README.md

Configuration file cheat sheet

Hereunder is a complete list of the configuration parameters with their description:

• model:
  • The class name of the model to train. You can for example change the code of ALFINet.py into a file BetterALFINet.py and then use it for meta likelihood free inference.
  • e.g. ALFINet, BetterALFINet
• simulator:
  • It contains the parameters related to the simulator.
  • name:
    • The class name of a simulator implementation.
    • e.g. PoissonSimulator, LinearRegressionSimiulator, WeinbergSimulator, ...
  • parameters:
    • The list of parameters taken by the constructor of the simulator class.
    • e.g. [], sigma: 0.5, ...
• train:
  • It contains the parameters related to the training configuration.
  • nb_epochs:
    • The number of training epochs.
    • e.g. 130
  • nb_iter:
    • The number of iteration of ALFI (T in the manuscript).
    • e.g. 15
  • nb_theta:
    • The number of thetas (simulation parameters) in the meta training set. Each of them constitute a problem.
    • e.g. 1000
  • nb_x_per_theta:
    • The number of observations for each theta in the meta training set.
    • e.g. 500
  • meta_batch_size:
    • The number of problem taken from the training to perform one gradient update.
    • e.g. 16
  • batch_size_theta:
    • The number of thetas drawn from the random distribution proposal on the true simulation parameters.
    • e.g. 8
  • batch_size_x:
    • The number of observations generated for each of these theta.
    • e.g. 64
  • lr:
    • The learning rate.
    • e.g. 0.0002
  • save_every:
    • The number of epochs between two saveguard.
    • e.g. 5
  • test_every:
    • The number of epochs between two tests (useful to do training plots).
    • e.g. 5
  • loss:
    • The type of loss used. If you want to write your own it must be relevant to compare the proposal distribution with the true parameters value.
    • e.g. normal, MSE, l1
  • weight:
    • The weighting on the loss which should controls the tradeoff exploration/exploitation of the iterative process.
    • e.g. exp, oi, last, constant
• test:
  • It contains the parameters related to the testing configuration.
  • nb_iter:
    • Same meaning as for train (but the value can be different).
  • nb_theta:
    • The number of different simulation parameters for test.
    • e.g. 1000
  • nb_x_per_theta:
    • Same meaning as for train (but the value can be different).
  • meta_batch_size:
    • Same meaning as for train (but the value can be different). It is only useful to limit memory requirements.
  • batch_size_x:
    • Same as for train.
  • batch_size_theta:
    • Same as for train.
• networks:
  • It contains the parameters related to the training configuration.
  • use_grad:
    • Put the variational gradient () as input or not.
    • e.g. true, false.
  • split_theta:
    • Whether consider the optimization over each parameter independently or not.
    • e.g. true, false.
  • x_data_agg:
    • Configuration parameters of the network that encodes the observations.
    • hidden_size:
      • The number of units by hidden layer.
      • e.g. 50.
    • output_size:
      • The number of output units.
      • e.g. 50.
  • theta_data_agg:
    • Configuration parameters of the network that aggregates the encoded observations and the gradients.
    • hidden_size:
      • The number of units by hidden layer.
      • e.g. 50.
    • output_size:
      • The number of output units.
      • e.g. 50.
  • RIM:
    • Configuration parameters of the RNN.
    • hidden_size:
      • The number of units by hidden layer.
      • e.g. 50.
    • st_size:
      • The number of memory units of the GRU.
      • e.g. 50.
    • bounded:
      • The bound on the output of the RNN which represents the update step on the proposal distribution (a negative number if no bound).
      • e.g. -1, 0.2
• proposal:
  • name:
    • The class name of the proposal distribution on parameters value.
    • e.g. ConstantProposal, FixedVarianceGaussianProposal, GaussianProposal
  • sigma:
    • A parameter value of the constructor of the proposal, it could be something else than sigma.`
    • e.g. 0.5.
• plot:
  • title:
    • Whether or not the figures should contains a title.`
    • e.g. true, false.
  • rmse_t:
    • nb_theta:
      • The number of theta to produce the plot that shows the evolution of the RMSE along iterations for these parameters.
      • e.g. 12.
  • hist-x:
    • nb_theta:
      • The number of theta to produce the plot that compares the observations generated with the final proposal distribution output by ALFI with the observations of the testing set.
      • e.g. 6.
    • nb_samples_real:
      • The number of observations used to do the plot for the true parameters.
      • e.g. 5000.
    • nb_samples_gen:
      • The number of observations used to do the plot for the estimated parameters.
      • e.g. 5000.
    • nb_bins:
      • The number of bins in the histograms.
      • e.g. 10.
  • init_box:
    • Box plot that shows what is the variation of the final proposal (after T iterations) depending on the starting proposal.
    • nb_theta:
      • The number of theta and so the number of box plot.
      • e.g. 12
  • init_rmse_t:
    • This plot shows what is the impact of different starting proposal on the evolution of the RMSE along iterations.
    • nb_theta:
      • The number of theta and so the number of sub plots.
      • e.g. 12
  • init_comparison:
    • Configuration to pre-compute data for the two plots above.
    • nb_inits:
      • The number of starting proposal.
      • e.g. 12.
    • nb_theta:
      • The number of theta (should be greater or equal than for the two previous plots configurations.)