output_model_path="./output_flickr" lr_scheduler="manual_warmup_decay" decay_steps="54360;72480" num_train_steps=90600 SAVE_STEPS=4530 WARMUP_STEPS=9060 BATCH_SIZE=4 LR_RATE=1e-5 WEIGHT_DECAY=0.01 MAX_LEN=48 hardest=False meansum=False use_circle_loss=True scale_circle=32.0 use_sigmoid=True margin=0.3