configs/pascal_voc_config.yaml

# Pascal VOC Training Configuration - Native Resolution (Variable Resolution with FluxResizer)
# Uses FLUX.1-dev and FluxResizer for optimal native resolutions (no padding)

# Experiment settings
experiment:
  name: "pascal_voc_native_resolution"
  description: "Pascal VOC with FLUX.1-dev at native resolutions using FluxResizer"

# Concept configurations (same as original)
concepts:
  basic: ["object", "background", "detail", "edges"]
  meta: ["living", "vehicle", "furniture", "object", "background"]
  meta_expanded: ["living", "vehicle", "furniture", "object", "background", "person", "car", "airplane", "table", "plant", "bird", "bicycle"]
  combined: ["object", "background", "living", "vehicle", "furniture", "detail", "edges"]
  with_classes: ["object", "background", "detail", "edges", "airplane", "bicycle", "bird", "boat", "bottle", "bus", "car", "cat", "chair", "cow", "table", "dog", "horse", "motorbike", "person", "plant", "sheep", "sofa", "train", "television"]
  all_comprehensive: ["living", "vehicle", "furniture", "object", "background", "detail", "edges", "airplane", "bicycle", "bird", "boat", "bottle", "bus", "car", "cat", "chair", "cow", "table", "dog", "horse", "motorbike", "person", "plant", "sheep", "sofa", "train", "television"]

# Training settings
training:
  concept_config: "basic"  # Which concept configuration to use
  epochs: 100  # Match baseline (was 20)
  learning_rate: 1e-4  # Match baseline (capped at 3e-5 in optimizer code)
  batch_size: 1  # REQUIRED: Must be 1 for variable resolution
  accumulate_grad_batches: 4  # Effective batch size = 4
  gradient_clip_val: 0.5
  precision: "16-mixed"
  
# Data settings
data:
  dataset: "pascal_voc"
  data_root: "${VOC_ROOT}"                       # dir holding JPEGImages/
  train_split_file: "${VOC_SPLITS}/train_aug.txt"
  val_split_file: "${VOC_SPLITS}/val.txt"
  mask_root: "${VOC_MASKS}"                      # dir holding SegmentationClassAug/ (masks)
  num_classes: 21
  # NOTE: No target_size! FluxResizer selects optimal resolution per image
  
# Model settings
model:
  flux_model: "black-forest-labs/FLUX.1-dev"  # Changed from schnell to dev
  dtype: "float16"
  dino_model: "dinov3_vitb16"  # DINOv3 base model
  feature_locations:
    transformer_blocks: [4, 9, 13, 18]
    single_transformer_blocks: [4, 15, 26, 37]
  decoder:
    features: 256
  hyperfeature_fusion:
    num_timesteps: 4
    fusion_type: "transformer"
    hidden_dim: 768
    num_transformer_layers: 3
    layer_scale_init: 1e-6
    
# FLUX settings
flux:
  timesteps: 28
  guidance_scale: 3.5
  num_inference_steps: 1
  concept_timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27]
  concept_layers: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17]

# Hardware settings
hardware:
  devices: 1  # Use 2 GPUs for faster training
  #strategy: "ddp_find_unused_parameters_true"  # Multi-GPU strategy
  num_sanity_val_steps: 1
  detect_anomaly: true
  accelerator: "gpu"

# Paths (timestamps will be automatically added)
paths:
  # All outputs live under ${MMDIFF_OUTPUT} (set it to a writable dir, e.g. /work/<user>/mmdiff_out).
  cache_base_dir: "${MMDIFF_OUTPUT}/cache"
  log_base_dir: "${MMDIFF_OUTPUT}/logs"
  checkpoint_base_dir: "${MMDIFF_OUTPUT}/checkpoints"
  use_timestamp: true
  permanent_cache_dir: "${MMDIFF_OUTPUT}/cache/pascal_feature_cache"  # Native resolution feature cache
  
# Logging
logging:
  log_every_n_steps: 10
  save_top_k: 3
  monitor: "val_loss"
  mode: "min"

# HuggingFace
huggingface:
  token: ""  # Leave empty and authenticate via `huggingface-cli login` or the HF_TOKEN env var