问题描述:
加载部分预训练好的权重到自己的模型上,fine-tune网络的时候,希望已经具有预训练权重的部分学习速率小一些,随机初始化的新添加的层学习速率大一些。
方法:
用 apply_gradients() 函数。
代码:
import tensorflow as tf
# the variables waiting for optimization
x = tf.Variable(tf.ones([]), name='fast/0')
y = tf.Variable(tf.zeros([]), name='slow/0')
loss = tf.square(x-y)
global_step = tf.Variable(0, name="global_step", trainable=False)
# the optimizer
opt = tf.train.AdamOptimizer(0.01)
# get all gradients
grads_and_vars = opt.compute_gradients(loss, [x, y]) # Return a list of (gradient, variable) pairs.
# Update rate of variables starting with 'fast' is 10 times normal
new_gradients = []
for item in grads_and_vars:
grad, var = item
var_name = var.name
if var_name.startswith('fast'):
print(var_name)
grad = grad*10
new_gradients.append((grad, var))
train_op = opt.apply_gradients(new_gradients, global_step=global_step)
init_op = tf.initialize_all_variables()
sess = tf.Session()
sess.run(init_op)
for i in range(5):
sess.run([train_op, loss, global_step])
print(sess.run([x, y]))