Add training configs for released checkpoints

Add the original config.yaml files from the VAE, generator, and ControlNet runs corresponding to the released checkpoints.

experiments/auto_encoder/2025-12-19_01-23-28-525/config.yaml

@@ -0,0 +1,140 @@
+accelerator: gpu
+accumulate_grad_batches: 1
+batch_size: !!python/tuple
+- 1
+benchmark: true
+check_val_every_n_epoch: 5
+data:
+  augmentation: true
+  bbox_max: 7.25
+  bbox_rand_shift: 1.0
+  bottom_padding: 3.0
+  category_exlude: !!python/tuple
+  - floor
+  - customizedceiling
+  - extrusioncustomizedbackgroundwall
+  - wallbottom
+  - wallouter
+  - wall
+  - ceiling
+  - REMOVE
+  - void
+  - ''
+  ceiling_clip: 2.2
+  chunk_shape: !!python/tuple
+  - 256
+  - 256
+  - 256
+  data_keys: !!python/tuple
+  - tsdf_p_1.0
+  data_list: !!python/tuple
+  - 3D-FRONT
+  - ScannetPP
+  fixed_len: 20000
+  known_ratio: 0.999
+  max_depth: 100.0
+  max_length: 100
+  min_depth: 0.0
+  num_categories: 250
+  overfit_patch: false
+  overfit_scene: null
+  pos_sample: importance
+  room_height: 2.8
+  sample_one_data_key: true
+  target: dataset.dataset.Dataset
+  truncation: null
+  voxel_size: 0.011
+ema_update_batch: 5
+gradient_clip_algorithm: norm
+gradient_clip_val: 1.0
+job_name: 2025-12-19_01-23-28-525_task_reconstr_spoi_60_vsst_0.011_fact_8_chan_8_dlis_3D-FRONT_dkey_tsdf_p_1_v
+learnable_weights: false
+learning_rate: 0.0001
+log_dir: /cluster/andram/qmeng/Experiments/scene_completion/auto_encoder/2025-12-19_01-23-28-525
+log_root: /cluster/andram/qmeng/Experiments/scene_completion
+lr_scheduler:
+  name: linear
+  num_warmup_steps: 1000
+max_epochs: 400
+max_steps: -1
+metrics: !!python/tuple
+- l1
+- l2
+model:
+  channels: 8
+  ckpt_path: /cluster/andram/qmeng/Experiments/scene_completion/auto_encoder/2025-12-19_01-23-28-525/checkpoint/vxl_0_011_last.ckpt
+  decoder: models.fvdb.decoder.Decoder
+  double_z: true
+  embeding_freq: -1
+  empty_token: true
+  encoder: models.fvdb.encoder.Encoder
+  f_maps: 64
+  factor: 8
+  feed_gt_structure: false
+  gaussian_tau: null
+  geometry_weight: 100.0
+  is_add_dec: false
+  kl_weight: 100.0
+  mask_unknown: true
+  max_down_time: 6
+  name: fVDB VAE
+  num_groups: 32
+  num_res_blocks: 1
+  order: gcs
+  structure_weight: 100.0
+  target: models.fvdb.sparse_vae.SparseVAE
+  unstable_cutoff_threshold: 0.15
+  use_attention: false
+  use_checkpoint: true
+  use_residual: true
+  with_color_branch: false
+  with_semantic_branch: false
+name: AutoEncoder
+num_sanity_val_steps: 1
+num_workers: 4
+precision: '32'
+profiler: simple
+record_memory: false
+rescale_factor: !!python/tuple
+- 1
+resume: /cluster/andram/qmeng/Experiments/scene_completion/auto_encoder/2025-12-19_01-23-28-525
+seed: 23
+slurm:
+  cluster: null
+  cpus_per_task: 4
+  folder: null
+  gpus_per_node: 3
+  mem_gb: 160
+  nodes: 1
+  slurm_constraint: '[a100]'
+  slurm_exclude: tarsonis,
+  slurm_job_name: 2025-12-19_01-23-28-525_task_reconstr_spoi_60_vsst_0.011_fact_8_chan_8_dlis_3D-FRONT_dkey_tsdf_p_1_v
+  slurm_mail_type: null
+  slurm_partition: submit
+  slurm_time: 4-00:00:00
+  stderr_to_stdout: true
+  tasks_per_node: 3
+  timeout_min: 10000000000
+slurm_folder: /cluster/andram/qmeng/Experiments/scene_completion/auto_encoder/2025-12-19_01-23-28-525
+stage_points: !!python/tuple
+- 60
+strategy: ddp
+target: models.autoencoder.Net
+task:
+  export_as: !!python/tuple
+  - tree
+  - npz
+  name: reconstruction
+  num_samples: 10
+  split: test
+test_data_keys: !!python/tuple
+- tsdf_p_1.0
+test_on_train_set: false
+timestamp: 2025-12-19_01-23-28-525
+train_data_keys: !!python/tuple
+- tsdf_p_1.0
+use_wandb: true
+val_data_keys: !!python/tuple
+- tsdf_p_1.0
+val_on_train_set: false
+voxel_size_start: 0.011

experiments/auto_encoder/2025-12-19_01-23-28-525/generator/2026-02-23_16-22-25-152/config.yaml

@@ -0,0 +1,190 @@
+accelerator: gpu
+accumulate_grad_batches: 1
+ae_log: 2025-12-19_01-23-28-525
+avg_over: region
+batch_size: 16
+benchmark: true
+check_val_every_n_epoch: 5
+ckpt_path: null
+data:
+  augmentation: true
+  bbox_rand_shift: 1.0
+  bottom_padding: 3.0
+  category_exclude: !!python/tuple
+  - floor
+  - ceiling
+  - wall
+  category_include: null
+  ceiling_clip: 2.2
+  csv_path: null
+  data_list: !!python/tuple
+  - 3D-FRONT
+  - ScannetPP
+  - ARKitScenes
+  fixed_len: 20000
+  full_scene: false
+  known_ratio: 0.999
+  num_categories: 150
+  overfit_patch: false
+  overfit_scene: null
+  patch_shape: !!python/tuple
+  - 256
+  - 256
+  - 256
+  target: seen2scene.dataset.dataset.Dataset
+  truncation: null
+  use_semantic_mapping: false
+  voxel_size: 0.011
+ema_update_batch: 5
+gradient_clip_algorithm: norm
+gradient_clip_val: 1.0
+ignore_keys: null
+job_name: 2026-02-23_16-22-25-152_prec_bf16-mix_nsvs_0_alog_2025-12-_lkey_tsdf_p_1_pend_100_msdi_udro_ulau
+latent_key: tsdf_p_1.0
+learning_rate: 0.0001
+log_dir: /cluster/andram/qmeng/Experiments/scene_completion/auto_encoder/2025-12-19_01-23-28-525/generator/2026-02-23_16-22-25-152
+log_root: /cluster/andram/qmeng/Experiments/scene_completion
+lr_scheduler:
+  lr: 0.0001
+  min_lr: 1.0e-07
+  name: cosine_warmup
+  num_warmup_steps: 1000.0
+mask_unknown: true
+max_epochs: 300
+max_steps: -1
+model:
+  band_weight: 0.5
+  cond_drop_prob: 0.1
+  logit_mean: 0.0
+  logit_std: 1.0
+  mode_scale: 1.29
+  model_cfg:
+    attn_mode: full
+    compression_block_size: 2
+    compression_version: v2
+    context_dim: null
+    factor: 1.0
+    in_channels: 8
+    mlp_ratio: 4
+    model_channels: 768
+    name: SparseDiT
+    num_blocks: 28
+    num_head_channels: 32
+    num_heads: 16
+    num_kv_heads: 2
+    out_channels: 8
+    pe_mode: rope
+    qk_rms_norm: true
+    qk_rms_norm_cross: true
+    selection_block_size: 4
+    share_mod: false
+    target: seen2scene.models.sparse.dit.SparseDiT
+    topk: 8
+    use_checkpoint: true
+    use_shift: true
+    use_ssa: false
+    window_size: 4
+    zero_out: true
+  name: FM
+  num_inference_steps: 50
+  precondition_outputs: false
+  scheduler:
+    num_train_timesteps: 1000
+    target: diffusers.FlowMatchEulerDiscreteScheduler
+  target: seen2scene.models.backbone.flow_matching.Model
+  use_distance_rope: true
+  use_ema: true
+  use_label_augmentation: true
+  weighting_scheme: logit_normal
+  zero_noise: false
+monitor_grad_freq: 50
+monitor_grad_histograms: false
+monitor_gradients: false
+monitor_semantic_distribution: true
+monitor_semantic_freq: 50
+monitor_semantic_top_k: 50
+name: Generator
+norm_by: std
+num_sanity_val_steps: 0
+num_workers: 4
+precision: bf16-mixed
+profiler: simple
+progressive_end: 100
+resume: /cluster/andram/qmeng/Experiments/scene_completion/auto_encoder/2025-12-19_01-23-28-525/generator/2026-02-23_16-22-25-152
+sample_posterior: true
+seed: 23
+slurm:
+  cluster: null
+  cpus_per_task: 4
+  folder: null
+  gpus_per_node: 4
+  mem_gb: 128
+  nodes: 1
+  slurm_constraint: '[rtx_a6000|a100]'
+  slurm_exclude: sorona,
+  slurm_job_name: 2026-02-23_16-22-25-152_prec_bf16-mix_nsvs_0_alog_2025-12-_lkey_tsdf_p_1_pend_100_msdi_udro_ulau
+  slurm_mail_type: null
+  slurm_partition: submit
+  slurm_time: 4-00:00:00
+  stderr_to_stdout: true
+  tasks_per_node: 4
+  timeout_min: 2880
+slurm_folder: /cluster/andram/qmeng/Experiments/scene_completion/auto_encoder/2025-12-19_01-23-28-525/generator/2026-02-23_16-22-25-152
+strategy: ddp_find_unused_parameters_true
+target: seen2scene.models.trainer.train_gen.Net
+task:
+  backend: pyrender
+  guidance_scale: 3.0
+  img_wh: 150
+  metrics: !!python/tuple
+  - fid
+  name: train_gen
+  num_for_plot: 1
+  num_samples: 128
+  num_variations: 2
+  num_views_for_fid: 10
+  scene_shape: !!python/tuple
+  - 256
+  - 256
+  - 256
+test_data_keys:
+- tsdf_p_1.0
+- mesh_gt
+test_split: test
+timestamp: 2026-02-23_16-22-25-152
+train_data_keys:
+- tsdf_p_1.0
+vae_cfg:
+  channels: 8
+  ckpt_path: /cluster/andram/qmeng/Experiments/scene_completion/auto_encoder/2025-12-19_01-23-28-525/checkpoint/vxl_0_011_last.ckpt
+  decoder: models.fvdb.decoder.Decoder
+  double_z: true
+  embeding_freq: -1
+  empty_token: true
+  encoder: models.fvdb.encoder.Encoder
+  f_maps: 64
+  factor: 8
+  feed_gt_structure: false
+  gaussian_tau: null
+  geometry_weight: 100.0
+  is_add_dec: false
+  kl_weight: 100.0
+  mask_unknown: true
+  max_down_time: 6
+  name: fVDB VAE
+  num_groups: 32
+  num_res_blocks: 1
+  order: gcs
+  structure_weight: 100.0
+  target: models.fvdb.sparse_vae.SparseVAE
+  unstable_cutoff_threshold: 0.15
+  use_attention: false
+  use_checkpoint: true
+  use_residual: true
+  with_color_branch: false
+  with_semantic_branch: false
+vae_ckpt: /cluster/andram/qmeng/Experiments/scene_completion/auto_encoder/2025-12-19_01-23-28-525/checkpoint/vxl_0_011_last.ckpt
+val_data_keys:
+- tsdf_p_1.0
+- mesh_gt
+val_split: val

experiments/auto_encoder/2025-12-19_01-23-28-525/generator/2026-02-23_16-22-25-152/control/2026-02-26_14-01-23-930/config.yaml

@@ -0,0 +1,204 @@
+accelerator: gpu
+accumulate_grad_batches: 1
+ae_log: 2025-12-19_01-23-28-525
+avg_over: region
+batch_size: 1
+benchmark: true
+check_val_every_n_epoch: 5
+ckpt_path: null
+data:
+  augmentation: true
+  bbox_rand_shift: 1.0
+  bottom_padding: 3.0
+  category_exclude: !!python/tuple
+  - floor
+  - ceiling
+  - wall
+  category_include: null
+  ceiling_clip: 2.2
+  csv_path: null
+  data_list: !!python/tuple
+  - 3D-FRONT
+  - ScannetPP
+  - ARKitScenes
+  fixed_len: 20000
+  full_scene: false
+  known_ratio: 0.999
+  num_categories: 150
+  overfit_patch: false
+  overfit_scene: null
+  patch_shape: !!python/tuple
+  - 256
+  - 256
+  - 256
+  target: seen2scene.dataset.dataset.Dataset
+  truncation: null
+  use_semantic_mapping: false
+  voxel_size: 0.011
+ema_update_batch: 5
+gen_log: 2026-02-23_16-22-25-152
+gradient_clip_algorithm: norm
+gradient_clip_val: 1.0
+ignore_keys: null
+job_name: 2026-02-26_14-01-23-930_prec_bf16-mix_nsvs_1_pend_30_mepo_60_alog_2025-12-_glog_2026-02-_skey_tsdf_p
+latent_key: tsdf_p_1.0
+learning_rate: 0.0001
+log_dir: /cluster/andram/qmeng/Experiments/scene_completion/auto_encoder/2025-12-19_01-23-28-525/generator/2026-02-23_16-22-25-152/control/2026-02-26_14-01-23-930
+log_root: /cluster/andram/qmeng/Experiments/scene_completion
+lr_scheduler:
+  lr: 0.0001
+  min_lr: 1.0e-07
+  name: cosine_warmup
+  num_warmup_steps: 1000.0
+mask_unknown: true
+max_epochs: 60
+max_steps: -1
+model:
+  gen_cfg:
+    band_weight: 0.5
+    cond_drop_prob: 0.1
+    logit_mean: 0.0
+    logit_std: 1.0
+    mode_scale: 1.29
+    model_cfg:
+      attn_mode: full
+      compression_block_size: 2
+      compression_version: v2
+      context_dim: null
+      factor: 1.0
+      in_channels: 8
+      mlp_ratio: 4
+      model_channels: 768
+      name: SparseDiT
+      num_blocks: 28
+      num_head_channels: 32
+      num_heads: 16
+      num_kv_heads: 2
+      out_channels: 8
+      pe_mode: rope
+      qk_rms_norm: true
+      qk_rms_norm_cross: true
+      selection_block_size: 4
+      share_mod: false
+      target: seen2scene.models.sparse.dit.SparseDiT
+      topk: 8
+      use_checkpoint: true
+      use_shift: true
+      use_ssa: false
+      window_size: 4
+      zero_out: true
+    name: FM
+    num_inference_steps: 50
+    precondition_outputs: false
+    scheduler:
+      num_train_timesteps: 1000
+      target: diffusers.FlowMatchEulerDiscreteScheduler
+    target: seen2scene.models.backbone.flow_matching.Model
+    use_distance_rope: true
+    use_ema: true
+    use_label_augmentation: true
+    weighting_scheme: logit_normal
+    zero_noise: false
+  gen_ckpt: /cluster/andram/qmeng/Experiments/scene_completion/auto_encoder/2025-12-19_01-23-28-525/generator/2026-02-23_16-22-25-152/checkpoint/vxl_0_011_last.ckpt
+  name: ControlNet
+  target: seen2scene.models.backbone.controlnet.Net
+  zero_proj_rank: 64
+monitor_grad_freq: 50
+monitor_grad_histograms: false
+monitor_gradients: false
+monitor_semantic_distribution: false
+monitor_semantic_freq: 50
+monitor_semantic_top_k: 50
+name: control
+norm_by: std
+num_sanity_val_steps: 1
+num_workers: 4
+precision: bf16-mixed
+profiler: simple
+progressive_end: 30
+resume: /cluster/andram/qmeng/Experiments/scene_completion/auto_encoder/2025-12-19_01-23-28-525/generator/2026-02-23_16-22-25-152/control/2026-02-26_14-01-23-930
+sample_posterior: true
+seed: 23
+slurm:
+  cluster: null
+  cpus_per_task: 4
+  folder: null
+  gpus_per_node: 4
+  mem_gb: 128
+  nodes: 1
+  slurm_constraint: '[rtx_a6000|a100]'
+  slurm_exclude: sorona,
+  slurm_job_name: 2026-02-26_14-01-23-930_prec_bf16-mix_nsvs_1_pend_30_mepo_60_alog_2025-12-_glog_2026-02-_skey_tsdf_p
+  slurm_mail_type: null
+  slurm_partition: submit
+  slurm_time: 4-00:00:00
+  stderr_to_stdout: true
+  tasks_per_node: 4
+  timeout_min: 2880
+slurm_folder: /cluster/andram/qmeng/Experiments/scene_completion/auto_encoder/2025-12-19_01-23-28-525/generator/2026-02-23_16-22-25-152/control/2026-02-26_14-01-23-930
+src_key: tsdf_p_0.1
+strategy: ddp_find_unused_parameters_true
+target: seen2scene.models.trainer.train_gen.Net
+task:
+  backend: pyrender
+  guidance_scale: 3.0
+  img_wh: 150
+  metrics: !!python/tuple
+  - iou
+  - l1
+  - l2
+  - tmd
+  name: train_control
+  num_for_plot: 1
+  num_samples: 8
+  num_variations: 2
+  num_views_for_fid: 10
+  repeat_num: 2
+  scene_shape: !!python/tuple
+  - 256
+  - 256
+  - 256
+test_data_keys:
+- tsdf_p_0.1
+- tsdf_p_1.0
+- mesh_gt
+test_split: test
+timestamp: 2026-02-26_14-01-23-930
+train_data_keys:
+- tsdf_p_0.1
+- tsdf_p_1.0
+vae_cfg:
+  channels: 8
+  ckpt_path: /cluster/andram/qmeng/Experiments/scene_completion/auto_encoder/2025-12-19_01-23-28-525/checkpoint/vxl_0_011_last.ckpt
+  decoder: models.fvdb.decoder.Decoder
+  double_z: true
+  embeding_freq: -1
+  empty_token: true
+  encoder: models.fvdb.encoder.Encoder
+  f_maps: 64
+  factor: 8
+  feed_gt_structure: false
+  gaussian_tau: null
+  geometry_weight: 100.0
+  is_add_dec: false
+  kl_weight: 100.0
+  mask_unknown: true
+  max_down_time: 6
+  name: fVDB VAE
+  num_groups: 32
+  num_res_blocks: 1
+  order: gcs
+  structure_weight: 100.0
+  target: models.fvdb.sparse_vae.SparseVAE
+  unstable_cutoff_threshold: 0.15
+  use_attention: false
+  use_checkpoint: true
+  use_residual: true
+  with_color_branch: false
+  with_semantic_branch: false
+vae_ckpt: /cluster/andram/qmeng/Experiments/scene_completion/auto_encoder/2025-12-19_01-23-28-525/checkpoint/vxl_0_011_last.ckpt
+val_data_keys:
+- tsdf_p_0.1
+- tsdf_p_1.0
+- mesh_gt
+val_split: val