output_model_path="output_hm"
lr_scheduler="manual_warmup_decay"
decay_steps="13308;19962"
lr_decay_ratio=0.1
num_train_steps=5000
SAVE_STEPS=1250
WARMUP_STEPS=500
BATCH_SIZE=8
VALID_STEPS=20000
LR_RATE=1e-5
WEIGHT_DECAY=0.01
MAX_LEN=128