# Created by BVH, Apr 2026. # 11-view driving world model (AnyDrive). import os from hydra.core.config_store import ConfigStore from custom.experiment.template import template_any4d_2b, S3_PRETRAINED_ROOT, S3_OUTPUT_ROOT ################################################################# job = dict( project='a4d2', group='example', name='wm11_gaia', prepend_datetime=True, local_root='', s3_root=S3_OUTPUT_ROOT, ) wandb = dict( enabled=True, entity='tri', project='a4d2', num_validation_logs=5, ) any4d_config = dict( dataloader='unified_anydrive', vae='a4d_vae', video_entries=[ # first entry = base / pretrained dict( rgb0=(0, 16, 'load', 'load'), rgb0_input_mask=(16, 17, 'load', None), rgb0_output_mask=(17, 18, 'load', None), ), # later entries = new viewpoints dict( 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, 'copy:rgb0_output_mask/load', None), ), dict( rgb2=(0, 16, 'copy:rgb1/load', 'copy:rgb1/load'), rgb2_input_mask=(16, 17, 'copy:rgb1_input_mask/load', None), rgb2_output_mask=(17, 18, 'copy:rgb1_output_mask/load', None), ), # copy assuming we resume from 3-view ckpt dict( rgb3=(0, 16, 'copy:rgb2/load', 'copy:rgb2/load'), rgb3_input_mask=(16, 17, 'copy:rgb2_input_mask/load', None), rgb3_output_mask=(17, 18, 'copy:rgb2_output_mask/load', None), ), dict( rgb4=(0, 16, 'copy:rgb2/load', 'copy:rgb2/load'), rgb4_input_mask=(16, 17, 'copy:rgb2_input_mask/load', None), rgb4_output_mask=(17, 18, 'copy:rgb2_output_mask/load', None), ), # OLD: copy assuming we resume from 5-view ckpt # NEW: copy assuming we resume from 3-view ckpt dict( rgb5=(0, 16, 'copy:rgb2/load', 'copy:rgb2/load'), rgb5_input_mask=(16, 17, 'copy:rgb2_input_mask/load', None), rgb5_output_mask=(17, 18, 'copy:rgb2_output_mask/load', None), ), dict( rgb6=(0, 16, 'copy:rgb2/load', 'copy:rgb2/load'), rgb6_input_mask=(16, 17, 'copy:rgb2_input_mask/load', None), rgb6_output_mask=(17, 18, 'copy:rgb2_output_mask/load', None), ), dict( rgb7=(0, 16, 'copy:rgb2/load', 'copy:rgb2/load'), rgb7_input_mask=(16, 17, 'copy:rgb2_input_mask/load', None), rgb7_output_mask=(17, 18, 'copy:rgb2_output_mask/load', None), ), dict( rgb8=(0, 16, 'copy:rgb2/load', 'copy:rgb2/load'), rgb8_input_mask=(16, 17, 'copy:rgb2_input_mask/load', None), rgb8_output_mask=(17, 18, 'copy:rgb2_output_mask/load', None), ), dict( rgb9=(0, 16, 'copy:rgb2/load', 'copy:rgb2/load'), rgb9_input_mask=(16, 17, 'copy:rgb2_input_mask/load', None), rgb9_output_mask=(17, 18, 'copy:rgb2_output_mask/load', None), ), dict( rgb10=(0, 16, 'copy:rgb2/load', 'copy:rgb2/load'), rgb10_input_mask=(16, 17, 'copy:rgb2_input_mask/load', None), rgb10_output_mask=(17, 18, 'copy:rgb2_output_mask/load', None), ), ], lowdim_adaln_entries=dict( # traj=(0, 41, 0, 2, 'zero/load'), # traj_input_mask=(0, 41, 2, 3, 'zero/load'), # NOTE(bvh): for MVP0 we want to demonstrate ~5 seconds (one shot) rollout traj=(0, 53, 0, 2, 'zero/load'), traj_input_mask=(0, 53, 2, 3, 'zero/load'), ), num_views=11, video_concat_mode='view', video_proj_mode='per_view', view_timestep_mode='per_view', block_gate_fix=True, train_cond_aug_sigma_range=(0.001, 0.01), val_cond_aug_sigma=0.001, loss_weights=dict( rgb0=1.0, rgb1=1.0, rgb2=1.0, rgb3=1.0, rgb4=1.0, rgb5=1.0, rgb6=1.0, rgb7=1.0, rgb8=1.0, rgb9=1.0, rgb10=1.0, ), harmonize_streams=True, harmonize_frames=True, data_train_overrides=dict( subsample=None, single_sample='random', ), data_val_overrides=dict( frame_stride=1, subsample=20, single_sample='first', ), use_views='all', shuffle_cams2views=False, task_probs=dict( cross_modal=0.0, dyn_view_synth=0.0, forecast=1.0, pose_est=0.0, inv_dyn=0.0, policy=0.0, world_model=0.9, ), train_directives=dict( ), val_directives=dict( tasks='forecast,world_model', perturb_traj='basile3', ), traj_multiplier=1.0, track_metrics=dict( rgb0=['psnr', 'ssim', 'lpips'], rgb1=['psnr', 'ssim', 'lpips'], rgb2=['psnr', 'ssim', 'lpips'], rgb3=['psnr', 'ssim', 'lpips'], rgb4=['psnr', 'ssim', 'lpips'], rgb5=['psnr', 'ssim', 'lpips'], rgb6=['psnr', 'ssim', 'lpips'], rgb7=['psnr', 'ssim', 'lpips'], rgb8=['psnr', 'ssim', 'lpips'], rgb9=['psnr', 'ssim', 'lpips'], rgb10=['psnr', 'ssim', 'lpips'], ), train_visuals_interval=99, train_visuals_detail=1, val_visuals_detail=1, visuals_quality=8, viz_input_blacklist=[], viz_extra_modes=['anydrive1', 'gaia1'], viz_mask_border_width=2, val_num_steps=35, ) ################################################################# model = dict( # pretrained / vanilla cosmos: # dit_path=f'{S3_PRETRAINED_ROOT}/nvidia/Cosmos-Predict2-2B-Video2World/model-480p-10fps.pt', # AnyDrive HR 1 (vidar): # dit_path='s3://tri-ml-sandbox-16011-us-west-2-datasets/sagemaker/cosmos-predict2/a4d2/debug/12-11-01-56_s27e_rwm3_drive40h/model/iter_000049000_012544000.pt', # AnyDrive HR 2 (vidar): dit_path='s3://tri-ml-sandbox-16011-us-west-2-datasets/sagemaker/cosmos-predict2/a4d2/debug/12-26-14-01_s27g_rwm3_drive40h/model/iter_000078000_019968000.pt', # GAIA 11-view (sa37d, only 5-iter save -- effectively untrained): # dit_path='s3://gaia-e2e-wfm-datasets/any4d/a4d2/sm_ad2/04-28-22-16_sa37d_rwm11_gaia25m_b2/model/iter_000000005_000000320.pt', text_encoder_path='', vae_path=f'{S3_PRETRAINED_ROOT}/nvidia/Cosmos-Predict2-2B-Video2World/tokenizer/tokenizer.pth', fsdp_shard_size=8, ############ run_validation=True, skip_first_validation='delay3', validation_iter=500, max_val_iter=3, max_iter=60000, grad_accum_iter=1, context_parallel_size=1, device_monitor=0, manual_gc_iter=288, manual_gc_warm_up=-1, ############ state_t=11, tokenizer_chunk_duration=41, ) checkpoint = dict( save_iter=1000, early_sanity_check=5, ) optimizer = dict( lr=1e-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/anydata/web/debug/gaia_iked11_41.yaml', # 'custom/config/anydata/web/train/gaia_pvm_ike5.yaml', # 'custom/config/anydata/web/train/gaia_wide_ike5.yaml', 'custom/config/anydata/web/train/gaia_ike11.yaml', ], num_workers=4, batch_size=2, ) dataset_val = dict( config=dict( # GAIA = 'custom/config/anydata/web/debug/gaia_iked11_41.yaml', # GAIA = 'custom/config/anydata/web/val/gaia_pvm_ike5.yaml', # GAIA = 'custom/config/anydata/web/val/gaia_wide_ike5.yaml', GAIA = 'custom/config/anydata/web/val/gaia_ike11.yaml', ), num_workers=0, batch_size=1, ) metrics = None ################################################################# cs = ConfigStore.instance() this_config = template_any4d_2b( job, wandb, any4d_config, model, checkpoint, optimizer, scheduler, metrics, dataset_train, dataset_val) experiment_name = 'any4d_' + os.path.splitext(os.path.basename(__file__))[0] cs.store( group='experiment', package='_global_', name=experiment_name, node=this_config, )