per_node_sigmoid_tcg.yaml

# @package _global_
#
# to execute this experiment run:
# python train.py experiment=tcg

defaults:
  - override /model: per_node_linear_tcg #hyper_tcg
  - override /datamodule: linear_unidentifiable_velocity #linear_velocity
  - override /logger:
      - csv
      - wandb
  - override /trainer: gpu
name: "per_node_sigmoid_tcg_gfn"

seed: 0

datamodule:
  batch_size: 500 #500
  T: 2
  p: 20
  vars_to_deidentify: [0, 1, 2]
  sparsity: 0.9 # 0.9 --> 1024 Nodes for p=20 and [0,1,2]
  system: "sigmoid_linear"
  sigma: 0
  burn_in: 1
  seed: 13

model:
  env_batch_size: 1024
  eval_batch_size: 5000
  full_posterior_eval: False
  uniform_backwards: True
  debug_use_shd_energy: False
  analytic_use_simple_mse_energy: True
  loss_fn: "detailed_balance"
  alpha: 0
  temperature: 0.01
  temper_period: 5
  prior_lambda: 150
  beta: 0.01
  confidence: 0.0
  hidden_dim: 128
  gfn_freq: 1
  energy_freq: 1
  pretraining_epochs: 0
  lr: 1e-4
  hyper: "mlp"
  bias: True

trainer:
  max_epochs: 1000
  min_epochs: 1000
  check_val_every_n_epoch: 5

logger:
  wandb:
    tags:
      ["kl", "analytic", "sigmoid", "per-node", "gfn", "${name}", "v_final_3"]