{
  "root": "/root/private/data/pallet",
  "annotation_sha256": "efa4bb2469e87c971527f3e1273c1394acaac51e1c127831b433b8b9dbcb1f0a",
  "code_sha256": "e17deec53e29d6eb1e9a2ce9c8f7483aadf06a22293a4dfaf0bb85c570d178c7",
  "protocol_sha256": "43e90b43366b704014165e65652f30ca8819c1ac2de0b7e49e5bc0bcd7ea9b7a",
  "protocol": "/root/private/project/pallet_detection_v2/outputs/protocol_v5",
  "split_seed": 42,
  "subset_sha256": "8cf1d9c88c4a422e5c6a951e26f016b749ce404fe529d3447df7a07ccf8252d3",
  "cache_index_sha256": "10db4abcbdbdb64452c616f9849ab783bcd7197d63a6dcf027ed452c4f4ef29e",
  "pretrained": false,
  "family": "rfdetr-small",
  "recipe": "official_research_constrained",
  "holdout_domain": null,
  "train_groups": [
    "2023_10_7/scene/plastic/395mm/1900mm",
    "2023_11_1/scene/plastic/393mm/1900mm",
    "2023_11_1/scene/wood/239mm/2200mm",
    "2023_11_1/scene/wood/393mm/1900mm",
    "2023_7_4/scene/185mm/1700mm/101",
    "2023_7_4/scene/235mm/2000mm/110",
    "2024_11_6/scene/pallet_4_test_2",
    "2025_3_18/scene/3",
    "2025_4_1/scene/3",
    "2025_4_22/scene/199"
  ],
  "val_groups": [
    "2023_10_8/scene/plastic/388mm/1900mm",
    "2023_7_4/scene/300mm/2000mm/101",
    "2024_10_26/scene/1",
    "2024_11_6/scene/pallet_3_test_2"
  ],
  "normalization": {
    "depth_scale": 1.0,
    "intensity_scale": 6.113261437416077
  },
  "seed": 42,
  "p2": false,
  "fusion": false,
  "edge": false,
  "augmentation": true,
  "mosaic": false,
  "size": 640,
  "batch_size": 4,
  "effective_batch": 16,
  "epochs": 2,
  "device_uuid": "3315c096-fef4-7350-483a-7c5f5ca2b619",
  "use_depth": true,
  "use_intensity": true,
  "cohort": "all",
  "recipe_spec": {
    "schema_version": 1,
    "family": "rfdetr-small",
    "recipe": "official_research_constrained",
    "training_status": "runnable_sensor_research_integration",
    "optimizer_preparation_status": "runnable",
    "sources": [
      {
        "package": "rfdetr",
        "version": "1.11.0",
        "file": "config.py",
        "sha256": "2162bf18e38057753aa59c5c93d26e8b39210587fe3be99d045ad3c2ebe2bb8b"
      },
      {
        "package": "rfdetr",
        "version": "1.11.0",
        "file": "training/param_groups.py",
        "sha256": "52e1bb979e95ba1908b912d4c6351342097dbe3a602555a944d4ef8b42462338"
      },
      {
        "package": "rfdetr",
        "version": "1.11.0",
        "file": "models/backbone/backbone.py",
        "sha256": "6dfb3003bab137aba181abe5e035d011654eca4879082088ffe4072c9e129499"
      },
      {
        "package": "rfdetr",
        "version": "1.11.0",
        "file": "training/module_model.py",
        "sha256": "48e57e6106e2e8a2fe09d876d2e237c9e1ed08756274890357bd74cc0a07c265"
      },
      {
        "package": "rfdetr",
        "version": "1.11.0",
        "file": "training/callbacks/ema.py",
        "sha256": "180a374169911b7331f7ebe0caadc26caa68b06cf6aed0d9357a6ef16bea2e5e"
      },
      {
        "package": "rfdetr",
        "version": "1.11.0",
        "file": "training/trainer.py",
        "sha256": "b274fb04e609decd0d3b2a86532b9391ab2763cc4c42581968823a6a5358da07"
      },
      {
        "package": "rfdetr",
        "version": "1.11.0",
        "file": "datasets/coco.py",
        "sha256": "e3b6443ae8b4b0707643aa4de6f7d5115ffe82b1f1658b2afcf7990100303a28"
      },
      {
        "package": "rfdetr",
        "version": "1.11.0",
        "file": "datasets/transforms.py",
        "sha256": "6e4138baa14fce2b33d5b616c5ecc585a47ab0eae721c2af2564b483ce059134"
      }
    ],
    "official_defaults": {
      "train": {
        "lr": 0.0001,
        "lr_encoder": 0.00015,
        "batch_size": 4,
        "grad_accum_steps": 1,
        "eval_batch_size": null,
        "pad_targets_to": null,
        "auto_batch_target_effective": 16,
        "auto_batch_max_targets_per_image": 100,
        "auto_batch_ema_headroom": 0.7,
        "epochs": 100,
        "resume": null,
        "ema_decay": 0.993,
        "ema_tau": 100,
        "checkpoint_interval": 10,
        "skip_best_epochs": 0,
        "smooth_alpha": 0.0,
        "warmup_epochs": 0.0,
        "lr_vit_layer_decay": 0.8,
        "lr_component_decay": 0.7,
        "drop_path": 0.0,
        "cls_loss_coef": 1.0,
        "keypoint_flip_pairs": [],
        "keypoint_l1_loss_coef": 0.0,
        "keypoint_findable_loss_coef": 0.0,
        "keypoint_visible_loss_coef": 0.0,
        "keypoint_nll_loss_coef": 0.0,
        "keypoint_oks_sigmas": null,
        "dataset_file": "roboflow",
        "square_resize_div_64": true,
        "dataset_dir": "/root/private/project/pallet_detection_v2/<dataset_dir>",
        "output_dir": "output",
        "multi_scale": "per-batch",
        "expanded_scales": true,
        "use_ema": true,
        "ema_update_interval": 1,
        "eval_base_model": false,
        "eval_ema_only": false,
        "num_workers": 2,
        "weight_decay": 0.0001,
        "amp_dtype": "auto",
        "best_model_metric": "map",
        "early_stopping": false,
        "early_stopping_patience": 10,
        "early_stopping_min_delta": 0.001,
        "early_stopping_use_ema": false,
        "progress_bar": null,
        "tensorboard": true,
        "wandb": false,
        "mlflow": false,
        "clearml": false,
        "project": null,
        "run": null,
        "class_names": null,
        "run_test": false,
        "eval_max_dets": 500,
        "eval_interval": 1,
        "log_per_class_metrics": false,
        "eval_backend": "vernier",
        "eval_masks_head_resolution": false,
        "aug_config": null,
        "scale_jitter": true,
        "augmentation_backend": "cpu",
        "save_dataset_grids": false,
        "notes": null,
        "accelerator": "auto",
        "clip_max_norm": 0.1,
        "seed": null,
        "sync_bn": false,
        "strategy": "auto",
        "devices": 1,
        "num_nodes": 1,
        "fp16_eval": false,
        "lr_scheduler": "step",
        "lr_scheduler_kwargs": {},
        "lr_scheduler_interval": "step",
        "lr_scheduler_monitor": "val/loss",
        "optimizer": "adamw",
        "optimizer_kwargs": {},
        "dont_save_weights": false,
        "train_log_sync_dist": false,
        "train_log_on_step": false,
        "compact_train_metrics": true,
        "compute_train_metrics": false,
        "num_sanity_val_steps": 0,
        "compute_val_loss": "auto",
        "compute_test_loss": true,
        "pin_memory": null,
        "persistent_workers": null,
        "prefetch_factor": null,
        "pack_targets": true
      },
      "model": {
        "encoder": "dinov2_windowed_small",
        "out_feature_indexes": [
          3,
          6,
          9,
          12
        ],
        "dec_layers": 3,
        "two_stage": true,
        "projector_scale": [
          "P4"
        ],
        "hidden_dim": 256,
        "patch_size": 16,
        "num_windows": 2,
        "sa_nheads": 8,
        "ca_nheads": 16,
        "dec_n_points": 2,
        "num_queries": 300,
        "num_select": 300,
        "postprocess_trace_alpha": 0.2,
        "bbox_reparam": true,
        "lite_refpoint_refine": true,
        "layer_norm": true,
        "amp": true,
        "num_channels": 3,
        "num_classes": 90,
        "pretrain_weights": "rf-detr-small.pth",
        "device": "cuda",
        "resolution": 512,
        "group_detr": 13,
        "gradient_checkpointing": false,
        "compile": false,
        "cuda_graphs": false,
        "fused_optimizer": true,
        "positional_encoding_size": 32,
        "ia_bce_loss": true,
        "segmentation_head": false,
        "use_grouppose_keypoints": false,
        "keypoint_cross_attn": true,
        "inter_instance_kp_attn": false,
        "grouppose_keypoint_dim_downscale": 1,
        "dual_projector": false,
        "dual_projector_kp_only": false,
        "num_keypoints_per_class": [],
        "num_decoder_registers": 0,
        "mask_downsample_ratio": 4,
        "backbone_lora": false,
        "freeze_encoder": false,
        "license": "Apache-2.0",
        "model_name": null
      }
    },
    "research_overrides": {
      "model": {
        "resolution": 640
      },
      "train": {
        "batch_size": 4,
        "grad_accum_steps": 4,
        "epochs": 300,
        "early_stopping": true,
        "early_stopping_patience": 50,
        "multi_scale": "off",
        "expanded_scales": false,
        "scale_jitter": false
      },
      "protocol": {
        "mosaic": false,
        "effective_batch": 16,
        "max_epochs": 300,
        "patience": 50,
        "size": 640
      }
    },
    "optimizer": {
      "name": "AdamW",
      "lr": 0.0001,
      "lr_encoder": 0.00015,
      "lr_component_decay": 0.7,
      "lr_vit_layer_decay": 0.8,
      "weight_decay": 0.0001,
      "betas": [
        0.9,
        0.999
      ],
      "groups": {
        "other": "lr",
        "decoder": "lr * lr_component_decay",
        "encoder": "lr_encoder * lr_component_decay**2 * layer_decay",
        "encoder_decay": "zero for gamma/pos_embed/rel_pos/bias/norm/embeddings; otherwise weight_decay",
        "other_decay": "optimizer weight_decay, including head/decoder biases"
      },
      "layer_decay": "num_layers=out_feature_indexes[-1]+1; exponent=num_layers+1-layer_id; embeddings=0; .layer.k.=k+1; otherwise num_layers+1",
      "fused": "model fused_optimizer=True, gated by upstream runtime/precision eligibility"
    },
    "schedule": {
      "name": "step",
      "warmup_epochs": 0.0,
      "lr_drop": 100,
      "gamma": 0.1,
      "interval": "optimizer_step",
      "cosine_min_factor_if_selected": 0.0,
      "accumulation": "fixed16_actual_example_mean_epoch_tail"
    },
    "ema": {
      "enabled": true,
      "decay": 0.993,
      "tau": 100,
      "update_interval": 1,
      "formula": "decay * (1-exp(-updates/tau))",
      "evaluation": "selected EMA by default"
    },
    "augmentation": {
      "backend_selector": "cpu",
      "aug_config": null,
      "default_optional_transform": "torchvision horizontal flip p=.5",
      "square_resize_div_64": true,
      "scale_jitter": false,
      "multi_scale": "off",
      "expanded_scales": false,
      "resize_crop": "direct random resize or resize[400,500,600] + random crop + selected target resize",
      "mosaic": "not present in upstream RF-DETR pipeline",
      "implementation": "rf_sensor_float_flip_v1",
      "native_resize_crop": false,
      "synchronized_sensor_flip_probability": 0.5
    },
    "blockers": [],
    "constraint_caveats": [
      "RF detection loss uses its native per-microbatch box normalization; the loop weights each microbatch mean by its real examples over the accumulation window and flushes epoch tails.",
      "RF-DETR disables multi_scale/scale_jitter to keep training inputs at 640; those are research overrides.",
      "Research patience=50 does not equalize upstream min_delta or checkpoint selection semantics."
    ],
    "sensor_research_deviations": [
      "Fixed sensor letterbox with supported float/mask interpolation replaces RF square resize, random resize/crop and RGB normalization; native horizontal flip probability .5 is synchronized with boxes and masks.",
      "Scale jitter and expanded multi-scale are disabled for fixed640; one-device effective16 accumulation uses actual-example weighted epoch tails, a named deviation from Lightning's fixed divisor on partial windows.",
      "Project original-coordinate AP50_95 selects EMA strict improvement, unlike RF's separate regular/EMA tracks, smoothing/skip epochs, min_delta and checkpoint metric policy.",
      "Official AveragedModel callback buffers/counter, step scheduler, grouping, fused eligibility, precision choice and clip norm .1 are used; model sensor adapter adds base-lr/decay group.",
      "CUDA deformable-attention training is numerically nondeterministic across identical seeds in this environment; RNG/scheduler/optimizer/EMA resume state is preserved, but bit-exact GPU weights cannot be guaranteed.",
      "ONNX export supports checkpoint loading but RF near-tied query ordering in missing-mask probes can fail strict parity; report parity case explicitly."
    ],
    "project_integration_sources": [
      {
        "file": "pallet/experiments/train.py",
        "sha256": "30378d4d652204ee1f9df2bb0907af94731ab08aa3fcaee1baa300e939005f97"
      },
      {
        "file": "pallet/experiments/rfdetr_recipe.py",
        "sha256": "45c6c9404d6bfc961559f22f08a7fba2b10daf84d23c1af6dcb3dbd413790ab9"
      },
      {
        "file": "pallet/experiments/dataset.py",
        "sha256": "0bf75e1370b7f9cabf82569f5c7cab9845d471d633e7602ebab80d5f7c50be83"
      }
    ]
  },
  "source_snapshot": "/root/private/project/pallet_detection_v2/outputs/source_snapshots/03453bf914fa01a06c4fb1224c28f951b4db6f348a8024f0685c29357bb53716.tar.gz",
  "source_snapshot_id": "03453bf914fa01a06c4fb1224c28f951b4db6f348a8024f0685c29357bb53716",
  "git_commit": "40dedb423e2c8df9b0f1675011cb087108d621aa",
  "git_dirty": false
}