# Created by BVH, Jul 2025. 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( # this becomes JobConfig project='a4d2', group='debug', name='every4', prepend_datetime=True, # becomes {now:%m-%d-%H-%M}_{config.job.name} local_root='', # empty = _DEFAULT_LOCAL_ROOT (/any4d on DGX, /tmp/a4d2 on SM) s3_root=S3_OUTPUT_ROOT, # default from template.py; set '' to disable S3 sync ) wandb = dict( enabled=True, entity='tri', project='a4d2', num_validation_logs=5, ) any4d_config = dict( # this is part of Any4DConfig, which inherits from Predict2Video2WorldModelConfig transforms='zego', # best for driving (supported by newer webdatasets only) 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 depth0=(18, 34, 'zero/load', 'rand/load'), depth0_input_mask=(34, 35, 'zero/load', None), depth0_output_mask=(35, 36, 'zero', None), cams0=(36, 68, 'zero/load', 'rand/load'), cams0_input_mask=(68, 69, 'zero/load', None), cams0_output_mask=(69, 70, '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), depth1=(18, 34, 'zero/load', 'rand/load'), depth1_input_mask=(34, 35, 'zero/load', None), depth1_output_mask=(35, 36, 'zero', None), cams1=(36, 68, 'zero/load', 'rand/load'), cams1_input_mask=(68, 69, 'zero/load', None), cams1_output_mask=(69, 70, 'zero', None), ), dict( # key: (channel_start, channel_end, in_proj_init, out_proj_init) rgb2=(0, 16, 'copy:rgb0/load', 'copy:rgb0/load'), rgb2_input_mask=(16, 17, 'copy:rgb0_input_mask/load', None), rgb2_output_mask=(17, 18, 'zero', None), depth2=(18, 34, 'zero/load', 'rand/load'), depth2_input_mask=(34, 35, 'zero/load', None), depth2_output_mask=(35, 36, 'zero', None), cams2=(36, 68, 'zero/load', 'rand/load'), cams2_input_mask=(68, 69, 'zero/load', None), cams2_output_mask=(69, 70, 'zero', None), ), dict( # key: (channel_start, channel_end, in_proj_init, out_proj_init) rgb3=(0, 16, 'copy:rgb0/load', 'copy:rgb0/load'), rgb3_input_mask=(16, 17, 'copy:rgb0_input_mask/load', None), rgb3_output_mask=(17, 18, 'zero', None), depth3=(18, 34, 'zero/load', 'rand/load'), depth3_input_mask=(34, 35, 'zero/load', None), depth3_output_mask=(35, 36, 'zero', None), cams3=(36, 68, 'zero/load', 'rand/load'), cams3_input_mask=(68, 69, 'zero/load', None), cams3_output_mask=(69, 70, 'zero', None), ), ], num_views=4, video_concat_mode='view', video_proj_mode='per_view', loss_weights=dict( rgb0=1.0, depth0=1.0, cams0=1.0, rgb1=1.0, depth1=1.0, cams1=1.0, rgb2=1.0, depth2=1.0, cams2=1.0, rgb3=1.0, depth3=1.0, cams3=1.0, ), harmonize_streams=True, harmonize_frames=True, load_modals=['rgb', 'language', 'depth', 'cams'], # , 'traj'], task_probs=dict( cross_modal=0.5, dyn_view_synth=0.5, forecast=0.5, pose_estimation=0.5, ), use_views='rand', track_metrics=dict( rgb0=['psnr', 'ssim'], depth0=['mse', 'mae'], cams0=['mse', 'mae'], rgb1=['psnr', 'ssim'], depth1=['mse', 'mae'], cams1=['mse', 'mae'], rgb2=['psnr', 'ssim'], depth2=['mse', 'mae'], cams2=['mse', 'mae'], rgb3=['psnr', 'ssim'], depth3=['mse', 'mae'], cams3=['mse', 'mae'], ), train_visuals_interval=99, train_visuals_detail=2, val_visuals_detail=2, viz_input_blacklist=[], # overridden to show cams ) ################################################################# 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', text_encoder_path=f'{S3_PRETRAINED_PREFIX}/google-t5/t5-11b', vae_path=f'{S3_PRETRAINED_PREFIX}/nvidia/Cosmos-Predict2-2B-Video2World/tokenizer/tokenizer.pth', fsdp_shard_size=8, ############ run_validation=True, validation_iter=300, # frequency / interval of validation runs max_val_iter=16, # number of steps per validation run max_iter=150000, # total number of training steps grad_accum_iter=2, 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 = [100], cycle_lengths = [model['max_iter']], f_start=[0.01], f_max=[1.0], f_min=[0.05], ) dataset_train = dict( config='custom/config/train/ddad_ego_all_4v.yaml', num_workers=6, batch_size=4, ) dataset_val = dict( config='custom/config/val/ddad_ego_some_4v.yaml', num_workers=0, # can crash debugger otherwise batch_size=1, ) metrics = 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, )