Question

长话短说。接下来的代码和平描述了期望的行为，即优化器

修改了变量值

  x = tf.Variable(5.0)
  # x = x.assign(tf.constant(5.0))
  opt = tf.train.MomentumOptimizer(learning_rate=0.0001, momentum=0.9)
  train_op = opt.minimize(x)
  with tf.Session() as sess:
    sess.run(tf.initialize_all_variables())
    sess.run(train_op)
    print(x.eval())            # => 4.9999 - Desired result

虽然在运行中将值赋值给变量但被认为不是梯度值得操作：

  x = tf.Variable(5.0)
  x_ = x.assign(tf.constant(5.0))
  opt = tf.train.MomentumOptimizer(learning_rate=0.0001, momentum=0.9)        
  train_op = opt.minimize(x_)
  with tf.Session() as sess:
    sess.run(tf.initialize_all_variables())
    sess.run(train_op)        # => ERROR: No gradients provided for any variable
    print(x.eval())

有没有办法将变量设置为某个值，但是由于列车运行，还是通过渐变修改了它的值？

编辑：代码修复

Answer 1

这个肮脏的黑客：

  x = tf.Variable(99.0)
  const = tf.constant(5.0)
  x_ = x + tf.stop_gradient(-x) + const # ARGHH
  opt = tf.train.MomentumOptimizer(learning_rate=0.0001, momentum=0.9)
  train = opt.minimize(x_)

  with tf.Session() as sess:
    sess.run(tf.initialize_all_variables())
    print(x_.eval())
    x_original = x.eval()
    sess.run(train)
    print(x.eval() - x_original + const.eval())

如何在赋值操作后将渐变传播到变量中？

1 个答案: