many_well.yaml

#defaults:
#  - override hydra/launcher: joblib

target:
  dim: 32

flow:
  layer_nodes_per_dim: 10
  n_layers: 10
  act_norm: false
  use_snf: false
  resampled_base: false
  snf: # config if snf is used.
    transition_operator_type: hmc # metropolis or hmc
    step_size: 1.0
    num_steps: 5 # leapfrog steps if hmc, else metropolis accept/reject steps
    it_snf_layer: 2

fab:
  # loss type options: fab_alpha_div for standard FAB loss
  # target_forward_kl: forward kl estimated with samples from the target
  # flow_reverse_kl, flow_alpha_2_div_nis for revers KL/alpha_2_div estimated with flow samples
  loss_type: fab_alpha_div
  alpha: 2.0 # null
  transition_operator:
    type: hmc
    n_inner_steps: 5
    init_step_size: 1.0
  n_intermediate_distributions: 4


training:
  tlimit: null # time limit in hours
  checkpoint_load_dir: null # or null
  seed: 0
  lr: 3e-4
  batch_size: 2048
  n_iterations: null
  n_flow_forward_pass: 10_000_000_000
  use_gpu: true
  use_64_bit: true
  use_buffer: false
  prioritised_buffer: false
  buffer_temp: 0.0 # rate that we weight new experience over old
  n_batches_buffer_sampling: 8
  maximum_buffer_length: 512000
  min_buffer_length: 65536
  log_w_clip_frac: null
  max_grad_norm: 100.0
  w_adjust_max_clip: null # clipping of weight adjustment factor for prioritised replay


evaluation:
  n_plots: 50 # number of times we visualise the model throughout training.
  n_eval: 50 # for calculating metrics of flow w.r.t target.
  eval_batch_size: 2048 # must be a multiple of inner batch size
  n_checkpoints: 10 # number of model checkpoints saved
  save_path:  ./results/many_well/seed${training.seed}/



logger:
  pandas_logger:
    save_period: 1000 # how often to save the pandas dataframe as a csv
#  wandb:
#    name: ManyWell32
#    project: fab
#    entity: flow-ais-bootstrap
#    tags: [alpha_2_loss,ManyWell32]
#  list_logger: