sweep-tests-Copy2.yaml

# ==============================================================================
# W&B smoke test: one EG-PCS-Net training run
# ==============================================================================
program: egpcs.cli.train
project: PCS_ET_v22
command:
  - python
  - -m
  - ${program}
  - ${args}

method: grid
metric:
  name: max_accuracy_validation
  goal: maximize

parameters:
  # ---------------------------------------------------------------------------
  # A) System / I/O
  # ---------------------------------------------------------------------------
  cuda:
    value: true
  cuda_id:
    value: 1
  multi_gpu:
    value: False
  gpu_ids:
    value: "0,1"

  seed:
    values: [912]
    #values: [187, 492, 678, 844, 998, 705, 912   ]
  log_wandb:
    value: true
  log_console:
    value: true

  dataset:
    value: images/printart/subjectivesafety_images/
  gaze_root:
    value: survey_eye_tracker/Eyetracker_attention_maps/864x508
  model_dir:
    value: models
  comparisons:
    value: comparisons_df.pickle # comparisons_df.pickle #comparisons_df_with_synthetic_berlin.pickle #comparisons_df.pickle
  resume:
    value: false
  epoch:
    value: 1
  resume_checkpoint:
    value: ""

  # ---------------------------------------------------------------------------
  # B) Dataset / Filtering
  # ---------------------------------------------------------------------------
  cities:
    value: "berlin" # , paris, munich, barcelona, london_uk_collideoscope, london_uk_gov"

  # ---------------------------------------------------------------------------
  # C) Model family / Objective
  # ---------------------------------------------------------------------------
  model:
    value: multitask_gaze
  ties:
    value: False
  # ---------------------------------------------------------------------------
  # D) Backbone & Architecture
  # ---------------------------------------------------------------------------
  backbone:
    values:
      #- "vit_small"
      - "dinov3_vitb16"
      #- "deit3_base_patch16_224"
      #- "vit_base_patch16_224"
      #- "vit_base_patch16_clip_224"
      #- "dinov2_reg_base"
      #- "vgg"
      #- "convnext_base"
      #- "efficientnet_v2_s"
      #- "regnet_y_8gf"
    

  pooling:
    value: cls
      
  pool_k:
    value: 50

  # ---------------------------------------------------------------------------
  # E) Fine-tuning policy
  # ---------------------------------------------------------------------------
  finetune:
    value: True
    
  num_ft_layers:
    values: [4]

  # ---------------------------------------------------------------------------
  # F) Batch / Training length
  # ---------------------------------------------------------------------------
  batch_size:
    value: 16
  k:
    value: 1
    
  # One epoch is enough to verify the complete training path.
  max_epochs:
    value: 5
  grad_clip:
    value: 0.0

  # ---------------------------------------------------------------------------
  # G) Optimizer
  # ---------------------------------------------------------------------------
  base_lr:
    value: 3.5e-5
  backbone_lr_scale:
    value: 1
  weight_decay:
    value: 0.05
  backbone_freeze_epochs:
    value: 0
    
  rank_dropout:
    value: 0.3
  cross_dropout:
    value: 0.3

  # ---------------------------------------------------------------------------
  # H) Ranking / Loss weights
  # ---------------------------------------------------------------------------
  ranking_margin:
    values: [5.6]
  rank_w:
    value: 1

  ties_w:
    value: 0.0
  attn_w:
    value: 3.5

  label_smoothing:
    value: 0.0
  use_class_weights:
    value: false
#  class_weights:
#    value: "[1, 10, 1]"



  # ---------------------------------------------------------------------------
  # I) Scheduler
  # ---------------------------------------------------------------------------
  scheduler:
    value: warmup_cosine
  warmup_frac:
    value: 0.15
  eta_min:
    value: 5.0e-6

  # ---------------------------------------------------------------------------
  # J) Augmentation
  # ---------------------------------------------------------------------------
  augment:
    values: ["heavy"]
  use_seg:
    value: false

  # ---------------------------------------------------------------------------
  # K) Accuracy reporting
  # ---------------------------------------------------------------------------
  full_accuracy:
    value: false

  # ---------------------------------------------------------------------------
  # L) Early stopping
  # ---------------------------------------------------------------------------
  early_stop:
    value: true
  early_stop_metric:
    value: loss_validation
  early_stop_mode:
    value: min
  early_stop_patience:
    value: 3
  early_stop_min_delta:
    value: 0.0
  early_stop_start_epoch:
    value: 1

  # ---------------------------------------------------------------------------
  # Gaze Configuration
  # ---------------------------------------------------------------------------
  # A single value keeps this grid sweep to exactly one run.
  model_variant:
    value: EG-PCS-Net

  gaze_align_target:
    value: attention  #patch_tokens

  attention_mode:
    values: ["raw"]

  attn_layer:
    values: [-1]

  gaze_map_size:
    value: auto

  #train_gaze_frac:
  #  values: [1]