Created
October 5, 2017 11:21
-
-
Save InnerPeace-Wu/6a26e00297e89134e3e8532053509d89 to your computer and use it in GitHub Desktop.
ways to do gradients clipping and learning rate decay in tensorflow
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
import tensorflow as tf | |
#aplly exponential decay on learning rate | |
global_step = tf.Variable(0, trainable=False) | |
stater_learning_rate = lr #for start | |
learning_rate = tf.train.exponential_decay(starter_learning_rate, global_step, | |
decay_steps, decay_rate, staircase=True) | |
optimizer = tf.train.AdamOptimizer(learning_rate) | |
#no clipping | |
train_op = optimizer.minimize(loss, global_step=global_step) | |
#global norm clipping. | |
grad_vars = optimizer.compute_gradients(loss) | |
grad = [x[0] for x in grad_vars] | |
vars = [x[1] for x in grad_vars] | |
grad, grad_norm = tf.clip_by_global_norm(grad, max_grad_norm) | |
train_op = optimizer.apply_gradients(zip(grad, vars), global_step=global_step) | |
#clip by value | |
clipped_gvs = [(tf.clip_by_value(grad, min_val, max_val), var) for grad, var in grad_vars] | |
train_op = optimizer.apply_gradients(clipped_gvs, global_step=global_step) |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment