# Created by BVH, Jul 2025. 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', # reference_camera now separate per dataset dataloader='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, # rgb1=1.0, ), harmonize_streams=True, harmonize_frames=True, load_modals=['rgb', 'cams'], # 'language', 'action', ], task_probs=dict( cross_modal=0.0, dyn_view_synth=1.0, forecast=0.0, pose_estimation=0.0, inverse_dynamics=0.0, policy=0.0, world_model=0.0, ), use_views='all', # shuffle_cams2views=True, shuffle_cams2views=False, train_directives=dict( num_pred_views=1, ), val_directives=dict( tasks='dyn_view_synth', num_pred_views=1, remove_cond=True, ), track_metrics=dict( rgb0=['psnr', 'ssim'], # rgb1=['psnr', 'ssim'], ), train_visuals_interval=99, train_visuals_detail=1, # 2, val_visuals_detail=1, # 2, viz_input_blacklist=[], # overridden to show cams val_num_steps=25, ) ################################################################# 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 dit_path=f'{S3_PRETRAINED_PREFIX}/nvidia/Cosmos-Predict2-2B-Video2World/model-480p-10fps.pt', # dit_path='', # dit_path=f's3://tri-ml-sandbox-16011-us-west-2-datasets/sagemaker/cosmos-predict2/a4d2/debug/08-27-03-18_s1b_dvs2_kub/model/iter_000042001_002688064.pt', # dit_path=f's3://tri-ml-sandbox-16011-us-west-2-datasets/sagemaker/cosmos-predict2/a4d2/cvpr/09-20-18-35_2cams/model/iter_000008000_000768000.pt', # dit_path=f's3://tri-ml-sandbox-16011-us-west-2-datasets/sagemaker/cosmos-predict2/a4d2/cvpr/10-01-14-57_s20_rdvs2_mix_b12/model/iter_000026000_019968000.pt', # text_encoder_path=f'{S3_PRETRAINED_PREFIX}/google-t5/t5-11b', text_encoder_path='', vae_path=f'{S3_PRETRAINED_PREFIX}/nvidia/Cosmos-Predict2-2B-Video2World/tokenizer/tokenizer.pth', fsdp_shard_size=8, ############ run_validation=True, validation_iter=500, # frequency / interval of validation runs # max_val_iter=8, # number of steps per validation run # max_iter=120000, # total number of training steps # max_iter=60000, # total number of training steps max_iter=50000, # total number of training steps # 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; = num_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 = [200], warm_up_steps = [1000], cycle_lengths = [model['max_iter']], f_start=[0.01], f_max=[1.0], f_min=[0.05], ) dataset_train = dict( # config='custom/config/train/kub4d_train_2v.yaml', config='custom/config/cvpr/mixed/train_2v.yaml', num_workers=4, # WARNING: 6 or 4 seems to cause hanging more often batch_size=6, ) dataset_val = dict( config=dict( ### DRIVING ArgoVerse2sync = 'custom/config/cvpr/mixed/val_2v/argoverse2sync.yaml', DDADfront = 'custom/config/cvpr/mixed/val_2v/ddad_front.yaml', DDADside = 'custom/config/cvpr/mixed/val_2v/ddad_side.yaml', Waymo = 'custom/config/cvpr/mixed/val_2v/waymo.yaml', LyftL5 = 'custom/config/cvpr/mixed/val_2v/lyftl5.yaml', ### ROBOTICS DROIDCalibID = 'custom/config/cvpr/mixed/val_2v/droidcalib_id.yaml', DROIDCalibOOD = 'custom/config/cvpr/mixed/val_2v/droidcalib_ood.yaml', LBMv12 = 'custom/config/cvpr/mixed/val_2v/lbmv12.yaml', ### 3D DL3DV = 'custom/config/cvpr/mixed/val_2v/dl3dv.yaml', MVImgNet = 'custom/config/cvpr/mixed/val_2v/mvimgnet.yaml', RE10K = 'custom/config/cvpr/mixed/val_2v/re10k.yaml', WildRGBD = 'custom/config/cvpr/mixed/val_2v/wildrgbd.yaml', ScanNet = 'custom/config/cvpr/mixed/val_2v/scannet.yaml', ### 4D Kubric4D = 'custom/config/cvpr/mixed/val_2v/kubric4d.yaml', # Kubric5D = 'custom/config/cvpr/mixed/val_2v/kubric5d.yaml', PD4D = 'custom/config/cvpr/mixed/val_2v/pd4d.yaml', GCDKubric4Dgrad = 'custom/config/cvpr/mixed/val_2v/gcd_kubric4d_grad.yaml', GCDPD4Dgrad = 'custom/config/cvpr/mixed/val_2v/gcd_pd4d_grad.yaml', ### TODO EgoExo4D = 'custom/config/cvpr/mixed/val_2v/egoexo4d.yaml', ), num_workers=0, # can crash debugger otherwise batch_size=1, ) metrics = dict( # gt_source='rec', 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, )