learning rate warmup实现
def noam_scheme(global_step, num_warmup_steps, num_train_steps, init_lr, warmup=True): """ decay learning rate if warmup > global step, the learning rate will be global_step/num_warmup_steps * init_lr if warmup < global step, the learning rate will be polynomial decay :param global_step: global steps :param num_warmup_steps: number of warm up steps :param num_train_steps: number of train steps :param init_lr: initial learning rate :param warmup: if True, it will warm up learning rate :return: learning rate """ learning_rate = tf.constant(value=init_lr, shape=[], dtype=tf.float32) learning_rate = tf.train.polynomial_decay(learning_rate, global_step, num_train_steps, end_learning_rate=0.0, power=1.0, cycle=False) if warmup: global_steps_int = tf.cast(global_step, tf.int32) warmup_steps_int = tf.constant(num_warmup_steps, dtype=tf.int32) global_steps_float = tf.cast(global_steps_int, tf.float32) warmup_steps_float = tf.cast(warmup_steps_int, tf.float32) warmup_percent_done = global_steps_float / warmup_steps_float warmup_learning_rate = init_lr * warmup_percent_done is_warmup = tf.cast(global_steps_int < warmup_steps_int, tf.float32) learning_rate = ((1.0 - is_warmup) * learning_rate + is_warmup * warmup_learning_rate) return learning_rate