configs/imagenet.yaml

wandb:
    entity: null

experiment:
    project: "muse"
    name: "imagenet"
    output_dir: "imagenet"
    max_train_examples: 1281167 # total number of imagenet examples
    max_eval_examples: 12800
    save_every: 1000
    eval_every: 500
    generate_every: 1000
    log_every: 50
    log_grad_norm_every: 100
    resume_from_checkpoint: False
    resume_lr_scheduler: True

model:
    vq_model:
        type: "maskgit_vqgan"
        pretrained: "openMUSE/maskgit-vqgan-imagenet-f16-256"

    transformer:
        vocab_size: 2048 # (1024 + 1000 + 1 = 2025 -> Vq + Imagenet + <mask>, use 2048 for even division by 8)
        max_position_embeddings: 264 # (256 + 1 for class id, use 264 for even division by 8)
        hidden_size: 768
        num_hidden_layers: 24
        num_attention_heads: 16
        intermediate_size: 3072
        codebook_size: 1024
        num_vq_tokens: 256
        num_classes: 1000
        initializer_range: 0.02
        norm_type: "layernorm"
        layer_norm_eps: 1e-6
        use_normformer: True
        use_encoder_layernorm: True
        use_mlm_layer: True
        use_mlm_layernorm: True
        use_bias: False
        hidden_dropout: 0.0
        attention_dropout: 0.0

    gradient_checkpointing: True
    enable_xformers_memory_efficient_attention: True


dataset:
    params:
        train_shards_path_or_url: "pipe:aws s3 cp s3://muse-datasets/imagenet-wds/imagenet-train-{000000..000320}.tar -"
        eval_shards_path_or_url: "pipe:aws s3 cp s3://muse-datasets/imagenet-wds/imagenet-val-{000000..000012}.tar -"
        batch_size: ${training.batch_size}
        shuffle_buffer_size: 1000
        num_workers: 4
        resolution: 256
        pin_memory: True
        persistent_workers: True
    preprocessing:
        resolution: 256
        center_crop: True
        random_flip: False


optimizer:
    name: fused_adamw 
    params: # default adamw params
        learning_rate: 1e-4
        scale_lr: False # scale learning rate by total batch size
        beta1: 0.9
        beta2: 0.999
        weight_decay: 0.01
        epsilon: 1e-8


lr_scheduler:
    scheduler: "constant_with_warmup"
    params:
        learning_rate: ${optimizer.params.learning_rate}
        warmup_steps: 1000


training:
    gradient_accumulation_steps: 1
    batch_size: 192 # on 80GB A100
    mixed_precision: "bf16"
    enable_tf32: True
    use_ema: False
    seed: 42
    max_train_steps: 50000
    overfit_one_batch: False
    min_masking_rate: 0.0
    label_smoothing: 0.0
    max_grad_norm: null
    # related to vae code sampling
    use_soft_code_target: False
    use_stochastic_code: False
    soft_code_temp: 1.0