnames = get_image_files(path) print(f"Total Images:{len(fnames)}") def label_func(x): return x.parent.name dls = ImageDataLoaders.from_path_func(path, fnames, label_func)
[7] dls.valid.show_batch(max_n=8, nrows=1)
[8] learn = cnn_learner(dls, resnet18, metrics=error_rate)
[9] lrs = learn.lr_find(suggest_funcs=(minimum, steep, valley, slide))
[10] learn.fine_tune(2)
[11] learn.path
[12] learn.export('mymodel.pkl')
The numbers infront of each segment representing the cells. The results of the AI might not be that good, due to the fact, that only 1 keypress is recognized and saved with an image. So if you play a racing game, you would want (w+d) or more keypresses.
Also you might want, that your keypresses/ mouseinputs and images are automaticly stored and sorted into the folders. I made it recognizing multiple keypresses/mouseinputs and sorting those into the folders, so i just need to compile those folders as .tar.zip
Also the trained Agend will also only be able to press 1 key at a time. You need to adjust this too. Otherwise this project would be a toy, to get into behavioural cloning.
This approach is heavily depending on the base code. e.g. The provided code always presses "w". In FallGuys you don't have multiple keystrokes, because you only press "a" or "d". So what you might want to add:
- multiple keystrokes
- mouse inputs
- automaticly sort the images into the folders, for those keys/mouseinputs used in the specific game
- make the trained agent able to press multiple keystrokes
Also [10] learn.fine_tune(2) : 2 epoch are more then enough, if your agent, don't behave like you want, then you should change the base code. The agent only can do stuff, based on the data it was given.
