I trained a 2D car driving model for roughly 12 million steps and it’s showing the behaviours I want, I however want to keep training it in order to improve the speed that it drives at. However, for some reason when I resumed training with --resume it reset the steps to 0. I still have the .onnx file, but it overwrote the checkpoints. I tried using the --initialize-from field in the CLI to give it the run ID from the onnx file but it did not retain any of the behaviours, despite using an identical configuration file except for run name. How can I resume training from an existing .onnx file?
mlagents-learn --initialize-from=“” --run-id=“”