Add LA-CATER, static camera checkpoints

lacater_static/ablations/isee_dino/settings.yaml

@@ -0,0 +1,187 @@
+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: 11
+    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
+    post_pos_embed:
+      name: ConcatPatchPosEmbed
+      n_patches: 1369
+      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: 1369
+    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: true
+  eval_mode_config: null
+  visualize_every_n_steps: 2500
+  masks_to_visualize: null
+  load_weights: null
+  modules_to_load: null
+dataset:
+  name: LaCaterVideoDataModule
+  root: /group/awm/datasets/la_cater
+  batch_size: 32
+  chunk_size: 8
+  frame_stride: 2
+  val_chunk_size: 24
+  val_batch_size: 4
+  val_max_clips: 250
+  num_workers: 16
+  num_val_workers: 4
+  samples_per_clip: 8
+  train_transforms:
+    name: movi_train
+    type: video
+    input_size: 518
+    use_movi_normalization: false
+    h_flip_prob: 0.5
+  val_transforms:
+    name: movi_val
+    type: video
+    input_size: 518
+    use_movi_normalization: false
+    num_classes: 1
+trainer:
+  precision: bf16
+  max_steps: 100000
+  log_every_n_steps: 500
+  val_check_interval: 2500
+  gradient_clip_val: 0.05
+  accelerator: auto
+train_metrics: null
+val_metrics: null
+activation_reg_schedule: null
+depth_beta_schedule: null
+globals:
+  NUM_SLOTS: 11
+  SLOT_DIM: 64
+  DINO_MODEL: vit_small_patch14_dinov2
+  FEAT_DIM: 384
+  POS_DIM: 4
+  BANDS: 1
+  INPUT_SIZE: 518
+  NUM_PATCHES: 1369
+  NUM_GPUS: 1
+  BATCH_SIZE_PER_GPU: 32
+  BASE_LR: 0.0001
+  CONTENT_LO: 0
+  CONTENT_HI: 60
+  POSE_LO: 60
+  POSE_HI: 64
+experiment_name: permanence_lacater_static_ours_dino
+experiment_group: slotcontrast
+seed: 20260826
+checkpoint_every_n_steps: 1000
+checkpoint_save_top_k: -1
+checkpoint_best: null
+wandb: null

lacater_static/ablations/isee_dino/step100000.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3a50c6a98b80bdc0af571c7a69013565cc9b92d2b7b322861d41e590d397f723
+size 125037707

lacater_static/ablations/isee_noapp/settings.yaml

@@ -0,0 +1,185 @@
+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: 11
+    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: 1369
+      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: 1369
+    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
+    content_cond: false
+  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: true
+  eval_mode_config: null
+  visualize_every_n_steps: 2500
+  masks_to_visualize: null
+  load_weights: null
+  modules_to_load: null
+dataset:
+  name: LaCaterVideoDataModule
+  root: /group/awm/datasets/la_cater
+  batch_size: 32
+  chunk_size: 8
+  frame_stride: 2
+  val_chunk_size: 24
+  val_batch_size: 4
+  val_max_clips: 250
+  num_workers: 16
+  num_val_workers: 4
+  samples_per_clip: 8
+  train_transforms:
+    name: movi_train
+    type: video
+    input_size: 518
+    use_movi_normalization: false
+    h_flip_prob: 0.5
+  val_transforms:
+    name: movi_val
+    type: video
+    input_size: 518
+    use_movi_normalization: false
+    num_classes: 1
+trainer:
+  precision: bf16
+  max_steps: 100000
+  log_every_n_steps: 500
+  val_check_interval: 2500
+  gradient_clip_val: 0.05
+  accelerator: auto
+train_metrics: null
+val_metrics: null
+activation_reg_schedule: null
+depth_beta_schedule: null
+globals:
+  NUM_SLOTS: 11
+  SLOT_DIM: 64
+  DINO_MODEL: vit_small_patch14_dinov2
+  FEAT_DIM: 384
+  POS_DIM: 4
+  BANDS: 1
+  INPUT_SIZE: 518
+  NUM_PATCHES: 1369
+  NUM_GPUS: 1
+  BATCH_SIZE_PER_GPU: 32
+  BASE_LR: 0.0001
+  CONTENT_LO: 0
+  CONTENT_HI: 60
+  POSE_LO: 60
+  POSE_HI: 64
+experiment_name: permanence_lacater_static_ours_noapp
+experiment_group: slotcontrast
+seed: 20260826
+checkpoint_every_n_steps: 1000
+checkpoint_save_top_k: -1
+checkpoint_best: null
+wandb: null

lacater_static/ablations/isee_noapp/step100000.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5a8a3f9d099095935a7637292e1b35a7f35e5bad73e02ab9220ce9034bac5623
+size 122935059

lacater_static/ablations/isee_ssall/settings.yaml

@@ -0,0 +1,184 @@
+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: 11
+    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: 1369
+      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: 1369
+    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
+      - 64
+  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: true
+  eval_mode_config: null
+  visualize_every_n_steps: 2500
+  masks_to_visualize: null
+  load_weights: null
+  modules_to_load: null
+dataset:
+  name: LaCaterVideoDataModule
+  root: /group/awm/datasets/la_cater
+  batch_size: 32
+  chunk_size: 8
+  frame_stride: 2
+  val_chunk_size: 24
+  val_batch_size: 4
+  val_max_clips: 250
+  num_workers: 16
+  num_val_workers: 4
+  samples_per_clip: 8
+  train_transforms:
+    name: movi_train
+    type: video
+    input_size: 518
+    use_movi_normalization: false
+    h_flip_prob: 0.5
+  val_transforms:
+    name: movi_val
+    type: video
+    input_size: 518
+    use_movi_normalization: false
+    num_classes: 1
+trainer:
+  precision: bf16
+  max_steps: 100000
+  log_every_n_steps: 500
+  val_check_interval: 2500
+  gradient_clip_val: 0.05
+  accelerator: auto
+train_metrics: null
+val_metrics: null
+activation_reg_schedule: null
+depth_beta_schedule: null
+globals:
+  NUM_SLOTS: 11
+  SLOT_DIM: 64
+  DINO_MODEL: vit_small_patch14_dinov2
+  FEAT_DIM: 384
+  POS_DIM: 4
+  BANDS: 1
+  INPUT_SIZE: 518
+  NUM_PATCHES: 1369
+  NUM_GPUS: 1
+  BATCH_SIZE_PER_GPU: 32
+  BASE_LR: 0.0001
+  CONTENT_LO: 0
+  CONTENT_HI: 60
+  POSE_LO: 60
+  POSE_HI: 64
+experiment_name: permanence_lacater_static_ours_ssall
+experiment_group: slotcontrast
+seed: 20260826
+checkpoint_every_n_steps: 1000
+checkpoint_save_top_k: -1
+checkpoint_best: null
+wandb: null

lacater_static/ablations/isee_ssall/step100000.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cb841ab72cd5ee4fc7b2ca3274298ac5866601cf9cd1798b986a917b03dd2632
+size 122940487

lacater_static/ablations/pre4_twostream/settings.yaml

@@ -0,0 +1,180 @@
+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: 11
+    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: 1369
+      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: none
+    pos_mlp: false
+    pos_init: identity
+  decoder:
+    name: CanonMLPDecoder
+    inp_dim: 64
+    outp_dim: 384
+    hidden_dims:
+    - 1024
+    - 1024
+    - 1024
+    n_patches: 1369
+    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
+  aux_decoders: null
+  pointmap: null
+  cross_view: null
+  anchor_supervision: null
+  visualize: true
+  eval_mode_config: null
+  visualize_every_n_steps: 2500
+  masks_to_visualize: null
+  load_weights: null
+  modules_to_load: null
+dataset:
+  name: LaCaterVideoDataModule
+  root: /group/awm/datasets/la_cater
+  batch_size: 32
+  chunk_size: 8
+  frame_stride: 2
+  val_chunk_size: 24
+  val_batch_size: 4
+  val_max_clips: 250
+  num_workers: 16
+  num_val_workers: 4
+  samples_per_clip: 8
+  train_transforms:
+    name: movi_train
+    type: video
+    input_size: 518
+    use_movi_normalization: false
+    h_flip_prob: 0.5
+  val_transforms:
+    name: movi_val
+    type: video
+    input_size: 518
+    use_movi_normalization: false
+    num_classes: 1
+trainer:
+  precision: bf16
+  max_steps: 100000
+  log_every_n_steps: 500
+  val_check_interval: 2500
+  gradient_clip_val: 0.05
+  accelerator: auto
+train_metrics: null
+val_metrics: null
+activation_reg_schedule: null
+depth_beta_schedule: null
+globals:
+  NUM_SLOTS: 11
+  SLOT_DIM: 64
+  DINO_MODEL: vit_small_patch14_dinov2
+  FEAT_DIM: 384
+  POS_DIM: 4
+  BANDS: 1
+  INPUT_SIZE: 518
+  NUM_PATCHES: 1369
+  NUM_GPUS: 1
+  BATCH_SIZE_PER_GPU: 32
+  BASE_LR: 0.0001
+  CONTENT_LO: 0
+  CONTENT_HI: 60
+  POSE_LO: 60
+  POSE_HI: 64
+experiment_name: lacater_pre4_twostream_100k_t8
+experiment_group: slotcontrast
+seed: 20260826
+checkpoint_every_n_steps: 1000
+checkpoint_save_top_k: -1

lacater_static/ablations/pre4_twostream/step100000.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:87034c9fa15a4329f8b12d880869e3df82a71add200307c34a993297d12459b0
+size 122946311

lacater_static/isee/encoder_step100000.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c8f10542eb7a7970fbc029cb1eda90a39e1ad595601dda9348d96b09d36e4e9b
+size 122946311

lacater_static/isee/settings.yaml

@@ -0,0 +1,184 @@
+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: 11
+    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: 1369
+      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: 1369
+    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: true
+  eval_mode_config: null
+  visualize_every_n_steps: 2500
+  masks_to_visualize: null
+  load_weights: null
+  modules_to_load: null
+dataset:
+  name: LaCaterVideoDataModule
+  root: /group/awm/datasets/la_cater
+  batch_size: 32
+  chunk_size: 8
+  frame_stride: 2
+  val_chunk_size: 24
+  val_batch_size: 4
+  val_max_clips: 250
+  num_workers: 16
+  num_val_workers: 4
+  samples_per_clip: 8
+  train_transforms:
+    name: movi_train
+    type: video
+    input_size: 518
+    use_movi_normalization: false
+    h_flip_prob: 0.5
+  val_transforms:
+    name: movi_val
+    type: video
+    input_size: 518
+    use_movi_normalization: false
+    num_classes: 1
+trainer:
+  precision: bf16
+  max_steps: 100000
+  log_every_n_steps: 500
+  val_check_interval: 2500
+  gradient_clip_val: 0.05
+  accelerator: auto
+train_metrics: null
+val_metrics: null
+activation_reg_schedule: null
+depth_beta_schedule: null
+globals:
+  NUM_SLOTS: 11
+  SLOT_DIM: 64
+  DINO_MODEL: vit_small_patch14_dinov2
+  FEAT_DIM: 384
+  POS_DIM: 4
+  BANDS: 1
+  INPUT_SIZE: 518
+  NUM_PATCHES: 1369
+  NUM_GPUS: 1
+  BATCH_SIZE_PER_GPU: 32
+  BASE_LR: 0.0001
+  CONTENT_LO: 0
+  CONTENT_HI: 60
+  POSE_LO: 60
+  POSE_HI: 64
+experiment_name: permanence_lacater_static_ours
+experiment_group: slotcontrast
+seed: 20260826
+checkpoint_every_n_steps: 1000
+checkpoint_save_top_k: -1
+checkpoint_best: null
+wandb: null

lacater_static/isee/walker_s0_args.json

@@ -0,0 +1,45 @@
+{
+ "run_name": "gws_loc_s0",
+ "runs": "/group/awm/slotcontrast/grid_walk_runs_latchev",
+ "feat": "/group/awm/slotcontrast/cache/grid_walk/train_slocked_f64",
+ "slots": "/group/awm/slotcontrast/cache/latch_evidence/walkcache_static_loc",
+ "feat_extra": "",
+ "device": "auto",
+ "threads": 0,
+ "seed": 0,
+ "resume": true,
+ "d": 32,
+ "no_mem": false,
+ "win": 11,
+ "scale_init": 20.0,
+ "stay_init": 5.0,
+ "sigma_init": 0.05,
+ "sigma_min": 0.02,
+ "hold_lag": 1,
+ "no_ckpt": false,
+ "ckpt_all": false,
+ "stats_every": 5,
+ "target_lag": 1,
+ "len_coef": 1.0,
+ "dmax": 0.35,
+ "pseudo": 8,
+ "pseudo_coef": 1.0,
+ "pseudo_min": 1,
+ "pseudo_max": 8,
+ "moved_boost": 4.0,
+ "steps": 4000,
+ "batch": 16,
+ "crop": 0,
+ "lr": 0.001,
+ "lr_floor": 0.05,
+ "lr_scalars": 0.0,
+ "warmup": 100,
+ "clip": 1.0,
+ "ema": 0.998,
+ "val_clips": 100,
+ "val_batch": 8,
+ "val_every": 500,
+ "save_every": 500,
+ "log_every": 25,
+ "bench": 0
+}

lacater_static/isee/walker_s0_step2000.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8893a3195fa0f4c29f1d5a1236509a1981039805d3f171eb1c94d54a21749c7e
+size 1381743

lacater_static/isee/walker_s1_args.json

@@ -0,0 +1,45 @@
+{
+ "run_name": "gws_loc_s1",
+ "runs": "/group/awm/slotcontrast/grid_walk_runs_latchev",
+ "feat": "/group/awm/slotcontrast/cache/grid_walk/train_slocked_f64",
+ "slots": "/group/awm/slotcontrast/cache/latch_evidence/walkcache_static_loc",
+ "feat_extra": "",
+ "device": "auto",
+ "threads": 0,
+ "seed": 1,
+ "resume": true,
+ "d": 32,
+ "no_mem": false,
+ "win": 11,
+ "scale_init": 20.0,
+ "stay_init": 5.0,
+ "sigma_init": 0.05,
+ "sigma_min": 0.02,
+ "hold_lag": 1,
+ "no_ckpt": false,
+ "ckpt_all": false,
+ "stats_every": 5,
+ "target_lag": 1,
+ "len_coef": 1.0,
+ "dmax": 0.35,
+ "pseudo": 8,
+ "pseudo_coef": 1.0,
+ "pseudo_min": 1,
+ "pseudo_max": 8,
+ "moved_boost": 4.0,
+ "steps": 4000,
+ "batch": 16,
+ "crop": 0,
+ "lr": 0.001,
+ "lr_floor": 0.05,
+ "lr_scalars": 0.0,
+ "warmup": 100,
+ "clip": 1.0,
+ "ema": 0.998,
+ "val_clips": 100,
+ "val_batch": 8,
+ "val_every": 500,
+ "save_every": 500,
+ "log_every": 25,
+ "bench": 0
+}

lacater_static/isee/walker_s1_step2000.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0b3e16bcb8879dc2a5dcb3627e0f93e81c27e51b39fd15bf79552b1807a3ffae
+size 1381832

lacater_static/isee/walker_s2_args.json

@@ -0,0 +1,45 @@
+{
+ "run_name": "gws_loc_s2",
+ "runs": "/group/awm/slotcontrast/grid_walk_runs_latchev",
+ "feat": "/group/awm/slotcontrast/cache/grid_walk/train_slocked_f64",
+ "slots": "/group/awm/slotcontrast/cache/latch_evidence/walkcache_static_loc",
+ "feat_extra": "",
+ "device": "auto",
+ "threads": 0,
+ "seed": 2,
+ "resume": true,
+ "d": 32,
+ "no_mem": false,
+ "win": 11,
+ "scale_init": 20.0,
+ "stay_init": 5.0,
+ "sigma_init": 0.05,
+ "sigma_min": 0.02,
+ "hold_lag": 1,
+ "no_ckpt": false,
+ "ckpt_all": false,
+ "stats_every": 5,
+ "target_lag": 1,
+ "len_coef": 1.0,
+ "dmax": 0.35,
+ "pseudo": 8,
+ "pseudo_coef": 1.0,
+ "pseudo_min": 1,
+ "pseudo_max": 8,
+ "moved_boost": 4.0,
+ "steps": 4000,
+ "batch": 16,
+ "crop": 0,
+ "lr": 0.001,
+ "lr_floor": 0.05,
+ "lr_scalars": 0.0,
+ "warmup": 100,
+ "clip": 1.0,
+ "ema": 0.998,
+ "val_clips": 100,
+ "val_batch": 8,
+ "val_every": 500,
+ "save_every": 500,
+ "log_every": 25,
+ "bench": 0
+}

lacater_static/isee/walker_s2_step2000.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:197363ebcd5801221038a8ca750bdaff0bd9609591178377c9223f693bdc8af5
+size 1381743

lacater_static/isee_fig1a/walker_vis_s0_args.json

@@ -0,0 +1,45 @@
+{
+ "run_name": "gws_vis_s0",
+ "runs": "/group/awm/slotcontrast/grid_walk_runs_slocked",
+ "feat": "/group/awm/slotcontrast/cache/grid_walk/train_slocked_f64",
+ "slots": "/group/awm/slotcontrast/cache/lacater57/perm_full9300",
+ "feat_extra": "",
+ "device": "auto",
+ "threads": 0,
+ "seed": 0,
+ "resume": true,
+ "d": 32,
+ "no_mem": false,
+ "win": 11,
+ "scale_init": 20.0,
+ "stay_init": 5.0,
+ "sigma_init": 0.05,
+ "sigma_min": 0.02,
+ "hold_lag": 1,
+ "no_ckpt": false,
+ "ckpt_all": false,
+ "stats_every": 5,
+ "target_lag": 1,
+ "len_coef": 1.0,
+ "dmax": 0.35,
+ "pseudo": 8,
+ "pseudo_coef": 1.0,
+ "pseudo_min": 1,
+ "pseudo_max": 8,
+ "moved_boost": 4.0,
+ "steps": 4000,
+ "batch": 16,
+ "crop": 0,
+ "lr": 0.001,
+ "lr_floor": 0.05,
+ "lr_scalars": 0.0,
+ "warmup": 100,
+ "clip": 1.0,
+ "ema": 0.998,
+ "val_clips": 100,
+ "val_batch": 8,
+ "val_every": 500,
+ "save_every": 500,
+ "log_every": 25,
+ "bench": 0
+}

lacater_static/isee_fig1a/walker_vis_s0_step2000.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:96a51f4c59b08ec59e81a3cc35b9b64442e2f95d235cffcfd34f200840db733f
+size 1381832

lacater_static/isee_fig1a/walker_vis_s1_args.json

@@ -0,0 +1,45 @@
+{
+ "run_name": "gws_vis_s1",
+ "runs": "/group/awm/slotcontrast/grid_walk_runs_slocked",
+ "feat": "/group/awm/slotcontrast/cache/grid_walk/train_slocked_f64",
+ "slots": "/group/awm/slotcontrast/cache/lacater57/perm_full9300",
+ "feat_extra": "",
+ "device": "auto",
+ "threads": 0,
+ "seed": 1,
+ "resume": true,
+ "d": 32,
+ "no_mem": false,
+ "win": 11,
+ "scale_init": 20.0,
+ "stay_init": 5.0,
+ "sigma_init": 0.05,
+ "sigma_min": 0.02,
+ "hold_lag": 1,
+ "no_ckpt": false,
+ "ckpt_all": false,
+ "stats_every": 5,
+ "target_lag": 1,
+ "len_coef": 1.0,
+ "dmax": 0.35,
+ "pseudo": 8,
+ "pseudo_coef": 1.0,
+ "pseudo_min": 1,
+ "pseudo_max": 8,
+ "moved_boost": 4.0,
+ "steps": 4000,
+ "batch": 16,
+ "crop": 0,
+ "lr": 0.001,
+ "lr_floor": 0.05,
+ "lr_scalars": 0.0,
+ "warmup": 100,
+ "clip": 1.0,
+ "ema": 0.998,
+ "val_clips": 100,
+ "val_batch": 8,
+ "val_every": 500,
+ "save_every": 500,
+ "log_every": 25,
+ "bench": 0
+}

lacater_static/isee_fig1a/walker_vis_s1_step2000.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7e35c49e235a8df768086fda14f82aca326daf750fd877a40459960f8db7802f
+size 1381832

lacater_static/isee_fig1a/walker_vis_s2_args.json

@@ -0,0 +1,45 @@
+{
+ "run_name": "gws_vis_s2",
+ "runs": "/group/awm/slotcontrast/grid_walk_runs_slocked",
+ "feat": "/group/awm/slotcontrast/cache/grid_walk/train_slocked_f64",
+ "slots": "/group/awm/slotcontrast/cache/lacater57/perm_full9300",
+ "feat_extra": "",
+ "device": "auto",
+ "threads": 0,
+ "seed": 2,
+ "resume": true,
+ "d": 32,
+ "no_mem": false,
+ "win": 11,
+ "scale_init": 20.0,
+ "stay_init": 5.0,
+ "sigma_init": 0.05,
+ "sigma_min": 0.02,
+ "hold_lag": 1,
+ "no_ckpt": false,
+ "ckpt_all": false,
+ "stats_every": 5,
+ "target_lag": 1,
+ "len_coef": 1.0,
+ "dmax": 0.35,
+ "pseudo": 8,
+ "pseudo_coef": 1.0,
+ "pseudo_min": 1,
+ "pseudo_max": 8,
+ "moved_boost": 4.0,
+ "steps": 4000,
+ "batch": 16,
+ "crop": 0,
+ "lr": 0.001,
+ "lr_floor": 0.05,
+ "lr_scalars": 0.0,
+ "warmup": 100,
+ "clip": 1.0,
+ "ema": 0.998,
+ "val_clips": 100,
+ "val_batch": 8,
+ "val_every": 500,
+ "save_every": 500,
+ "log_every": 25,
+ "bench": 0
+}

lacater_static/isee_fig1a/walker_vis_s2_step2000.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1aec9cd81116483f0bb0563b611afd7bf983f65ffb6c5dd0f26e42c6d461eeb5
+size 1381832

lacater_static/loci_looped/cfg.json

@@ -0,0 +1,84 @@
+{
+ "model_path": "lacater_level2",
+ "datatype": "lacater",
+ "dataset": "LACATER",
+ "learning_rate": 3.333e-05,
+ "num_updates": 800000,
+ "max_epochs": 1000,
+ "max_updates": 1150000,
+ "phases": {
+  "start_inner_loop": 60000,
+  "shufleslots_end": 30000,
+  "entity_pretraining_phase2_end": 50000,
+  "entity_pretraining_phase1_end": 30000,
+  "background_pretraining_end": 0
+ },
+ "defaults": {
+  "num_workers": 8,
+  "prefetch_factor": 2,
+  "statistics_offset": 10,
+  "load_optimizers": false,
+  "teacher_forcing": 10,
+  "skip_frames": 1,
+  "error_dropout": 0.1,
+  "lacater_stride": 2,
+  "seed": 20260918,
+  "save_every": 5000
+ },
+ "bptt": {
+  "bptt_start_timestep": 0,
+  "bptt_steps": 2,
+  "bptt_steps_max": 2,
+  "increase_bptt_steps_every": 200000
+ },
+ "model": {
+  "level": 2,
+  "batch_size": 16,
+  "num_objects": 10,
+  "img_channels": 3,
+  "input_size": [
+   240,
+   320
+  ],
+  "latent_size": [
+   15,
+   20
+  ],
+  "gestalt_size": 96,
+  "bottleneck": "binar",
+  "position_regularizer": 0.01,
+  "time_regularizer": 0.1,
+  "encoder_regularizer": 0.333333,
+  "inner_loop_enabled": true,
+  "encoder": {
+   "channels": 48,
+   "level1_channels": 24,
+   "num_layers": 3,
+   "reg_lambda": 1e-10
+  },
+  "predictor": {
+   "heads": 2,
+   "layers": 2,
+   "channels_multiplier": 2,
+   "reg_lambda": 1e-10,
+   "transformer_type": "standard"
+  },
+  "decoder": {
+   "channels": 48,
+   "level1_channels": 24,
+   "num_layers": 5
+  },
+  "background": {
+   "learning_rate": 0.0001,
+   "learning_rate_old": 0.0001,
+   "num_layers": 1,
+   "latent_channels": 48,
+   "level1_channels": 24,
+   "gestalt_size": 8,
+   "flow": false
+  },
+  "update_module": {
+   "reg_lambda": 5e-06
+  }
+ }
+}

lacater_static/loci_looped/net_900000.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:924387eee96d33fc0b290893c66ac73d464248896def5d4e70163e7a4ccfcb22
+size 33536941

lacater_static/randsfq/0042.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d84462b977e01f3a08deebe1eb87e6ef0f7401fe719449a58c57e8d4d1a2d92a
+size 139678903

lacater_static/randsfq/config.py

@@ -0,0 +1,293 @@
+"""`permanence_randsfq_lacater_static` -- RandSF.Q (Zhao, Li, Kannala, Pajarinen, AAAI 2026,
+arXiv:2508.01345) trained from scratch on LA-CATER, STATIC camera, for T1. Not upstream: this repo,
+2026-09-15. Its moving-camera twin is `permanence_randsfq_lacater_moving.py`, which differs in
+`data_root` alone (asserted by tests/test_randsfq_lacater.py).
+
+PROTOCOL (user decision 2026-09-15, "match ours" = slotcontrast/configs/slotcontrast/lacater.yaml):
+518 px input -> DINOv2 ViT-S/14 -> 37x37 = 1,369 tokens; K = 11; 8-frame windows at frame stride 2
+drawn from the 300 annotated frames; ImageNet normalisation; horizontal flip p 0.5 at train with the
+whole window flipped together. EVERYTHING ELSE is the authors' MOVi-C recipe
+`config-randsfq/randsfq_c-movi_c.py`. The complete delta from it:
+
+| key | recipe | here | why |
+|---|---|---|---|
+| resolut0 | [256, 256] | [518, 518] | protocol |
+| resolut1 | [16, 16] | [37, 37] | 518 // 14; sizes the decoder's learnt position embedding |
+| encode_backbone.modules[0] | Interpolate x0.875 bicubic | Identity | 518 goes straight into the ViT (its native img_size); x0.875 would feed it 453 px. Identity keeps the ViT at index 1, so parameter names match the recipe |
+| DINO2ViT.in_size | int(256 * 0.875) = 224 | 518 | same |
+| dataset_t / dataset_v | MOVi lmdb + segment + bbox | LaCater(root, split, clip_len 8, stride 2) | the dataset |
+| transform0 | StridedRandomSliceSequence(size=6) | window drawn inside LaCater | protocol: 8 frames at stride 2 from [0, 300) |
+| samples_per_clip | (one window per clip per epoch) | 1 | same semantics -> 1,163 steps/epoch vs MOVi-C's 1,218, so train.py's epoch-granular val and SaveModel cadence match the recipe |
+| dataset_v | MOVi-C val, 250 clips, full 24 frames | first 250 val clips, one 8x2 window from frame 0 | loss-only val at the train window shape; a 300-frame clip at 518 px is not a val batch |
+| transform_t RandomCrop(size=None, scale=[.75, 1]) | present | REMOVED | on a 4:3 frame its area is 56-75% of the image, i.e. never the full frame; ours has no crop; T1 needs the full frame |
+| transform_v CenterCrop(size=None) | present | REMOVED | the max square drops the left/right 40 px of every 320x240 frame, where the snitch can be |
+| Resize keys / size | video + segment, 256 | video, 518 (bilinear, antialias, as recipe) | no segment. `[!]` 320x240 -> 518x518 is an ANISOTROPIC squash, exactly as ours (`movi_train` Resize) |
+| RandomFlip / Normalize keys | video + segment / video | video / video | no segment |
+| collate_fn_t / _v | ClPadToMax1(segment) + ClPadTo1(bbox) + DefaultCollate | DefaultCollate | no segment / bbox |
+| initializ | MLP(in_dim=4) on GT frame-0 boxes (`condit="batch.bbox"`) | NormalShared(num=11, dim=256) | `[!]` the `c-` recipes are CONDITIONAL on ground-truth boxes, which LA-CATER does not have per object; NormalShared is the authors' own unconditional init (`randsfq_r-ytvis.py`, "# >NormalSeparat") |
+| model_imap | input + condit | input | same |
+| transit.dt | 6 | 8 | the authors set dt = training window length in BOTH released recipes (MOVi 6/6, YTVIS 5/5) |
+| decode.num_keep_fixed | (literal int(256 * 0.1) - 1 = 24) | int(1369 * 0.1) - 1 = 135 | the literal is ~10% of the recipe's 256 tokens; at 1369 it would be 1.8% (model/dias.py) |
+| num_work | 4 | 8 | loader only, no effect on the objective: live MJPEG decode + 518 px resize measured 55 windows/s at 4 workers vs 118 at 8 (0.144 vs 0.068 s per batch of 8, static train, n=40 batches); an H200 at 2-3x the A6000's measured 0.51 s/step compute would wait on 4 |
+| acc_fn_t / acc_fn_v | mBO / ARI, ARI-fg, mBO, mIoU | {} / {} | LA-CATER has no masks: training and validation are loss-only |
+| after_forward | argmax decoder attention -> one-hot `output.segment` | removed | it only fed those metrics (and is ~2 GB of transient tensors at 518 px) |
+| before_step .cuda() keys | video, segment, bbox | video | no segment / bbox |
+
+UNCHANGED, deliberately (the recipe): emb_dim 256, vfm_dim 384, backbone
+`vit_small_patch14_reg4_dinov2.lvd142m` (DINOv2 ViT-S/14 WITH 4 register tokens -- ours uses the
+register-free `vit_small_patch14_dinov2`), frozen; encode_project MLP; SlotAttention 3 iters on
+frame 0 / 1 after; RSFQTransit nhead 4, expansion 4, dropout 0.5; decoder 4 layers, 4 heads;
+Adam lr 2e-4 / 4 = 5e-5, linear warm-up total_step // 20 then cosine to lr / 1e3; grad-norm clip
+0.05; GradScaler + autocast; loss = recon MSE + 0.5 * SlotContrast; total_step 50,000; batch 8;
+val_interval total_step // 40; SaveModel every val epoch past 50%; seed via train.py.
+
+`[!]` Run it through `scripts/randsfq/permanence_randsfq_lacater_train.sh {static,moving} <seed>`,
+which refuses gcp nodes (reading /group there starves the loader; stage to /scratch first) and turns
+on train.py's opt-in resume with the cross-cluster claim in `scripts/randsfq/randsfq_claim.sh`.
+"""
+
+from object_centric_bench.datum import (
+    Resize,
+    RandomFlip,
+    Normalize,
+    Lambda,
+    LaCater,
+    DefaultCollate,
+)
+from object_centric_bench.learn import (
+    Adam,
+    GradScaler,
+    ClipGradNorm,
+    MSELoss,
+    SlotContrastLoss,
+    CbLinearCosine,
+    Callback,
+    AverageLog,
+    SaveModel,
+)
+from object_centric_bench.model import (
+    RandSFQ,
+    Sequential,
+    DINO2ViT,
+    Identity,
+    MLP,
+    SlotAttention,
+    NormalShared,
+    RSFQTransit,
+    ARRandTransformerDecoder,
+    LearntPositionalEmbedding,
+    Linear,
+    LayerNorm,
+    TransformerDecoder,
+    TransformerDecoderLayer,
+)
+from object_centric_bench.util import Compose, ComposeNoStar
+
+### global
+
+max_num = 10 + 1  # K = 11: protocol, and the recipe's own value
+resolut0 = [518, 518]  # CHANGED 256 -> 518 (protocol)
+resolut1 = [37, 37]  # CHANGED 16 -> 37 (518 // 14)
+emb_dim = 256
+vfm_dim = 384
+
+total_step = 50000  # 100000 better
+val_interval = total_step // 40
+batch_size_t = 32 // 4  # 64 better
+batch_size_v = batch_size_t
+num_work = 8  # CHANGED 4 -> 8: live MJPEG decode; measured 55 windows/s at 4 workers, 118 at 8
+lr = 2e-4 / 4  # scale with batch_size
+
+### datum
+
+data_root = "la_cater"  # joined onto train.py's --data_dir (/group/awm/datasets)
+clip_len = 8  # CHANGED (protocol): 8-frame windows
+stride = 2  # CHANGED (protocol): frame stride 2 -> a window spans 15 of the 300 annotated frames
+
+IMAGENET_MEAN = [[[123.675]], [[116.28]], [[103.53]]]
+IMAGENET_STD = [[[58.395]], [[57.12]], [[57.375]]]
+transform_t = [
+    dict(type=Resize, keys=["video"], size=resolut0, interp="bilinear"),
+    dict(type=RandomFlip, keys=["video"], dims=[-1], p=0.5),
+    dict(type=Normalize, keys=["video"], mean=[IMAGENET_MEAN], std=[IMAGENET_STD]),
+]
+transform_v = [
+    dict(type=Resize, keys=["video"], size=resolut0, interp="bilinear"),
+    dict(type=Normalize, keys=["video"], mean=[IMAGENET_MEAN], std=[IMAGENET_STD]),
+]
+dataset_t = dict(
+    type=LaCater,
+    root=data_root,
+    split="train",
+    clip_len=clip_len,
+    stride=stride,
+    samples_per_clip=1,
+    train=True,
+    seed=0,
+    transform=dict(type=Compose, transforms=transform_t),
+    base_dir=...,
+)
+dataset_v = dict(
+    type=LaCater,
+    root=data_root,
+    split="val",
+    clip_len=clip_len,
+    stride=stride,
+    samples_per_clip=1,
+    train=False,
+    start=0,
+    max_clips=250,
+    transform=dict(type=Compose, transforms=transform_v),
+    base_dir=...,
+)
+collate_fn_t = dict(
+    type=ComposeNoStar,
+    transforms=[
+        dict(type=DefaultCollate),
+    ],
+)
+collate_fn_v = collate_fn_t
+
+### model
+
+model = dict(
+    type=RandSFQ,
+    encode_backbone=dict(
+        type=Sequential,
+        modules=[
+            dict(type=Identity),  # CHANGED: recipe Interpolate(scale_factor=0.875, bicubic)
+            dict(
+                type=DINO2ViT,
+                model_name="vit_small_patch14_reg4_dinov2.lvd142m",
+                in_size=resolut0[0],  # CHANGED: recipe int(resolut0[0] * 0.875)
+                rearrange=True,
+                norm_out=False,
+            ),
+        ],
+    ),
+    encode_posit_embed=dict(type=Identity),
+    encode_project=dict(
+        type=MLP, in_dim=vfm_dim, dims=[vfm_dim, vfm_dim], ln="pre", dropout=0.0
+    ),
+    # CHANGED: recipe MLP(in_dim=4) on GT boxes; LA-CATER has none -> the authors' unconditional init
+    initializ=dict(type=NormalShared, num=max_num, dim=emb_dim),
+    aggregat=dict(
+        type=SlotAttention,
+        num_iter=3,  # 3 > 2, 1
+        embed_dim=emb_dim,
+        ffn_dim=emb_dim * 4,
+        dropout=0,
+        kv_dim=vfm_dim,
+        trunc_bp=None,  # > bi-level
+    ),
+    transit=dict(
+        type=RSFQTransit,
+        dt=clip_len,  # CHANGED 6 -> 8: dt = training window length, as in both released recipes
+        ci=vfm_dim,
+        c=emb_dim,
+        nhead=4,
+        expanz=4,
+        pdo=0.5,  # XXX XXX XXX
+        norm_first=False,
+        bias=False,
+    ),
+    decode=dict(
+        type=ARRandTransformerDecoder,
+        emb_dim=vfm_dim,
+        posit_embed=dict(
+            type=LearntPositionalEmbedding,
+            resolut=[resolut1[0] * resolut1[1]],
+            embed_dim=vfm_dim,
+        ),
+        project1=dict(  # fc>fc+ln
+            type=Sequential,
+            modules=[
+                dict(
+                    type=Linear, in_features=vfm_dim, out_features=vfm_dim, bias=False
+                ),
+                dict(type=LayerNorm, normalized_shape=vfm_dim),
+            ],
+        ),
+        project2=dict(  # fc+ln>fc
+            type=Sequential,
+            modules=[
+                dict(
+                    type=Linear, in_features=emb_dim, out_features=vfm_dim, bias=False
+                ),
+                dict(type=LayerNorm, normalized_shape=vfm_dim),
+            ],
+        ),
+        backbone=dict(
+            type=TransformerDecoder,
+            decoder_layer=dict(
+                type=TransformerDecoderLayer,
+                d_model=vfm_dim,
+                nhead=4,  # 4@384, 6@768
+                dim_feedforward=vfm_dim * 4,
+                dropout=0.0,
+                activation="gelu",
+                batch_first=True,
+                norm_first=True,
+                bias=False,
+            ),
+            num_layers=4,
+        ),
+        readout=dict(type=Identity),
+        # CHANGED: the recipe's literal int(256 * 0.1) - 1 is ~10% of ITS 256 tokens
+        num_keep_fixed=int(resolut1[0] * resolut1[1] * 0.1) - 1,
+    ),
+)
+model_imap = dict(input="batch.video")  # CHANGED: no condit (no GT boxes)
+model_omap = ["feature", "slotz", "attenta", "recon", "attentd"]
+ckpt_map = []  # target<-source
+freez = [r"^m\.encode_backbone\..*"]
+
+### learn
+
+param_groups = None
+optimiz = dict(type=Adam, params=param_groups, lr=lr)
+gscale = dict(type=GradScaler)
+gclip = dict(type=ClipGradNorm, max_norm=0.05)
+
+loss_fn_t = loss_fn_v = dict(
+    recon=dict(
+        metric=dict(type=MSELoss),
+        map=dict(input="output.recon", target="output.feature"),
+        transform=dict(type=Lambda, ikeys=[["target"]], func=lambda _: _.detach()),
+    ),
+    ssc=dict(
+        metric=dict(type=SlotContrastLoss),
+        map=dict(input="output.slotz"),
+        weight=0.5,
+    ),
+)
+# CHANGED: LA-CATER has no segmentation masks -> no mBO / ARI / mIoU; train and val are loss-only
+acc_fn_t = dict()
+acc_fn_v = dict()
+
+before_step = [
+    dict(
+        type=Lambda,
+        ikeys=[["batch.video"]],  # CHANGED: no batch.segment / batch.bbox
+        func=lambda _: _.cuda(),
+    ),
+    dict(
+        type=CbLinearCosine,
+        assigns=["optimiz.param_groups[0]['lr']=value"],
+        nlin=total_step // 20,
+        ntotal=total_step,
+        vstart=0,
+        vbase=lr,
+        vfinal=lr / 1e3,
+    ),
+]
+# CHANGED: no after_forward (it only built `output.segment` for the removed metrics)
+callback_t = [
+    dict(type=Callback, before_step=before_step),
+    dict(type=AverageLog, log_file=...),
+]
+callback_v = [
+    dict(type=Callback, before_step=before_step[:1]),
+    callback_t[1],
+    dict(type=SaveModel, save_dir=..., since_step=total_step * 0.5),
+]

lacater_static/slotcontrast/settings.yaml

@@ -0,0 +1,138 @@
+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: 11
+    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: 1369
+  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: true
+  eval_mode_config: null
+  visualize_every_n_steps: 2500
+  masks_to_visualize: null
+  load_weights: null
+  modules_to_load: null
+dataset:
+  name: LaCaterVideoDataModule
+  root: /group/awm/datasets/la_cater
+  batch_size: 32
+  chunk_size: 8
+  frame_stride: 2
+  val_chunk_size: 24
+  val_batch_size: 4
+  val_max_clips: 250
+  num_workers: 16
+  num_val_workers: 4
+  samples_per_clip: 8
+  train_transforms:
+    name: movi_train
+    type: video
+    input_size: 518
+    use_movi_normalization: false
+    h_flip_prob: 0.5
+  val_transforms:
+    name: movi_val
+    type: video
+    input_size: 518
+    use_movi_normalization: false
+    num_classes: 1
+trainer:
+  precision: bf16
+  max_steps: 100000
+  log_every_n_steps: 500
+  val_check_interval: 2500
+  gradient_clip_val: 0.05
+  accelerator: auto
+train_metrics: null
+val_metrics: null
+activation_reg_schedule: null
+depth_beta_schedule: null
+globals:
+  NUM_SLOTS: 11
+  SLOT_DIM: 64
+  DINO_MODEL: vit_small_patch14_dinov2
+  FEAT_DIM: 384
+  INPUT_SIZE: 518
+  NUM_PATCHES: 1369
+  NUM_GPUS: 1
+  BATCH_SIZE_PER_GPU: 32
+  BASE_LR: 0.0001
+experiment_name: lacater_base_100k_t8
+experiment_group: slotcontrast
+seed: 20260826
+checkpoint_every_n_steps: 1000
+checkpoint_save_top_k: -1
+checkpoint_best: null
+wandb: null

lacater_static/slotcontrast/step100000.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9f3c15fb7715f31f0dca9118b9bdf7608ab09a1030634f97ea4fa92def53b1ca
+size 125739709