config_seml.yaml

seml:
  executable: 'train_seml.py'
  name: "gemnet"
  output_dir: "slurm_logs"
  project_root_dir: "."

slurm:
  experiments_per_job: 1
  sbatch_options:
    gres: 'gpu:1'
    mem: 40G
    cpus-per-task: 2
    time: 07-00:00
    partition: gpu_all

fixed:
  num_spherical: 7
  num_radial: 6
  num_blocks: 4

  emb_size_atom: 128
  emb_size_edge: 128
  emb_size_trip: 64
  emb_size_quad: 32
  emb_size_rbf: 16
  emb_size_cbf: 16
  emb_size_sbf: 32
  emb_size_bil_trip: 64
  emb_size_bil_quad: 32

  num_before_skip: 1
  num_after_skip: 1
  num_concat: 1
  num_atom: 2

  cutoff: 5.0
  int_cutoff: 10.0

  mve: False
  loss: "rmse"
  forces_coupled: False
  envelope_exponent: 5
  extensive: True

  rho_force: 0.999
  ema_decay: 0.999
  weight_decay: 0.000002

  learning_rate: 0.001
  decay_steps: 4500000
  decay_rate: 0.01
  staircase: False
  decay_patience: 5
  decay_factor: 0.5
  decay_cooldown: 5
  agc: False
  grad_clip_max: 10.0

  restart: null
  tfseed: 1234
  data_seed: 42
  scale_file: "scaling_factors.json"
  comment: "GemNet"
  output_init: "HeOrthogonal"

  logdir: "logs"
  dataset: "data/coll_v1.2_train.npz"
  val_dataset: "data/coll_v1.2_val.npz"
  num_train: 0 # derived from dataset
  num_val: 0 # derived from dataset

  patience: 5
  evaluation_interval: 7500
  save_interval: 7500
  warmup_steps: 3750
  batch_size: 32
  num_steps: 1500000

grid:
  triplets_only:
    type: choice
    options:
      - True
      - False

  direct_forces:
    type: choice
    options:
      - True
      - False