output_model_path="output_refcoco_plus" lr_scheduler="manual_warmup_decay" decay_steps="3290;4700;7050" lr_decay_ratio=0.2 num_train_steps=26640 SAVE_STEPS=470 WARMUP_STEPS=940 BATCH_SIZE=32 VALID_STEPS=20000 LR_RATE=2e-5 WEIGHT_DECAY=0.01 MAX_LEN=80 layer_decay_rate=0.9