We provide an overview of the important files and their relationships.

  • Training: train.py is the main file during training. It will run every epoch, save network parameters. It will call other files one by one that are used for training.

    • Before training:

      • Load optimizer network congifuration: util.py. The train.py will call util.py in order to red optimizer net configuration.
      • Load optimizee problem: util.py -> problems.py. The train.py will call util.py, which will call problems.py in order to load the optimizee problems used for training.
      • Load optimizer network: meta.py. The train.py will use the optimizer net configuration information to build an optimizer in meta.py.
      • optimizer-optimizee graph: meta.py. The meta.py will also build a complete computational graph with both optimizer and optimizee.

    • During training:

      • run every epoch: util.py. This file will run every epoch during training.
    • After training:
      • save meta net parameters: train.py. This file will save the meta net parameters.

  • Evaluation: evaluate.py is the main file during evaluation. It will run a whole evaluation trajectory and save the trajectory. It will call other files one by one that are used for evaluation.

    • Before evaluation:

      • Load optimizer network congifuration: util.py. The evaluate.py will call util.py in order to red optimizer net configuration.
      • Load optimizee problem: util.py -> problems.py. The evaluate.py will call util.py, which will call problems.py in order to load the optimizee problems used for evaluation.
      • Load optimizer network: meta.py. The evaluate.py will use the optimizer net configuration information to build an optimizer in meta.py.
      • optimizer-optimizee graph: meta.py. The meta.py will also build a complete computational graph with both optimizer and optimizee.

    • During evaluation:

      • run the trajectory: util.py. This file will run the trajectory during evaluation.
    • After evaluation:
      • save the trajectory: evaluate.py. This file will save the trajectory.