pytorch

Training code for CNN with PyTorch.

Installation

On my Linux box with NVIDIA gfx card (your setup may vary):

pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128

Running pipeline

• Go to cmd/mlproducer and do go build
• Collect lots of data with autoplay and copy /tmp/autoplay.txt somewhere
• Activate your venv then cat /tmp/autoplay.txt | ../cmd/mlproducer/mlproducer | python3 training.py
• You can pipe in pv -br after the call to mlproducer to see what the byte rate is. I see like 700 MiB on my computer. It can probably be made faster with more efficient mlproducer code.
• Wait a long time (maybe not that long, depends on how many games you used).
• Profit!

Convert

export.py - export to onnx onnx-to-tensorrt.py - make sure the python tensorrt version matches whatever the triton container expects. this part is a pain in the ass. (pip install tensorrt==10.11.0.33 for example)