ml.yaml

# Config file specifying the setup of a Boltzmann Generator


data:
  transform: data/position_min_energy.pt
  test: data/val.pt
  train: data/train.pt


system:                             # Properties of molecular system

  temperature: 300                  # Double, temperature of the system
  energy_cut: 1.e+8                 # Double, energy level at which regularization shall be applied
  energy_max: 1.e+20                # Double, maximum level at which energies will be clamped
  n_threads: 20                     # Int, number of threads to be used, number of cores if null
  transform: internal               # String, type of the transformation
  shift_dih: False                  # Bool, flag whether to shift dihedral angles
  env: implicit                     # String, environment of the molecule, can be implicit or vacuum


flow:                               # Properties of the flow model

  type: circular-coup-nsf           # String, type of the flow
  base:                             # Base distribution
    type: gauss-uni                 # Type of the base dist
    params: null
  blocks: 12                        # Int, number of Real NVP blocks, consisting of an ActNorm layer
                                    # if specified, a permutation, and a affine coupling layer
  actnorm: False                    # Bool, flag whether to include an ActNorm layers
  mixing: null                      # String, how features are mixed
  circ_shift: random                # String, whether to shift circular coordinates, can be none,
                                    # constant, or random
  blocks_per_layer: 1               # Int, number of blocks per layer
  hidden_units: 256                 # Int, number of hidden units of the NN in neural spline layers
  num_bins: 8                       # Int, number of bins of the neural splines
  init_identity: True               # Bool, flag whether to initialize layers as identity map
  dropout: 0.                       # Float, dropout probability for the NN layers


fab:

  transition_type: hmc              # String, type of transition operator used
  n_int_dist: 8                     # Int, number of intermediate distributions
  n_inner: 4                        # Int, number of steps between intermediate distributions
  epsilon: 0.1                      # Double, step size of HMC
  adjust_step_size: True            # Bool, flag whether to adjust step size
  loss_type: forward_kl             # String, loss to be used
  alpha: null



training:                           # Properties of the training procedure

  max_iter: 250000                  # Int, maximum number of iteration
  warmup_iter: 1000                 # Int, number of iterations of linearly warm up learning rate
  optimizer: adam                   # String, name of the optimizer
  batch_size: 1024                  # Int, batch size used during training
  learning_rate: 5.e-4              # Double, learning rate used during training
  lr_scheduler:
    type: cosine                    # String, kind of LR scheduler, can be exponential, cosine
  max_grad_norm: 1.e3               # Double, limit for gradient clipping
  weight_decay: 1.e-5               # Double, regularization parameter
  log_iter: 20000                   # Int, number of iterations after which loss is saved
  checkpoint_iter: 200000           # Int, number of iterations after which checkpoint is saved
  eval_samples: 1000000             # Int, number of samples to draw when evaluating the model
  filter_chirality: none            # String, whether to filter chirality during training or evaluation
  seed: 0                           # Int, seed to be used for the random number generator
  save_root: experiments/aldp/ml/00 # String, root directory for saving the output