scripts/nn/optim/rmsprop.dml

#-------------------------------------------------------------
#
# Licensed to the Apache Software Foundation (ASF) under one
# or more contributor license agreements.  See the NOTICE file
# distributed with this work for additional information
# regarding copyright ownership.  The ASF licenses this file
# to you under the Apache License, Version 2.0 (the
# "License"); you may not use this file except in compliance
# with the License.  You may obtain a copy of the License at
#
#   http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing,
# software distributed under the License is distributed on an
# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
# KIND, either express or implied.  See the License for the
# specific language governing permissions and limitations
# under the License.
#
#-------------------------------------------------------------

/*
 * RMSprop optimizer.
 */

update = function(matrix[double] X, matrix[double] dX, double lr, double decay_rate,
                  double epsilon, matrix[double] cache)
    return (matrix[double] X, matrix[double] cache) {
  /*
   * Performs an RMSprop update.
   *
   * This is an adaptive learning rate optimizer that can be viewed
   * as an adjustment of the Adagrad method to use a moving average
   * of the sum of squared gradients in order to improve convergence.
   *
   * Reference:
   *  - Neural Networks for Machine Learning, Lecture 6a, Hinton,
   *    slide 29.
   *    - http://www.cs.toronto.edu/~tijmen/csc321/slides/lecture_slides_lec6.pdf
   *
   * Inputs:
   *  - X: Parameters to update, of shape (any, any).
   *  - dX: Gradient wrt `X` of a loss function being optimized, of
   *      same shape as `X`.
   *  - lr: Learning rate.
   *  - decay_rate: Term controlling the rate of the moving average.
   *      Typical values are in the range of [0.9, 0.999].
   *  - epsilon: Smoothing term to avoid divide by zero errors.
   *      Typical values are in the range of [1e-8, 1e-4].
   *  - cache: State that maintains the moving average of the squared
   *      gradients, of same shape as `X`.
   *
   * Outputs:
   *  - X: Updated parameters `X`, of same shape as input `X`.
   *  - cache: Updated state that maintains the moving average of the
   *      squared gradients, of same shape as `X`.
   */
  cache = decay_rate*cache + (1-decay_rate)*dX^2
  X = X - (lr * dX / (sqrt(cache)+epsilon))
}

init = function(matrix[double] X)
    return (matrix[double] cache) {
  /*
   * Initialize the state for this optimizer.
   *
   * Note: This is just a convenience function, and state
   * may be initialized manually if needed.
   *
   * Inputs:
   *  - X: Parameters to update, of shape (any, any).
   *
   * Outputs:
   *  - cache: State that maintains the moving average of the squared
   *      gradients, of same shape as `X`.
   */
  cache = matrix(0, rows=nrow(X), cols=ncol(X))
}