Astral is a machine learning AI for game automation. Astral recognizes and observes game objects in the same way as any other human would, using the Yolo algorithm, powerful GPUs, a trained model, and a screen as input.
💡 Unfortunately, none of the compatible CUDA versions work with Windows 7. For operating systems below Windows 10, they can opt in for CPU.
💡 Unfortunately, AMD GPUs are not supported; for GPU support, an Nvidia GPU with CUDA is required.
Full instructions for installing required dependencies at https://github.com/maalik0786/FastYolo 💡 Although different CUDA versions greater than or 11.0 should work, tested with CUDA 11.0 with cuDNN 8.4.0 and it worked.
The implementation of YoloV5 uses ML.NET, for GPU support follow the official instructions at https://docs.microsoft.com/en-us/dotnet/machine-learning/how-to-guides/install-gpu-model-builder the implementation also works on CPU.
🔬 A GPU is favored for faster predictions, while a CPU can still do predictions.
🌐 If the input is through network, it is preferred that both computers be on the same lan network to avoid high latency; nevertheless, another machine on different network (via the internet) may create latency and delays.
🦥 The most playable delay for both input and predictions at the moment is 10 predictions per second 🔮 or 100ms delay (both input and predictions), which is not a big issue with newer GPUs and decent settings.
⚡ Below is a list of graphics cards tested on resolutions and its prediction speed (predictions/second).
⚠️ Different factors can cause the GPU to perform better or worse; please take this table with a grain of salt and don't place too much faith in it; different hardware and environment can result in different results, such as having different video settings in a game, different drivers, room temperature, or having background processes, for example.
Predictor GPU | Game GPU | Game | Setup | 960x540 | 1280x720 | 1920x1080 |
---|---|---|---|---|---|---|
GTX 1050 | GTX 1050 | CSGO | 🖥️ Same Machine | 10 p/sec | 11 p/sec | 12 p/sec |
MX 330 | MX 330 | CSGO | 🖥️ Same Machine | ? | ? | 5 p/sec |
💡 As you can see, a mid-level gaming desktop GPU from 2016 may get decent results; but, newer GPUs can produce better and quicker predictions.
💡 Picture resizing can take some time; if the configurations demand additional image altering, this can slow down the prediction.
🖼️ Astral uses Yolo for object detection, more information on creating models here: https://github.com/AlexeyAB/darknet#how-to-train-tiny-yolo-to-detect-your-custom-objects
💡 There are several versions and sizes of Yolo; currently, the C# implementation may function on any model trained, for the quickest predictions, we choose the smallest Yolo model size. The larger the model size, the more accurate the prediction, but slower; the Yolo implementation used can be seen at https://github.com/AlexeyAB/darknet.