output_model_path="output_vcr"
lr_scheduler="manual_warmup_decay"
decay_steps="13308;19962"
lr_decay_ratio=0.1
num_train_steps=200
SAVE_STEPS=100
WARMUP_STEPS=6654
BATCH_SIZE=8
VALID_STEPS=20000
LR_RATE=2e-5
WEIGHT_DECAY=0.01
MAX_LEN=80
