python train_daVinci.py \
    --dataset_names dataset_name\
    --ckpt_dir ./model_ckpts/hl/suturing_hl_3\
    --gpu 0 \
    --recovery_probability 0.6 \
    --batch_size 16 \
    --num_epochs 2000 \
    --lr 4e-4 \
    --min_lr 1e-5 \
    --lr_cycle 25 \
    --warmup_epochs 5 \
    --weight_decay 0.05 \
    --validation_interval 10 \
    --prediction_offset 15 \
    --history_len 4 \
    --save_ckpt_interval 5 \
    --history_step_size 30 \
    --one_hot_flag \
    --early_stopping_interval 300 \
    --seed 5 \
    --plot_val_images_flag \
    --max_num_images 5 \
    --cameras_to_use left_img_dir \
    --backbone_model swin-t \
    --model_init_weights imagenet \
    --image_dim 224 224 \
    --freeze_backbone_until none \
    --multitask_loss_weight 0.6 \
    --uniform_sampling_flag \
    --extra_repeated_phase_last_frame_sampling_flag \
    --extra_repeated_phase_last_frame_sampling_probability 0.15 \
    --add_center_crop_view_flag \
    --global_pool_image_features_flag \
    --dataset_mean_std_file_names "dataset_mean_std_camera_type='left_img_dir'_image_step_size=1.json" \
    --val_split_number 0 \
    --use_complexer_multitask_mlp_head_flag \
    --selected_multitasks dominant_moving_direction 
