retrain_freq would allow the ML stage to only run on certain DDMD iterations. This optimization would make it easier to train a model from scratch without needing a pretrained model. The current problem is that the agent stage requires the model_selection stage to run each iteration where it expects a trained model at each iteration. This can be fixed by passing retrain_freq to the model_selection stage so it knows when to update the latest model weights versus when to use old model weights.
braceal
commented
3 years ago
retrain_freqwould allow the ML stage to only run on certain DDMD iterations. This optimization would make it easier to train a model from scratch without needing a pretrained model. The current problem is that the agent stage requires the model_selection stage to run each iteration where it expects a trained model at each iteration. This can be fixed by passing retrain_freq to the model_selection stage so it knows when to update the latest model weights versus when to use old model weights.