optim_type: adamw
lr_choice: layerwise_decay
lr_schedule: cosine_decay
lr: 0.0001
lr_decay: 0.9
end_lr: 0
lr_mult: 1
weight_decay: 0.001
warmup_steps: 0.1
validation_metric_name: accuracy
custom_metric_func: null
efficient_finetune: null
trainable_param_names: []
mixup_fn: null
mixup_off_epoch: 5
skip_final_val: false