Some brains make use of a global class state of the brain class (usually for sparse matrices). As far as I can tell this global class state is not preserved at the end of an experiment. This is no problem for logarithmic sparse matrices, as they are deterministic to begin with.
But it will probably be a significant problem for random sparse matrices.
It could still be possible that they will work because of the fixing of random seeds at the beginning of the experiment, but if so, than it would be very prone to errors.
I suggest the following solution:
[ ] add brain_class_state.pkl the output dir, which contains the global class state of the experiment (you can get this from IBrain.get_class_state()
[ ] in render_hof.py read this variable and pass it to Experiment
[ ] in Experiment initiate the brain class state from the variable and only create a new brain class state for new experiments.
Some brains make use of a global class state of the brain class (usually for sparse matrices). As far as I can tell this global class state is not preserved at the end of an experiment. This is no problem for logarithmic sparse matrices, as they are deterministic to begin with.
But it will probably be a significant problem for random sparse matrices.
It could still be possible that they will work because of the fixing of random seeds at the beginning of the experiment, but if so, than it would be very prone to errors.
I suggest the following solution:
brain_class_state.pkl
the output dir, which contains the global class state of the experiment (you can get this fromIBrain.get_class_state()
render_hof.py
read this variable and pass it toExperiment
Experiment
initiate the brain class state from the variable and only create a new brain class state for new experiments.related issues: #34