View source on GitHub
|
Tensor normalizer classses.
These encapsulate variables and function for tensor normalization.
Example usage:
observation = tf.placeholder(tf.float32, shape=[]) tensor_normalizer = StreamingTensorNormalizer( tensor_spec.TensorSpec([], tf.float32), scope='normalize_observation') normalized_observation = tensor_normalizer.normalize(observation) update_normalization = tensor_normalizer.update(observation)
with tf.Session() as sess: for o in observation_list: # Compute normalized observation given current observation vars. normalizedobservation = sess.run( normalized_observation, feed_dict = {observation: o})
# Update normalization params for next normalization op.
sess.run(update_normalization, feed_dict = {observation: o})
# Do something with normalized_observation_
...
Classes
class EMATensorNormalizer: TensorNormalizer with exponential moving avg. mean and var estimates.
class StreamingTensorNormalizer: Normalizes mean & variance based on full history of tensor values.
class TensorNormalizer: Encapsulates tensor normalization and owns normalization variables.