{
  "preprocessing_version": "1.0.0",
  "model_family": "car_body_shared",
  "supported_candidates": [
    {
      "name": "EfficientNet-B0",
      "timm_id": "efficientnet_b0",
      "role": "primary_baseline",
      "image_size": 224,
      "crop_size": 224,
      "notes": "Primary baseline. ~5.3M params, ~21 MB FP32. Mild augmentation sufficient."
    },
    {
      "name": "MobileNetV3-Large",
      "timm_id": "mobilenetv3_large_100",
      "role": "active_candidate",
      "image_size": 224,
      "crop_size": 224,
      "notes": "Lightweight CNN. ~5.5M params, ~22 MB FP32. Shares same preprocessing contract."
    },
    {
      "name": "EfficientNetV2-S",
      "timm_id": "tf_efficientnetv2_s",
      "role": "active_candidate",
      "image_size": 224,
      "crop_size": 224,
      "notes": "Improved EfficientNet. ~21.5M params, ~86 MB FP32. Within 95 MB limit."
    },
    {
      "name": "ConvNeXt-V2-Atto",
      "timm_id": "convnextv2_atto",
      "role": "active_candidate",
      "image_size": 224,
      "crop_size": 224,
      "notes": "Hybrid CNN/ViT macro-design. ~3.7M params, ~15 MB FP32. Ideal for small datasets."
    },
    {
      "name": "EfficientNet-B3",
      "timm_id": "efficientnet_b3",
      "role": "active_candidate",
      "image_size": 224,
      "crop_size": 224,
      "notes": "Larger EfficientNet. ~12M params, ~48 MB FP32. Within 95 MB limit."
    },
    {
      "name": "ResNet-50",
      "timm_id": "resnet50",
      "role": "fallback",
      "fallback_flag_required": true,
      "image_size": 224,
      "crop_size": 224,
      "notes": "Classic CNN baseline. ~25M params, ~100 MB FP32 – exceeds 95 MB; requires quantization or FP16."
    },
    {
      "name": "ConvNeXt-Tiny",
      "timm_id": "convnext_tiny",
      "role": "fallback",
      "fallback_flag_required": true,
      "image_size": 224,
      "crop_size": 224,
      "notes": "~28M params, ~112 MB FP32 – exceeds 95 MB; requires FP16 cast."
    },
    {
      "name": "Swin-Transformer-Tiny",
      "timm_id": "swin_tiny_patch4_window7_224",
      "role": "fallback",
      "fallback_flag_required": true,
      "image_size": 224,
      "crop_size": 224,
      "notes": "~28M params, ~112 MB FP32 – exceeds 95 MB; requires FP16 cast (model.half())."
    }
  ],
  "image_size": 224,
  "crop_size": 224,
  "interpolation": "bicubic",
  "interpolation_rationale": "Bicubic uses 4x4 pixel neighborhoods with cubic polynomials, preserving derivative continuity at vehicle edges (roofline, C-pillar, wheel arches). Bilinear blurs these high-frequency spatial features.",
  "normalization": {
    "type": "imagenet_zscore",
    "mean": [0.485, 0.456, 0.406],
    "std": [0.229, 0.224, 0.225],
    "channels": "RGB",
    "rationale": "ImageNet Z-Score normalization conditions the Hessian matrix (kappa(H) ≈ 1), producing a near-spherical loss landscape for stable gradient descent convergence."
  },
  "augmentation": {
    "vertical_flip_enabled": false,
    "vertical_flip_rationale": "FORBIDDEN. Creates physically impossible upside-down vehicles. Neural networks use gravity direction as implicit spatial reference; vertical flip destroys the structural hierarchy (wheels-down, roof-up) that distinguishes body types.",
    "horizontal_flip": {
      "probability": 0.5,
      "rationale": "Safe. Aerodynamic symmetry means left/right views are label-invariant. Effectively doubles dataset size at zero semantic cost."
    },
    "random_resized_crop": {
      "scale_min": 0.8,
      "scale_max": 1.0,
      "aspect_ratio_min": 0.95,
      "aspect_ratio_max": 1.05,
      "rationale": "Aspect-ratio locked to [0.95, 1.05]. Vehicle body type is defined by length-to-height ratio; extreme distortion can make a Sedan appear as a MICRO or Hatchback."
    },
    "rotation": {
      "max_degrees": 10.0,
      "padding_mode": "reflect",
      "rationale": "Simulates camera tilt (±10°). Reflection padding avoids zero-pad corner artefacts that create spurious edge features in early conv layers. Rotation >15° risks same label-invariance violation as vertical flip."
    },
    "color_jitter": {
      "brightness": 0.2,
      "contrast": 0.2,
      "saturation": 0.2,
      "hue": 0.1,
      "probability": 0.5,
      "rationale": "Breaks spurious color correlations: MICRO cars tend to have bright colors; F1/Open-Wheel cars are often red (sponsor liveries). Forces model to learn chassis geometry, not color shortcuts."
    }
  },
  "transform_contract": {
    "train": "get_train_transforms() – includes all augmentation above",
    "validation": "get_val_transforms() – deterministic: Resize + CenterCrop + ToTensor + Normalize",
    "evaluation": "get_eval_transforms() – identical to validation",
    "inference_api": "get_inference_transforms() – identical to validation; Phase 5 API imports this"
  },
  "class_count": 8,
  "class_slugs": [
    "SUV",
    "VAN",
    "STATION_WAGON",
    "MICRO",
    "ACIK_TEKERLEKLI_F1_ARACLARI",
    "SEDAN",
    "HATCHBACK",
    "PICK_UP"
  ]
}