# BVH, Feb 2026. # CVPR ablation: train only on GCDKubric4D + GCDParDom4D (vs. full mixture in rdvs2_mix_b16.py). import os from hydra.core.config_store import ConfigStore from custom.experiment.template import template_any4d_2b ################################################################# job = dict( # this becomes JobConfig project='a4d2', group='cvpr', name='dvs2', prepend_datetime=True, # becomes {now:%m-%d-%H-%M}_{config.job.name} ) wandb = dict( enabled=True, entity='tri', project='a4d2_cvpr', num_validation_logs=5, ) any4d_config = dict( # this is part of Any4DConfig, which inherits from Predict2Video2WorldModelConfig transforms='default', # NOTE(bvh): reference_camera now separate per dataset vidar2a4d='basile', vae='a4d_vae', video_entries=[ # first entry = base / pretrained dict( # key: (channel_start, channel_end, in_proj_init, out_proj_init) # NOTE: there will be T*H*W tokens with this information rgb0=(0, 16, 'load', 'load'), # do not change rgb0_input_mask=(16, 17, 'load', None), # do not change rgb0_output_mask=(17, 18, 'zero', None), # do not change cams0=(18, 50, 'zero/load', 'zero/load'), cams0_input_mask=(50, 51, 'zero/load', None), cams0_output_mask=(51, 52, 'zero', None), ), # later entries = new viewpoints dict( # key: (channel_start, channel_end, in_proj_init, out_proj_init) rgb1=(0, 16, 'copy:rgb0/load', 'copy:rgb0/load'), rgb1_input_mask=(16, 17, 'copy:rgb0_input_mask/load', None), rgb1_output_mask=(17, 18, 'zero', None), cams1=(18, 50, 'zero/load', 'zero/load'), cams1_input_mask=(50, 51, 'zero/load', None), cams1_output_mask=(51, 52, 'zero', None), ), ], num_views=2, video_concat_mode='view', video_proj_mode='per_view', view_timestep_mode='per_view', loss_weights=dict( rgb0=1.0, # no rgb1 since predicted view is always rgb0 ), harmonize_streams=True, harmonize_frames=True, load_modals=['rgb', 'cams'], task_probs=dict( cross_modal=0.0, dyn_view_synth=1.0, forecast=0.0, pose_est=0.0, inv_dyn=0.0, policy=0.0, world_model=0.0, ), use_views='all', shuffle_cams2views=False, # AV train pairs are already canonical L/R; no shuffle train_directives=dict( num_pred_views=1, ), val_directives=dict( tasks='dyn_view_synth', num_pred_views=1, remove_cond=True, ), # NOTE(bvh): pre-refactor flat fields are deprecated; semantics now in data_*_overrides. # override_zero_origin=True, # makes more sense for 1->1 DVS setting? # override_temporal_stride_val=1, data_train_overrides=dict(zero_origin=True), data_val_overrides=dict(frame_stride=1, zero_origin=True), track_metrics=dict( rgb0=['psnr', 'ssim'], # no rgb1 since predicted view is always rgb0 ), train_visuals_interval=99, train_visuals_detail=0, val_visuals_detail=3, visuals_quality=8, viz_input_blacklist=[], # overridden to show cams viz_mask_border_width=0, # overridden because DVS is temporally invariant val_num_steps=35, ) ################################################################# S3_PRETRAINED_PREFIX = r's3://tri-ml-sandbox-16011-us-west-2-datasets/cosmos-predict-2/checkpoints' model = dict( # this makes up parts of Predict2Video2WorldModelConfig and Predict2ModelManagerConfig # ablation so start from vanilla cosmos # dit_path=f'{S3_PRETRAINED_PREFIX}/nvidia/Cosmos-Predict2-2B-Video2World/model-480p-10fps.pt', dit_path=f's3://tri-ml-sandbox-16011-us-west-2-datasets/sagemaker/cosmos-predict2/a4d2/cvpr/02-25-23-14_s30_avab_gcd/model/iter_000011000_002816000.pt', # text_encoder_path=f'{S3_PRETRAINED_PREFIX}/google-t5/t5-11b', text_encoder_path='', # disable vae_path=f'{S3_PRETRAINED_PREFIX}/nvidia/Cosmos-Predict2-2B-Video2World/tokenizer/tokenizer.pth', fsdp_shard_size=8, # fsdp_shard_size=4, ############ run_validation=True, skip_first_validation='delay', # to check VRAM validation_iter=500, # frequency / interval of validation runs # max_val_iter=8, # number of steps per validation run max_iter=40000, # total number of training steps # NOTE: ^ AV paper says: 30k LR + 10k HR # grad_accum_iter=2, grad_accum_iter=1, context_parallel_size=1, device_monitor=0, manual_gc_iter=288, manual_gc_warm_up=-1, # never disable automatic GC to be safe (weird VRAM issue) ############ state_t=11, # for noise level; = latent # frames for now tokenizer_chunk_duration=41, # for VAE; = raw # frames for now ) checkpoint = dict( save_iter=1000, s3_folder='s3://tri-ml-sandbox-16011-us-west-2-datasets/sagemaker/cosmos-predict2/', early_sanity_check=5, ) optimizer = dict( lr=5.0e-5, ) scheduler = dict( warm_up_steps = [500], cycle_lengths = [model['max_iter']], f_start=[0.01], f_max=[1.0], f_min=[0.05], ) dataset_train = dict( # config='custom/config/cvpr/mixed/train_2v_32gpu.yaml', config='custom/config/ablation/gcd_train_2v.yaml', num_workers=4, # per GPU # batch_size=4, # per GPU # batch_size=8, # per GPU batch_size=16, # per GPU # batch_size=32, # per GPU # batch_size=64, # per GPU ) dataset_val = dict( config=dict( ### 3D DL3DVQL = 'custom/config/cvpr/qual_fast_LR/dl3dv.yaml', # 64 MVImgNetQL = 'custom/config/cvpr/qual_fast_LR/mvimgnet.yaml', # 64 RE10KQL = 'custom/config/cvpr/qual_fast_LR/re10k.yaml', # 64 ScanNetQL = 'custom/config/cvpr/qual_fast_LR/scannet.yaml', # 64 TartanAirQL = 'custom/config/cvpr/qual_fast_LR/tartanair.yaml', # 64 WildRGBDQL = 'custom/config/cvpr/qual_fast_LR/wildrgbd.yaml', # 64 ### 4D AssemblyHandsQL = 'custom/config/cvpr/qual_fast_LR/assemblyhands.yaml', # 64 DyCheckMQL = 'custom/config/cvpr/qual_fast_LR/dycheckm.yaml', # 64 EgoExo4DIDQL = 'custom/config/cvpr/qual_fast_LR/egoexo4d_ID.yaml', # 64 EgoExo4DOODQL = 'custom/config/cvpr/qual_fast_LR/egoexo4d_OOD.yaml', # 64 GCDK4DDirQL = 'custom/config/cvpr/qual_fast_LR/kubric4d_dir.yaml', # 64 GCDK4DGradQL = 'custom/config/cvpr/qual_fast_LR/kubric4d_grad.yaml', # 64 # Kubric4DQL = 'custom/config/cvpr/qual_fast_LR/kubric4d.yaml', # freezes? Kubric5DQL = 'custom/config/cvpr/qual_fast_LR/kubric5d.yaml', # 64 ### DRIVING ArgoverseQL = 'custom/config/cvpr/qual_fast_LR/argoverse.yaml', # 64 DDADQL = 'custom/config/cvpr/qual_fast_LR/ddad.yaml', # 64 LyftL5QL = 'custom/config/cvpr/qual_fast_LR/lyftl5.yaml', # 64 PD4DQL = 'custom/config/cvpr/qual_fast_LR/pd4d.yaml', # 64 GCDPD4DDirQL = 'custom/config/cvpr/qual_fast_LR/pd4d_dir.yaml', # 64 GCDPD4DGradQL = 'custom/config/cvpr/qual_fast_LR/pd4d_grad.yaml', # 64 WaymoQL = 'custom/config/cvpr/qual_fast_LR/waymo.yaml', # 64 ### ROBOTICS DROIDIDQL = 'custom/config/cvpr/qual_fast_LR/droid_ID.yaml', # 64 DROIDOODQL = 'custom/config/cvpr/qual_fast_LR/droid_OOD.yaml', # 64 LBMv12QL = 'custom/config/cvpr/qual_fast_LR/lbmv12.yaml', # 64 ), num_workers=0, # can crash debugger otherwise batch_size=1, ) metrics = dict( gt_source='gt', modes=['r','rpv','apv','ipv'], rgb=dict(), ) ################################################################# cs = ConfigStore.instance() this_config = template_any4d_2b( job, wandb, any4d_config, model, checkpoint, optimizer, scheduler, metrics, dataset_train, dataset_val) # Use the filename (without extension) as the experiment name experiment_name = 'any4d_' + os.path.splitext(os.path.basename(__file__))[0] cs.store( group='experiment', package='_global_', name=experiment_name, node=this_config, )