Add KITTI tracking checkpoints

kitti/isee/encoder_step100000.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fc5c6d8bcd0905b229912b418672df6af7bb0cf1fb7b096af45a7cca1d365d80
+size 122785543

kitti/isee/encoder_step22500.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:63df28f02365e655d990cef77e780ffaaacd06a67c0328a937c5eb115bd7bf18
+size 122785479

kitti/isee/settings.yaml

@@ -0,0 +1,214 @@
+optimizer:
+  name: Adam
+  lr: 0.0004
+  lr_scheduler:
+    name: exp_decay_with_warmup
+    warmup_steps: 2500
+    decay_steps: 100000
+model:
+  initializer:
+    name: FixedLearnedInit
+    n_slots: 15
+    dim: 64
+  encoder:
+    backbone:
+      name: NoPEDinoExtractor
+      registry_name: nope_dinov2_s
+      checkpoint_path: /group/awm/pyext/dinosaw_ckpt/nope_coco_dv2_vits14_reg4.pth
+      feature: vit_output
+      frozen: true
+    output_transform:
+      name: networks.two_layer_mlp
+      inp_dim: 384
+      outp_dim: 64
+      hidden_dim: 768
+      layer_norm: true
+    post_pos_embed:
+      name: ConcatPatchPosEmbed
+      n_patches: 729
+      mode: fourier
+      bands: 1
+      include_raw: false
+      norm_features: false
+  grouper:
+    name: SlotAttention
+    inp_dim: 68
+    slot_dim: 64
+    n_iters: 2
+    use_mlp: true
+    stream_split: 60
+    inp_split: 64
+    pos_logit_gain: 1.0
+    pos_norm: rms
+    pos_mlp: false
+    pos_init: identity
+  decoder:
+    name: CanonMLPDecoder
+    inp_dim: 64
+    outp_dim: 384
+    hidden_dims:
+    - 1024
+    - 1024
+    - 1024
+    n_patches: 729
+    pose_dims:
+    - 60
+    - 64
+    content_dims:
+    - 0
+    - 60
+    pose_hidden: 64
+    scale_range:
+    - 0.01
+    - 4.0
+    abs_table: false
+    h_init: default
+    h_hidden_dims: []
+    fourier_bands: 8
+    delta: 5.0
+    pos_gain: 1.0
+    gate_mode: none
+  dynamics_predictor: null
+  predictor:
+    name: networks.SplitSlotPredictor
+    dim: 64
+    content_dims:
+    - 0
+    - 60
+    pose_dims:
+    - 60
+    - 64
+    appearance_blocks: 1
+    appearance_heads: 4
+    kinematics_blocks: 1
+    kinematics_heads: 4
+    content_cond_dim: 8
+    detach_content_cond: true
+  target_encoder: null
+  latent_processor:
+    first_step_corrector_args:
+      n_iters: 3
+  mask_resizers: null
+  losses:
+    loss_featrec:
+      name: MSELoss
+      pred_dims:
+      - 0
+      - 384
+    loss_ss:
+      name: Slot_Slot_Contrastive_Loss
+      pred_key: processor.state
+      temperature: 0.1
+      batch_contrast: true
+      patch_inputs: false
+      keep_input_dim: true
+      dims:
+      - 0
+      - 60
+  loss_weights:
+    loss_featrec: 1.0
+    loss_ss: 0.5
+  input_type: video
+  target_type: features
+  target_encoder_input: null
+  encoder_input: null
+  decoder_source: state
+  aux_decoders: null
+  pointmap: null
+  cross_view: null
+  occlusion: null
+  anchor_supervision: null
+  visualize: false
+  eval_mode_config: null
+  visualize_every_n_steps: 10000
+  masks_to_visualize: null
+  load_weights: null
+  modules_to_load: null
+dataset:
+  name: OOMemmapDataModule
+  video_path: /group/awm/datasets/kitti_video_ram384x1280/video_u8.npy
+  index_json: /group/awm/datasets/kitti_video_ram384x1280/index.json
+  dataset_root: /group/awm/datasets/kitti_video_ram384x1280
+  train_split: train
+  val_split: val
+  segmentations:
+    kind: none
+  batch_size: 32
+  chunk_size: 8
+  frame_stride: 1
+  val_chunk_size: 24
+  val_batch_size: 4
+  num_workers: 16
+  num_val_workers: 4
+  samples_per_clip: 32
+  train_transforms:
+    name: movi_train
+    type: video
+    input_size: 378
+    window_crop:
+      resize:
+      - 378
+      - 1260
+      x_starts:
+      - 0
+      - 294
+      - 588
+      - 882
+    use_movi_normalization: false
+    h_flip_prob: 0.5
+  val_transforms:
+    name: movi_val
+    type: video
+    input_size: 378
+    window_crop:
+      resize:
+      - 378
+      - 1260
+      x_starts:
+      - 441
+    use_movi_normalization: false
+trainer:
+  precision: bf16
+  max_steps: 100000
+  log_every_n_steps: 250
+  val_check_interval: 1000
+  gradient_clip_val: 0.05
+  accelerator: auto
+train_metrics: null
+val_metrics: null
+activation_reg_schedule: null
+depth_beta_schedule: null
+globals:
+  NUM_SLOTS: 15
+  SLOT_DIM: 64
+  DINO_MODEL: vit_small_patch14_dinov2
+  FEAT_DIM: 384
+  INPUT_SIZE: 378
+  NUM_PATCHES: 729
+  NUM_GPUS: 1
+  BATCH_SIZE_PER_GPU: 32
+  TOTAL_BATCH_SIZE: 32
+  BASE_LR: 0.0001
+  CROP_RESIZE:
+  - 378
+  - 1260
+  CROP_X_STARTS:
+  - 0
+  - 294
+  - 588
+  - 882
+  VAL_CROP_X_STARTS:
+  - 441
+  POS_DIM: 4
+  BANDS: 1
+  CONTENT_LO: 0
+  CONTENT_HI: 60
+  POSE_LO: 60
+  POSE_HI: 64
+experiment_name: permanence_kitti_ours_100k
+experiment_group: slotcontrast
+seed: 20260826
+checkpoint_every_n_steps: 2500
+checkpoint_save_top_k: -1
+checkpoint_best: null
+wandb: null

kitti/isee/walker_seed0/config.json

@@ -0,0 +1,80 @@
+{
+ "ckpt": "22k5",
+ "arm": "kwv_blank",
+ "seed": 0,
+ "run_name": "",
+ "device": "cuda",
+ "threads": 6,
+ "steps": 800,
+ "batch": 32,
+ "lr": 0.001,
+ "lr_floor": 0.05,
+ "warmup": 50,
+ "clip": 1.0,
+ "ema": 0.995,
+ "win": 11,
+ "dmax": 0.15,
+ "eval_every": 100,
+ "resume_every": 100,
+ "stats_every": 5,
+ "log_every": 10,
+ "bench": 0,
+ "max_train": 0,
+ "no_eval": false,
+ "run": "kwv_blank_22k5_seed0",
+ "host": "hala",
+ "src": "s22500",
+ "pstar": "/home/pramish_paudel/codes/moog/outputs/kitti/_shared/walk72/kw/s0/ceiling/pstar27_ours22k5.npy",
+ "pstar_fine": "/home/pramish_paudel/codes/moog/outputs/kitti/_shared/walk72/kw/s0/ceiling/pstarfine105_ours22k5.npy",
+ "hp": {
+  "feat_dim": 64,
+  "d": 32,
+  "mem": true,
+  "win": 11,
+  "scale_init": 20.0,
+  "stay_init": 5.0,
+  "sigma_init": 0.05,
+  "sigma_min": 0.02,
+  "hold_lag": 1,
+  "ckpt_steps": true
+ },
+ "n_params": 59835,
+ "train_counts": {
+  "src": "s22500",
+  "split": "train",
+  "arm": "kwv_blank",
+  "n_specs": 1954,
+  "classes": [
+   "car"
+  ],
+  "n_rows": 1954,
+  "n_stretches": 49,
+  "n_tiles": 46,
+  "replaced_cells_median": 129.5,
+  "rows_with_zero_replaced": 0,
+  "seconds": 2.0
+ },
+ "eval_counts": {
+  "occ": {
+   "src": "s22500",
+   "split": "sval",
+   "arm": "kw_occ",
+   "n_find": 118,
+   "n_read": 118,
+   "n_no_j_dropped": 1,
+   "n_rows": 117,
+   "seconds": 0.3
+  }
+ },
+ "train_rows_masked_by_dmax": 3,
+ "train_rows_masked_by_dmax_share": 0.0015,
+ "pins": {
+  "sandbox/grid_walk/gw.py": "214ea446",
+  "sandbox/kitti_walk/gate_matcher.py": "4086ef64",
+  "scripts/kitti/kitti_track_dump.py": "995c5480",
+  "scripts/kitti/kitti_tilescore69.py": "323ffaa7"
+ },
+ "label_policy": "stretches selected with train-half MOTS visible masks; features, poses and targets label-free",
+ "torch": "2.13.0+cu130",
+ "omp": "6"
+}

kitti/isee/walker_seed0/ema_step800.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3bfe31eb1deee4b04c89a44e4b62be799de2557b0863e4a6426a89fd6d8b6738
+size 354337

kitti/isee/walker_seed0/walker_sd.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:09404cbcde68373bbbabe4ceb41b4adb0fc66d4f921dd14de469d3e4997cca66
+size 351644

kitti/slotcontrast/settings.yaml

@@ -0,0 +1,167 @@
+optimizer:
+  name: Adam
+  lr: 0.0004
+  lr_scheduler:
+    name: exp_decay_with_warmup
+    warmup_steps: 2500
+    decay_steps: 100000
+model:
+  initializer:
+    name: FixedLearnedInit
+    n_slots: 15
+    dim: 64
+  encoder:
+    backbone:
+      name: TimmExtractor
+      model: vit_small_patch14_dinov2
+      features:
+      - vit_block12
+      frozen: true
+      pretrained: true
+      model_kwargs:
+        dynamic_img_size: true
+    output_transform:
+      name: networks.two_layer_mlp
+      inp_dim: 384
+      outp_dim: 64
+      hidden_dim: 768
+      layer_norm: true
+  grouper:
+    name: SlotAttention
+    inp_dim: 64
+    slot_dim: 64
+    n_iters: 2
+    use_mlp: true
+  decoder:
+    name: MLPDecoder
+    inp_dim: 64
+    outp_dim: 384
+    hidden_dims:
+    - 1024
+    - 1024
+    - 1024
+    n_patches: 729
+  dynamics_predictor: null
+  predictor:
+    name: networks.TransformerEncoder
+    dim: 64
+    n_blocks: 1
+    n_heads: 4
+  target_encoder: null
+  latent_processor:
+    first_step_corrector_args:
+      n_iters: 3
+  mask_resizers: null
+  losses:
+    loss_featrec:
+      name: MSELoss
+      pred_dims:
+      - 0
+      - 384
+    loss_ss:
+      name: Slot_Slot_Contrastive_Loss
+      pred_key: processor.state
+      temperature: 0.1
+      batch_contrast: true
+      patch_inputs: false
+      keep_input_dim: true
+  loss_weights:
+    loss_featrec: 1.0
+    loss_ss: 0.5
+  input_type: video
+  target_type: features
+  target_encoder_input: null
+  encoder_input: null
+  decoder_source: state
+  aux_decoders: null
+  pointmap: null
+  cross_view: null
+  occlusion: null
+  anchor_supervision: null
+  visualize: false
+  eval_mode_config: null
+  visualize_every_n_steps: 10000
+  masks_to_visualize: null
+  load_weights: null
+  modules_to_load: null
+dataset:
+  name: OOMemmapDataModule
+  video_path: /group/awm/datasets/kitti_video_ram384x1280/video_u8.npy
+  index_json: /group/awm/datasets/kitti_video_ram384x1280/index.json
+  dataset_root: /group/awm/datasets/kitti_video_ram384x1280
+  train_split: train
+  val_split: val
+  segmentations:
+    kind: none
+  batch_size: 32
+  chunk_size: 8
+  frame_stride: 1
+  val_chunk_size: 24
+  val_batch_size: 4
+  num_workers: 16
+  num_val_workers: 4
+  samples_per_clip: 32
+  train_transforms:
+    name: movi_train
+    type: video
+    input_size: 378
+    window_crop:
+      resize:
+      - 378
+      - 1260
+      x_starts:
+      - 0
+      - 294
+      - 588
+      - 882
+    use_movi_normalization: false
+    h_flip_prob: 0.5
+  val_transforms:
+    name: movi_val
+    type: video
+    input_size: 378
+    window_crop:
+      resize:
+      - 378
+      - 1260
+      x_starts:
+      - 441
+    use_movi_normalization: false
+trainer:
+  precision: bf16
+  max_steps: 100000
+  log_every_n_steps: 250
+  val_check_interval: 1000
+  gradient_clip_val: 0.05
+  accelerator: auto
+train_metrics: null
+val_metrics: null
+activation_reg_schedule: null
+depth_beta_schedule: null
+globals:
+  NUM_SLOTS: 15
+  SLOT_DIM: 64
+  DINO_MODEL: vit_small_patch14_dinov2
+  FEAT_DIM: 384
+  INPUT_SIZE: 378
+  NUM_PATCHES: 729
+  NUM_GPUS: 1
+  BATCH_SIZE_PER_GPU: 32
+  BASE_LR: 0.0001
+  CROP_RESIZE:
+  - 378
+  - 1260
+  CROP_X_STARTS:
+  - 0
+  - 294
+  - 588
+  - 882
+  VAL_CROP_X_STARTS:
+  - 441
+experiment_name: permanence_kitti_sc_100k
+experiment_group: slotcontrast
+seed: 20260826
+checkpoint_every_n_steps: 2500
+checkpoint_save_top_k: -1
+checkpoint_best: null
+wandb: null

kitti/slotcontrast/step100000.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1833600404f6c947ae1e380dd7a7a9bcbe84eb6b5436ce4e64258b44aaafddf8
+size 125250749

kitti/slotcontrast/step22500.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:520614b004d3310fa0a607c6675f80bc78ec029dcce1c8fbc7b4e0b27b4c1932
+size 125250749