# Network MODEL_USE: ban GLIMPSE: 4 HIDDEN_SIZE: 1024 K_TIMES: 3 DROPOUT_R: 0.2 CLASSIFER_DROPOUT_R: 0.5 FLAT_OUT_SIZE: 2048 USE_BBOX_FEAT: True BBOXFEAT_EMB_SIZE: 1024 IMG_FEAT_SIZE: 1024 # Execution BATCH_SIZE: 512 LR_BASE: 0.002 LR_DECAY_R: 0.25 LR_DECAY_LIST: [10, 12] WARMUP_EPOCH: 3 MAX_EPOCH: 13 GRAD_NORM_CLIP: 0.25 GRAD_ACCU_STEPS: 8 LOSS_FUNC: ce LOSS_REDUCTION: sum OPT: Adamax OPT_PARAMS: {betas: '(0.9, 0.999)', eps: '1e-9'}