#-------------------------------------------------------------
#
# 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.
#
#-------------------------------------------------------------

/*
 * 2D Scale & Shift layer.
 */
source("scripts/nn/util.dml") as util

forward = function(matrix[double] X, matrix[double] gamma, matrix[double] beta,
                   int C, int Hin, int Win)
    return (matrix[double] out) {
  /*
   * Computes the forward pass for a 2D scale & shift layer.  The input
   * data has N examples, each represented as a 3D volume unrolled into
   * a single vector.
   *
   * A 2D scale & shift layer introduces learnable parameters
   * (gamma, beta) to scale and shift the input on a per-channel basis.
   *
   *   `y = x*gamma + beta`
   *
   * Inputs:
   *  - X: Inputs, of shape (N, C*Hin*Win).
   *  - gamma: Scale parameters, of shape (C, 1).
   *  - beta: Shift parameters, of shape (C, 1).
   *  - C: Number of input channels (dimensionality of input depth).
   *  - Hin: Input height.
   *  - Win: Input width.
   *
   * Outputs:
   *  - out: Outputs, of shape (N, C*Hin*Win).
   */
  # Scale and shift
  scaled = bias_multiply(X, gamma)  # shape (N, C*Hin*Win)
  out = bias_add(scaled, beta)  # shape (N, C*Hin*Win)
}

backward = function(matrix[double] dout, matrix[double] out,
                    matrix[double] X, matrix[double] gamma, matrix[double] beta,
                    int C, int Hin, int Win)
      return (matrix[double] dX, matrix[double] dgamma, matrix[double] dbeta) {
  /*
   * Computes the backward pass for a 2D scale & shift layer.
   *
   * Inputs:
   *  - dout: Gradient wrt `out` from upstream, of shape (N, C*Hin*Win).
   *  - out: Outputs from the forward pass, of shape (N, C*Hin*Win).
   *  - X: Input data matrix to the forward pass, of
   *      shape (N, C*Hin*Win).
   *  - gamma: Scale parameters, of shape (C, 1).
   *  - beta: Shift parameters, of shape (C, 1).
   *  - C: Number of input channels (dimensionality of input depth).
   *  - Hin: Input height.
   *  - Win: Input width.
   *
   * Outputs:
   *  - dX: Gradient wrt `X`, of shape (N, C*Hin*Win).
   *  - dgamma: Gradient wrt `W`, of shape (C, 1).
   *  - dbeta: Gradient wrt `b`, of shape (C, 1).
   *
   */
  # Compute gradients during training
  dgamma = util::channel_sums(dout*X, C, Hin, Win)  # shape (C, 1)
  dbeta = util::channel_sums(dout, C, Hin, Win)  # shape (C, 1)
  dX = bias_multiply(dout, gamma)  # shape (N, C*Hin*Win)
}

init = function(int C)
    return (matrix[double] gamma, matrix[double] beta) {
  /*
   * Initialize the parameters of this layer.
   *
   * By default, we initialize to an identity function, with a scale
   * filler of `1`, and a shift filler of `0`.
   *
   * Note: This is just a convenience function, and parameters
   * may be initialized manually if needed.
   *
   * Inputs:
   *  - C: Number of input channels (dimensionality of input depth).
   *
   * Outputs:
   *  - gamma: Scale parameters, of shape (C, 1).
   *  - beta: Shift parameters, of shape (C, 1).
   */
   gamma = matrix(1, rows=C, cols=1)
   beta = matrix(0, rows=C, cols=1)
}