# Robotics (DROID) training with Wan 2.1 Fun 1.3B InP backbone.
# Based on rwm3_robot.py (Cosmos 2B) adapted for Wan following the
# pattern established in examples_vidar/rwm3_drive40h_wan.py.

import os
from hydra.core.config_store import ConfigStore
from custom.experiment.template import template_any4d_wan_1_3b

#################################################################

job = dict(
    project='a4d2',
    group='debug',
    name='wm3_rob_wan',
    prepend_datetime=True,
)

wandb = dict(
    enabled=True,
    entity='tri',
    project='a4d2',
    num_validation_logs=5,
)

any4d_config = dict(
    dataloader='unified_anyact',
    vae='a4d_vae_wan',
    use_wan_backbone=True,
    video_entries=[
        # Wan 2.1 Fun 1.3B InP: 34-ch Any4D stream, reshaped to 36-ch by _to_wan_36ch.
        #   ch [0:16]  rgb latent (noisy target)
        #   ch [16:17] input_mask (1 at given frames, 0 elsewhere)
        #   ch [17:18] output_mask (0 at given frames, 1 elsewhere)
        #   ch [18:34] reference latent (Wan ref slot, built by a4d_vae_wan)
        dict(
            rgb0=(0, 16, 'load', 'load'),
            rgb0_input_mask=(16, 17, 'load', None),
            rgb0_output_mask=(17, 18, 'load', None),
            rgb0_ref=(18, 34, 'load', None),
        ),
        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),
            rgb1_ref=(18, 34, 'copy:rgb0_ref/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),
            rgb2_ref=(18, 34, 'copy:rgb1_ref/load', None),
        ),
    ],
    lowdim_adaln_entries=dict(
        action=(0, 41, 0, 20, 'zero/load'),
        action_input_mask=(0, 41, 20, 21, 'zero/load'),
    ),
    num_views=3,
    video_concat_mode='view',
    video_proj_mode='per_view',
    view_timestep_mode='per_view',
    block_gate_fix=True,
    disable_risky_sharding=True,  # keep final_layer / t_embedder un-sharded to avoid multi-view hangs
    loss_weights=dict(
        rgb0=1.0,
        rgb1=1.0,
        rgb2=1.0,
    ),
    harmonize_streams=True,
    harmonize_frames=True,
    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',
    ),
    action_multiplier=3.0,
    data_train_overrides=dict(
        subsample=None,
        single_sample='random',
    ),
    data_val_overrides=dict(
        frame_stride=1,
        subsample=20,
        single_sample='first',
    ),
    track_metrics=dict(
        rgb0=['psnr', 'ssim'],
        rgb1=['psnr', 'ssim'],
        rgb2=['psnr', 'ssim'],
    ),
    train_visuals_interval=99,
    train_visuals_detail=1,
    val_visuals_detail=1,
    visuals_quality=8,
    viz_input_blacklist=[],
    viz_extra_modes=['anyact1'],
    viz_mask_border_width=2,
    val_num_steps=35,
)

#################################################################

WAN_CKPT_DIR = os.environ.get(
    'WAN_1_3B_INP_CKPT_DIR',
    # S3 default so SageMaker runs work out of the box (mirrors rwm3_robot41d_wan14b.py).
    # Override with WAN_1_3B_INP_CKPT_DIR=checkpoints/Wan2.1-Fun-1.3B-InP for local runs.
    's3://tri-ml-sandbox-16011-us-west-2-datasets/wan/Wan2.1-Fun-1.3B-InP',
)
WAN_VAE_PATH = os.environ.get(
    'WAN_VAE_PATH',
    f'{WAN_CKPT_DIR}/Wan2.1_VAE.pth',
)

model = dict(
    dit_path=WAN_CKPT_DIR,
    text_encoder_path='',
    vae_path=WAN_VAE_PATH,
    fsdp_shard_size=8,
    ############
    run_validation=True,
    skip_first_validation='delay3',
    validation_iter=500,
    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=3.0e-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/anydata/web/debug/droid_ikela3_41.yaml',
        'custom/config/anydata/web/debug/droid_ikela3_41.yaml',
        'custom/config/anydata/web/debug/droid_ikela3_41.yaml',
    ],
    num_workers=2,
    batch_size=2,
)

dataset_val = dict(
    config=dict(
        DROID='custom/config/anydata/web/debug/droid_ikela3_41.yaml',
    ),
    num_workers=0,
    batch_size=1,
)

metrics = None

#################################################################

cs = ConfigStore.instance()

this_config = template_any4d_wan_1_3b(
    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,
)