logging:
    project: pixel_titok_video_4
    run_name: TL128-BS16-128_32-LPIPS1.0
    logging_interval: 50

    save_path: out_deploy
    save_step_interval: 1000
    keep_prior_checkpoints: 2 # -1 to keep all

    resume_from_checkpoint: lpips-90k.ckpt
    init_from_checkpoint:

    init_is_latent_ckpt: False

    discard_disc_on_resume: False

model:
    titok:
        temporal_patch_size: 4
        spatial_patch_size: 16

        fsq_levels: [7, 5, 5, 5, 5]

        num_latent_tokens: 128

        encoder_size: base
        decoder_size: base
        exp_residual: False

        lpips_weight: 1.0
        lpips_subsample: 8 # calculate lpips score from lpips_subsample frames in each clip. Use -1 to sample all.

    disc:
        use_disc: True

        disc_start: 90000
        disc_factor: 1.0
        disc_weight: 0.1
        lecam_weight: 0.001 # 0.0 = disabled, 0.001 normal. Disabled for single-pass disc

        every_n: 2 # train disc every n generator steps
        adapt_disc_weight: False

dataset:
    train_dataset: ["hf://datasets/sailvideo/MiraData-v1/**/*.tar", "hf://datasets/sailvideo/webvid10m_resize128/**/*.tar"]
    eval_dataset: "03.tar"
    resolution: 128
    num_frames: 32
    frames_per_second: 8
    workers: 4
    pin_memory: False # uses extra VRAM when enabled.

optimizer:
    titok:
        learning_rate: 1e-4
        beta1: 0.9
        beta2: 0.99
        weight_decay: 1e-4
        warmup_steps: 2000
        end_lr: 1e-5

    disc:
        learning_rate: 1e-4
        beta1: 0.9
        beta2: 0.99
        weight_decay: 1e-4 
        warmup_steps: 0
        end_lr: 1e-5

training:
    torch_compile: True
    seed: 42
    max_grad_norm: 1.0
    
    batch_size: 16
    enable_tf32: True
    precision: bf16-mixed
    train_devices: 1
    accelerator: 'gpu'

    max_steps: 300000
    val_step_interval: 500

    eval_log_codebook: True
    eval_recon_log_num: 4 # should be less than eval_sample_size
    eval_sample_size: 32
    eval_batch_size: 4

    eval_clear_cache: True