scripts/nn/optim/adam.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.
#
#-------------------------------------------------------------

/*
 * Adam optimizer.
 */

update = function(matrix[double] X, matrix[double] dX, double lr, double beta1, double beta2,
                  double epsilon, int t, matrix[double] m, matrix[double] v)
    return (matrix[double] X, matrix[double] m, matrix[double] v) {
  /*
   * Performs an Adam update.
   *
   * Reference:
   *  - Adam: A Method for Stochastic Optimization, Kingma, Ba.
   *    - http://arxiv.org/abs/1412.6980
   *
   * 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.  Recommended value is 0.001.
   *  - beta1: Exponential decay rate for the 1st moment estimates.
   *      Recommended value is 0.9.
   *  - beta2: Exponential decay rate for the 2nd moment estimates.
   *      Recommended value is 0.999.
   *  - epsilon: Smoothing term to avoid divide by zero errors.
   *      Recommended value is 1e-8.
   *  - t: Timestep, starting at 0.
   *  - m: State containing the 1st moment (mean) estimate by
   *      maintaining exponential moving averages of the gradients, of
   *      same shape as `X`.
   *  - v: State containing the 2nd raw moment (uncentered variance)
   *      estimate by maintaining exponential moving averages of the
   *      squared gradients, of same shape as `X`.
   *
   * Outputs:
   *  - X: Updated parameters `X`, of same shape as input `X`.
   *  - m: Updated state containing the 1st moment (mean) estimate by
   *      maintaining exponential moving averages of the gradients, of
   *      same shape as `X`.
   *  - v: Updated state containing the 2nd raw moment (uncentered
   *      variance) estimate by maintaining exponential moving averages
   *      of the squared gradients, of same shape as `X`.
   */
  t = t + 1
  m = beta1*m + (1-beta1)*dX  # update biased 1st moment estimate
  v = beta2*v + (1-beta2)*dX^2  # update biased 2nd raw moment estimate
  # m = m / (1-beta1^t)  # compute bias-corrected 1st moment estimate
  # v = v / (1-beta2^t)  # compute bias-corrected 2nd raw moment estimate
  # X = X - (lr * m / (sqrt(v)+epsilon))  # param update
  # Simplified for computational efficiency:
  lr = lr * sqrt(1-beta2^t) / (1-beta1^t)
  X = X - (lr * m / (sqrt(v)+epsilon))
}

init = function(matrix[double] X)
    return (matrix[double] m, matrix[double] v) {
  /*
   * 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:
   *  - m: Initial state containing the 1st moment (mean) estimate by
   *      maintaining exponential moving averages of the gradients, of
   *      same shape as `X`.
   *  - v: Initial state containing the 2nd raw moment (uncentered
   *      variance) estimate by maintaining exponential moving averages
   *      of the squared gradients, of same shape as `X`.
   */
  m = matrix(0, rows=nrow(X), cols=ncol(X))
  v = matrix(0, rows=nrow(X), cols=ncol(X))
}