linear_sigmoid_svgd.yaml

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

defaults:
  - override /model: bayesian_linear_velocity
  - override /datamodule: linear_unidentifiable_velocity.yaml
  - override /logger:
      - wandb
      - csv
  - override /trainer: gpu
name: "linear_sigmoid_svgd"

seed: 13

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

# best
model:
  lr: 1e-3
  alpha: 1e-4
  svgd_reg: 0 # 1e5
  l1_reg: 10 # 1e-1
  l2_reg: 0
  kl_reg: 0 # 1e-5
  svgd_gamma: 3000 # -1 to set to med
  temperature: 0.01
  n_ens: 1024
  eval_batch_size: 1024
  k_hidden: 20
  hyper: "linear"
  hyper_hidden_dim: [64, 64, 64]
  bias: True
  svgd: True

  optimizer: "adam"

trainer:
  max_epochs: 1000
  check_val_every_n_epoch: 5

logger:
  wandb:
    tags: ["kl", "analytic", "sigmoid", "svgd", "${name}", "v_final_3"]