How do I capture weight updates while using tf.train.optimizer

Question

How do I capture weight updates after a train step while using tf.train.optimizer?

One can get gradients, but I could not find any effective way other than finding the difference of weights from their previous copy. It becomes tedious as I have stored weights as a part of a dictionary, key being hidden layer, values being list of weights.

weights = {
'h1': tf.Variable(tf.random_normal([n_input, n_hidden_1], 0, 0.1)),
'h2': tf.Variable(tf.random_normal([n_hidden_1, n_hidden_2], 0, 0.1)),
'out': tf.Variable(tf.random_normal([n_hidden_2, n_classes], 0, 0.1))
}
biases = {
'b1': tf.Variable(tf.random_normal([n_hidden_1], 0, 0.1)),
'b2': tf.Variable(tf.random_normal([n_hidden_2], 0, 0.1)),
'out': tf.Variable(tf.random_normal([n_classes], 0, 0.1))
}

Answer 1

print(sess.run(weights['h1']))

在每个时期的循环结束时。

Answer 2

I think the easiest way is to capture the snapshot of all trainable variables before and after a training step and get the difference:

with tf.Session() as session:
  ...

  def get_snapshot():
    return {var: value for var, value in zip(tf.trainable_variables(), session.run(tf.trainable_variables()))}

  ...

  snapshot_before = get_snapshot()
  session.run([optimizer], feed_dict={x: train_x, y: train_y})
  snapshot_after = get_snapshot()
  for var in tf.trainable_variables():
    print var.name, ' -> ', (snapshot_after[var] - snapshot_before[var])

This can look even nicer if implemented via a context manager . In terms of efficiency, this requires two extra session.run() invocations, however it's much better than evaluating each variable one by one.