# horovodrun -np 1 -H localhost:1 \ # python train_re.py --config config/hps-refcoco+.json # horovodrun -np 2 -H localhost:2 \ # python train_re.py --config config/hps-refcoco+.json horovodrun -np 1 -H localhost:1 \ python train_re.py \ --train_txt_db /db/refcoco+_train_base-cased.db \ --train_img_dir /img/visual_grounding_coco_gt \ --val_txt_db /db/refcoco+_val_base-cased.db \ --val_img_dir /img/visual_grounding_coco_gt \ --checkpoint /pretrain/bert-base_weak/ckpt/model_step_420000.pt \ --cut_bert -1 \ --output_dir /storage/refcoco+/bert-base_mlm+itm+mrfr_pretrain-refcoco+_lr1e-4 \ --max_txt_len 60 \ --train_batch_size 128 \ --val_batch_size 128 \ --learning_rate 1e-4 \ --optim adamw \ --betas 0.9 0.98 \ --weight_decay 0.01 \ --dropout 0.1 \ --grad_norm 2.0 \ --decay linear \ --num_train_steps 24000 \ --warmup_steps 1500 \ --gradient_accumulation_steps 1 \ --seed 24 \ --fp16