yslan commited on
Commit
3ac68de
·
verified ·
1 Parent(s): dc04269

Add immutable checkpoint backup cosmos3_nano_stage_one_lora_r256_replicate_only_500k_step_006000_20261104 data_config_provenance

Browse files
cosmos3_nano_stage_one_lora_r256_replicate_only_500k_20261103/attempt01/step_006000/source_windows.yaml ADDED
@@ -0,0 +1,387 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ vgg_repository: /vgg/yushi/Repo/vgg_forecast_4d_cosmos_pinned_source_20261102
2
+ vgg_commit: f5a59cd511f20c20cf09e2edd21781ca6907d4c8
3
+ recipe_source_config: /scratch/shared/engs2907/logs/vgg_forecast_4d/train/flexpi_four_d_pretraining_wan22_5b_separate_streams_full_parameters_video_depth_camera_dino_and_sparse_tracking_video_from_the_15000_update_four_d_run_20261101/attempt02/run/config.yaml
4
+ recipe_source_config_sha256: 72590f4598cae1b329c31b3cb45d059bdaeca23642faa10ca3f31372f9ebc0aa
5
+ window_block:
6
+ _target_: four_d_windows.FourDPretrainingWindows
7
+ sources:
8
+ - name: four_d_pretraining_relaunch_syn4d_sources
9
+ kind: track4d_source_config
10
+ config: train/mot/wan_video_depth_vae_from_video_four_d_pretraining_relaunch_rendered_depth_sources_21_frames_512
11
+ weight: 15.2
12
+ left_out_sources:
13
+ - omniworld_game
14
+ - tartanair
15
+ - mvs_synth
16
+ condition_only_streams: []
17
+ - name: four_d_pretraining_relaunch_omniworld_game_source
18
+ kind: track4d_source_config
19
+ config: train/mot/wan_video_depth_vae_from_video_four_d_pretraining_relaunch_rendered_depth_sources_21_frames_512
20
+ weight: 8.75
21
+ left_out_sources:
22
+ - syn4d_v1
23
+ - syn4d_v2_mixed
24
+ - syn4d_v2_mixed_no_sim
25
+ - tartanair
26
+ - mvs_synth
27
+ condition_only_streams: []
28
+ - name: four_d_pretraining_relaunch_synthverse_source
29
+ kind: track4d_source_config
30
+ config: train/mot/wan_video_depth_vae_from_video_four_d_pretraining_relaunch_synthverse_dynamic_replica_vkitti_midair_spring_sources_without_midair_quality_masks_21_frames_512
31
+ weight: 11.05
32
+ left_out_sources:
33
+ - dynamic_replica
34
+ - vkitti
35
+ - midair
36
+ - spring
37
+ condition_only_streams: []
38
+ - name: four_d_pretraining_relaunch_vkitti_source
39
+ kind: track4d_source_config
40
+ config: train/mot/wan_video_depth_vae_from_video_four_d_pretraining_relaunch_synthverse_dynamic_replica_vkitti_midair_spring_sources_without_midair_quality_masks_21_frames_512
41
+ weight: 6.0
42
+ left_out_sources:
43
+ - synthverse_depth
44
+ - dynamic_replica
45
+ - midair
46
+ - spring
47
+ condition_only_streams: []
48
+ - name: four_d_pretraining_relaunch_tartanair_sky_then_far_percentile_source
49
+ kind: track4d_source_config
50
+ config: train/mot/wan_video_depth_vae_from_video_four_d_pretraining_relaunch_sky_and_far_percentile_depth_validity_rules_21_frames_512
51
+ weight: 6.0
52
+ left_out_sources:
53
+ - dynamic_replica
54
+ - spring
55
+ - midair
56
+ - mvs_synth
57
+ - tartanground
58
+ - paralleldomain4d
59
+ condition_only_streams: []
60
+ - name: four_d_pretraining_relaunch_dynamic_replica_stored_depth_source
61
+ kind: track4d_source_config
62
+ config: train/mot/wan_video_depth_vae_from_video_four_d_pretraining_relaunch_sky_and_far_percentile_depth_validity_rules_21_frames_512
63
+ weight: 5.0
64
+ left_out_sources:
65
+ - tartanair
66
+ - spring
67
+ - midair
68
+ - mvs_synth
69
+ - tartanground
70
+ - paralleldomain4d
71
+ condition_only_streams: []
72
+ - name: four_d_pretraining_relaunch_spring_stored_depth_source
73
+ kind: track4d_source_config
74
+ config: train/mot/wan_video_depth_vae_from_video_four_d_pretraining_relaunch_sky_and_far_percentile_depth_validity_rules_21_frames_512
75
+ weight: 8.0
76
+ left_out_sources:
77
+ - tartanair
78
+ - dynamic_replica
79
+ - midair
80
+ - mvs_synth
81
+ - tartanground
82
+ - paralleldomain4d
83
+ condition_only_streams: []
84
+ - name: four_d_pretraining_relaunch_midair_within_200_meters_source
85
+ kind: track4d_source_config
86
+ config: train/mot/wan_video_depth_vae_from_video_four_d_pretraining_relaunch_sky_and_far_percentile_depth_validity_rules_21_frames_512
87
+ weight: 3.0
88
+ left_out_sources:
89
+ - tartanair
90
+ - dynamic_replica
91
+ - spring
92
+ - mvs_synth
93
+ - tartanground
94
+ - paralleldomain4d
95
+ condition_only_streams: []
96
+ - name: four_d_pretraining_relaunch_mvs_synth_1000_then_far_percentile_source
97
+ kind: track4d_source_config
98
+ config: train/mot/wan_video_depth_vae_from_video_four_d_pretraining_relaunch_sky_and_far_percentile_depth_validity_rules_21_frames_512
99
+ weight: 3.0
100
+ left_out_sources:
101
+ - tartanair
102
+ - dynamic_replica
103
+ - spring
104
+ - midair
105
+ - tartanground
106
+ - paralleldomain4d
107
+ condition_only_streams: []
108
+ - name: four_d_pretraining_relaunch_tartanground_sky_then_far_percentile_source
109
+ kind: track4d_source_config
110
+ config: train/mot/wan_video_depth_vae_from_video_four_d_pretraining_relaunch_sky_and_far_percentile_depth_validity_rules_21_frames_512
111
+ weight: 3.0
112
+ left_out_sources:
113
+ - tartanair
114
+ - dynamic_replica
115
+ - spring
116
+ - midair
117
+ - mvs_synth
118
+ - paralleldomain4d
119
+ condition_only_streams: []
120
+ - name: four_d_pretraining_relaunch_paralleldomain4d_sky_then_far_percentile_source
121
+ kind: track4d_source_config
122
+ config: train/mot/wan_video_depth_vae_from_video_four_d_pretraining_relaunch_sky_and_far_percentile_depth_validity_rules_21_frames_512
123
+ weight: 10.0
124
+ left_out_sources:
125
+ - tartanair
126
+ - dynamic_replica
127
+ - spring
128
+ - midair
129
+ - mvs_synth
130
+ - tartanground
131
+ condition_only_streams: []
132
+ frames: 9
133
+ raster_hw:
134
+ - 288
135
+ - 512
136
+ raymap_camera:
137
+ origin_fraction: 0.3333333333333333
138
+ origin_range: 1.0
139
+ decode_margin: 16
140
+ origin_box_grid: 1
141
+ pointmap_log_factor: 0.4
142
+ pointmap_hole_fill:
143
+ synthverse_depth: far_clip
144
+ syn4d: far_clip
145
+ omniworld_game: far_clip
146
+ dynamic_replica: nearest_labelled_depth_of_3x3_neighbours
147
+ tartanair: far_clip
148
+ dl3dv: far_clip
149
+ scannet: telea_log_depth
150
+ virtual_kitti2: far_clip
151
+ midair: far_clip
152
+ pointodyssey: far_clip
153
+ mvs_synth: far_clip
154
+ wildrgbd: telea_log_depth
155
+ hypersim: far_clip
156
+ spring: far_clip
157
+ kubric49: far_clip
158
+ kubric81: far_clip
159
+ droid_stereo_depth: telea_log_depth
160
+ tartanground: far_clip
161
+ paralleldomain4d: far_clip
162
+ syn4d_v2: far_clip
163
+ pointmap_gauge: window_labelled_median_depth
164
+ geometry_image: genception_log_depth
165
+ context_len: 512
166
+ batch_size: 16
167
+ stream_seed: 42
168
+ epoch_windows: 32000000
169
+ val_set_proportion: 0.0
170
+ resume_state: null
171
+ gradient_accumulation_steps: 1
172
+ loader_workers: 12
173
+ loader_opencv_threads: 1
174
+ window_captions:
175
+ caption_dropout: 0.1
176
+ missing_caption: empty_prompt
177
+ caches:
178
+ syn4d_v1_gemini_holo4d:
179
+ embeddings: /scratch/shared/engs2907/logs/vgg_forecast_4d/cache/wan_umt5_captions/four_d_window_captions_512/syn4d_v1_gemini_holo4d_local_and_global_with_repaired_globals_20261031/captions_t5_emb.safetensors
180
+ meta: /scratch/shared/engs2907/logs/vgg_forecast_4d/cache/wan_umt5_captions/four_d_window_captions_512/syn4d_v1_gemini_holo4d_local_and_global_with_repaired_globals_20261031/captions_t5_emb_meta.json
181
+ syn4d_v2_gemini_holo4d:
182
+ embeddings: /scratch/shared/engs2907/logs/vgg_forecast_4d/cache/wan_umt5_captions/four_d_window_captions_512/syn4d_v2_gemini_holo4d_20261030/captions_t5_emb.safetensors
183
+ meta: /scratch/shared/engs2907/logs/vgg_forecast_4d/cache/wan_umt5_captions/four_d_window_captions_512/syn4d_v2_gemini_holo4d_20261030/captions_t5_emb_meta.json
184
+ tartanair_gemini_holo4d:
185
+ embeddings: /scratch/shared/engs2907/logs/vgg_forecast_4d/cache/wan_umt5_captions/four_d_window_captions_512/tartanair_gemini_holo4d/captions_t5_emb.safetensors
186
+ meta: /scratch/shared/engs2907/logs/vgg_forecast_4d/cache/wan_umt5_captions/four_d_window_captions_512/tartanair_gemini_holo4d/captions_t5_emb_meta.json
187
+ mvs_synth_gemini_holo4d:
188
+ embeddings: /scratch/shared/engs2907/logs/vgg_forecast_4d/cache/wan_umt5_captions/four_d_window_captions_512/mvs_synth_gemini_holo4d/captions_t5_emb.safetensors
189
+ meta: /scratch/shared/engs2907/logs/vgg_forecast_4d/cache/wan_umt5_captions/four_d_window_captions_512/mvs_synth_gemini_holo4d/captions_t5_emb_meta.json
190
+ omniworld_game_gemini_local_official_text_chunks_and_gemini_global:
191
+ embeddings: /scratch/shared/engs2907/logs/vgg_forecast_4d/cache/wan_umt5_captions/four_d_window_captions_512/omniworld_game_gemini_local_official_text_chunks_and_gemini_global_20261030b/captions_t5_emb.safetensors
192
+ meta: /scratch/shared/engs2907/logs/vgg_forecast_4d/cache/wan_umt5_captions/four_d_window_captions_512/omniworld_game_gemini_local_official_text_chunks_and_gemini_global_20261030b/captions_t5_emb_meta.json
193
+ synthverse_gemini_holo4d:
194
+ embeddings: /scratch/shared/engs2907/logs/vgg_forecast_4d/cache/wan_umt5_captions/four_d_window_captions_512/synthverse_gemini_holo4d_local96_and_global_20261031/captions_t5_emb.safetensors
195
+ meta: /scratch/shared/engs2907/logs/vgg_forecast_4d/cache/wan_umt5_captions/four_d_window_captions_512/synthverse_gemini_holo4d_local96_and_global_20261031/captions_t5_emb_meta.json
196
+ dynamic_replica_gemini_holo4d:
197
+ embeddings: /scratch/shared/engs2907/logs/vgg_forecast_4d/cache/wan_umt5_captions/four_d_window_captions_512/dynamic_replica_gemini_holo4d/captions_t5_emb.safetensors
198
+ meta: /scratch/shared/engs2907/logs/vgg_forecast_4d/cache/wan_umt5_captions/four_d_window_captions_512/dynamic_replica_gemini_holo4d/captions_t5_emb_meta.json
199
+ vkitti_gemini_holo4d:
200
+ embeddings: /scratch/shared/engs2907/logs/vgg_forecast_4d/cache/wan_umt5_captions/four_d_window_captions_512/vkitti_gemini_holo4d/captions_t5_emb.safetensors
201
+ meta: /scratch/shared/engs2907/logs/vgg_forecast_4d/cache/wan_umt5_captions/four_d_window_captions_512/vkitti_gemini_holo4d/captions_t5_emb_meta.json
202
+ midair_gemini_holo4d:
203
+ embeddings: /scratch/shared/engs2907/logs/vgg_forecast_4d/cache/wan_umt5_captions/four_d_window_captions_512/midair_gemini_holo4d_20261030/captions_t5_emb.safetensors
204
+ meta: /scratch/shared/engs2907/logs/vgg_forecast_4d/cache/wan_umt5_captions/four_d_window_captions_512/midair_gemini_holo4d_20261030/captions_t5_emb_meta.json
205
+ spring_gemini_holo4d:
206
+ embeddings: /scratch/shared/engs2907/logs/vgg_forecast_4d/cache/wan_umt5_captions/four_d_window_captions_512/spring_gemini_holo4d/captions_t5_emb.safetensors
207
+ meta: /scratch/shared/engs2907/logs/vgg_forecast_4d/cache/wan_umt5_captions/four_d_window_captions_512/spring_gemini_holo4d/captions_t5_emb_meta.json
208
+ tartanground_gemini_holo4d:
209
+ embeddings: /scratch/shared/engs2907/logs/vgg_forecast_4d/cache/wan_umt5_captions/four_d_window_captions_512/tartanground_gemini_holo4d_20261030/captions_t5_emb.safetensors
210
+ meta: /scratch/shared/engs2907/logs/vgg_forecast_4d/cache/wan_umt5_captions/four_d_window_captions_512/tartanground_gemini_holo4d_20261030/captions_t5_emb_meta.json
211
+ paralleldomain4d_gemini_holo4d:
212
+ embeddings: /scratch/shared/engs2907/logs/vgg_forecast_4d/cache/wan_umt5_captions/four_d_window_captions_512/paralleldomain4d_gemini_holo4d_20261030/captions_t5_emb.safetensors
213
+ meta: /scratch/shared/engs2907/logs/vgg_forecast_4d/cache/wan_umt5_captions/four_d_window_captions_512/paralleldomain4d_gemini_holo4d_20261030/captions_t5_emb_meta.json
214
+ stanzas:
215
+ syn4d_v1:
216
+ product: syn4d_v1_gemini_holo4d
217
+ join: local_covering_then_global
218
+ syn4d_v2_mixed:
219
+ product: syn4d_v2_gemini_holo4d
220
+ join: local_covering_then_global
221
+ syn4d_v2_mixed_no_sim:
222
+ product: syn4d_v2_gemini_holo4d
223
+ join: local_covering_then_global
224
+ tartanair:
225
+ product: tartanair_gemini_holo4d
226
+ join: local_covering
227
+ mvs_synth:
228
+ product: mvs_synth_gemini_holo4d
229
+ join: local_covering
230
+ omniworld_game:
231
+ product: omniworld_game_gemini_local_official_text_chunks_and_gemini_global
232
+ join: local_covering_then_second_local_covering_then_global
233
+ pointodyssey_test: none
234
+ kubric49_validation: none
235
+ wildrgbd_test: none
236
+ vkitti_test:
237
+ product: vkitti_gemini_holo4d
238
+ join: local_covering_then_global
239
+ droid_held_out: none
240
+ synthverse_depth:
241
+ product: synthverse_gemini_holo4d
242
+ join: local_covering_then_global
243
+ dynamic_replica:
244
+ product: dynamic_replica_gemini_holo4d
245
+ join: global
246
+ vkitti:
247
+ product: vkitti_gemini_holo4d
248
+ join: local_covering_then_global
249
+ midair:
250
+ product: midair_gemini_holo4d
251
+ join: local_covering_then_global
252
+ spring:
253
+ product: spring_gemini_holo4d
254
+ join: global
255
+ tartanground:
256
+ product: tartanground_gemini_holo4d
257
+ join: local_covering_then_global
258
+ paralleldomain4d:
259
+ product: paralleldomain4d_gemini_holo4d
260
+ join: global
261
+ paralleldomain4d_validation:
262
+ product: paralleldomain4d_gemini_holo4d
263
+ join: global
264
+ training_scene_exclusions:
265
+ reason: rendered depth passes through glass (raw EXR depth through glass panes
266
+ at 190-260 m), owner 10-29; evidence /scratch/shared/engs2907/logs/vgg_forecast_4d/tools/syn4d_glass_depth_check_20261029;
267
+ scifiroom_bald added 10-30 from /scratch/shared/engs2907/logs/vgg_forecast_4d/tools/depth_clipping_audit_20261030
268
+ stanzas:
269
+ syn4d_v1:
270
+ - bigoffice_v1
271
+ - lab_bald
272
+ - office_bald
273
+ - scifiroom_bald
274
+ syn4d_v2_mixed:
275
+ - citynight
276
+ - mall
277
+ - metro
278
+ syn4d_v2_mixed_no_sim:
279
+ - citynight
280
+ - mall
281
+ - metro
282
+ tracking_video:
283
+ query_grid_px: 10
284
+ disc_radius_px: 2
285
+ occluded_style: ring
286
+ ring_radius_px: 3
287
+ ring_width_px: 1
288
+ xyz_normalization: quantile_0.02_0.98
289
+ colour_floor: 0.2
290
+ min_frame_zero_points: 16
291
+ surface_test:
292
+ relative_tolerance: 0.1
293
+ absolute_tolerance: 0.05
294
+ stanza_rules:
295
+ tartanair: static_scene_points
296
+ midair: static_scene_points
297
+ mvs_synth: static_scene_points
298
+ tartanground: static_scene_points
299
+ syn4d_v1: tracked_points
300
+ syn4d_v2_mixed: tracked_points
301
+ syn4d_v2_mixed_no_sim: tracked_points
302
+ omniworld_game: no_tracking_video
303
+ synthverse_depth: no_tracking_video
304
+ dynamic_replica: tracked_points
305
+ vkitti: no_tracking_video
306
+ spring: no_tracking_video
307
+ paralleldomain4d: no_tracking_video
308
+ scannet: static_scene_points
309
+ wildrgbd_test: static_scene_points
310
+ pointodyssey_test: tracked_points
311
+ kubric49_validation: no_tracking_video
312
+ vkitti_test: no_tracking_video
313
+ droid_held_out: no_tracking_video
314
+ paralleldomain4d_validation: no_tracking_video
315
+ dynamic_replica_test: tracked_points
316
+ stanza_track_reading:
317
+ syn4d_v1:
318
+ stanza:
319
+ capability: dynamic_cross_time
320
+ supervision:
321
+ - pointmap
322
+ - source_motion
323
+ annotation_provenance: syn4d_mesh_surface_forward_track_from_frame0
324
+ allow_missing_tracks: false
325
+ kwargs:
326
+ return_tracks: true
327
+ depth_only: false
328
+ metadata_root: /vgg/chenzhao/syn4d_metadata
329
+ wrapped_face_distance_thresh: 0.1
330
+ query_inconsistent_distance_thresh: 0.1
331
+ pointodyssey_test:
332
+ stanza:
333
+ supervision:
334
+ - pointmap
335
+ - source_motion
336
+ kwargs:
337
+ return_tracks: true
338
+ dynamic_replica_test:
339
+ stanza:
340
+ supervision:
341
+ - pointmap
342
+ - source_motion
343
+ kwargs:
344
+ return_tracks: true
345
+ syn4d_v2_mixed:
346
+ loader: syn4d_v2
347
+ stanza:
348
+ capability: dynamic_cross_time
349
+ supervision:
350
+ - pointmap
351
+ - source_motion
352
+ annotation_provenance: syn4d_v2_mesh_surface_forward_track_from_frame0
353
+ allow_missing_tracks: false
354
+ kwargs:
355
+ return_tracks: true
356
+ depth_only: false
357
+ metadata_root: /vgg/chenzhao/syn4d_metadata
358
+ tracking_format: safetensor
359
+ mask_bg: false
360
+ trusted_background_depth_max: 80.0
361
+ wrapped_face_distance_thresh: 0.1
362
+ query_inconsistent_distance_thresh: 0.1
363
+ syn4d_v2_mixed_no_sim:
364
+ loader: syn4d_v2
365
+ stanza:
366
+ capability: dynamic_cross_time
367
+ supervision:
368
+ - pointmap
369
+ - source_motion
370
+ annotation_provenance: syn4d_v2_mesh_surface_forward_track_from_frame0
371
+ allow_missing_tracks: false
372
+ kwargs:
373
+ return_tracks: true
374
+ depth_only: false
375
+ metadata_root: /vgg/chenzhao/syn4d_metadata
376
+ tracking_format: safetensor
377
+ mask_bg: false
378
+ trusted_background_depth_max: 80.0
379
+ wrapped_face_distance_thresh: 0.1
380
+ query_inconsistent_distance_thresh: 0.1
381
+ dynamic_replica:
382
+ stanza:
383
+ supervision:
384
+ - pointmap
385
+ - source_motion
386
+ kwargs:
387
+ return_tracks: true