eps_ibp_mnist.yaml

# @package _global_

# to execute this experiment run:
# python train.py experiment=example

defaults:
  - override /datamodule: mnist.yaml
  - override /model: eps_ibp_mnist_bn.yaml
  - override /callbacks: default.yaml
  - override /trainer: default.yaml

seed: 100

trainer:
  devices: 1
  accelerator: gpu
  strategy: null
  auto_select_gpus: true

  min_epochs: 80
  max_epochs: 100

  gradient_clip_val: 10.0
  gradient_clip_algorithm: norm
  track_grad_norm: 2.0
  check_val_every_n_epoch: 1

datamodule:
  batch_size: 128
  num_workers: 8

model:
  kappa: 0.0
  init_method: ibp
  num_stn_epochs: 1
  num_reg_epochs: 20

  weight_decay_l1: 0.0
  min_eps_for_regularizers: 1.e-6
  reg_clip_with_robust: True
  reg_tight_weight: 0.5
  reg_relu_weight: 0.5

  optimizer:
    name: Adam
    amsgrad: false
    lr: 1.e-3   # 5.e-4   # 0.001
    weight_decay: 0.0

  lr_scheduler:
    name: OneCycleLR
    max_lr: 1.e-3
    # total_steps: 78200
    epochs: 100
    steps_per_epoch: 469
    pct_start: 0.3
    anneal_strategy: linear          # cos
    div_factor: 25.0              # init_lr = max_lr / div_factor
    final_div_factor: 100.      # min_lr = max_lr / final_div_factor
    three_phase: True
    interval: step   # step

  # eps test
  eps_test: [0.3, 0.4]
  # epsilon finding in train step
  max_eps: 0.4
  xtol: 1.e-6
  rtol: 1.e-4
  maxiter: 10
  ftol: 1.e-5
  
  # loss function
  standard_loss: CrossEntropy
  standard_label_smoothing: 0.0
  robust_loss: CrossEntropy
  robust_label_smoothing: 0.0

callbacks:
  model_checkpoint:
    monitor: "val/acceps"
    dirpath: ${paths.output_dir}/../../checkpoints/kappa=${model.kappa}
    # filename: "epoch={epoch:04d}-step={step}-val_acceps={val/acceps:.4f}"
    filename: "best_acceps"
    mode: "max"
    save_last: True
    auto_insert_metric_name: False
    verbose: True
    save_top_k: 1

  early_stopping:
    monitor: "val/acc"
    patience: 50
    mode: "max"

tags: ["mnist", "cnn7",]

task_name: "train/eps_ibp_mnist"