[GitHub] eric-haibin-lin commented on a change in pull request #9540: refactor regression ops to nnvm interface

[GitBox <gi...@apache.org>]

eric-haibin-lin commented on a change in pull request #9540: refactor regression ops to nnvm interface
URL: https://github.com/apache/incubator-mxnet/pull/9540#discussion_r163678720
 
 

 ##########
 File path: src/operator/regression_output.cc
 ##########
 @@ -18,86 +18,115 @@
  */
 
 /*!
- * Copyright (c) 2015 by Contributors
- * \file regression_output.cc
- * \brief regression output operator
+ * \file regression_ouput.cc
+ * \brief Regression output operator.
 */
+
 #include "./regression_output-inl.h"
-#include "./mshadow_op.h"
 
 namespace mxnet {
 namespace op {
 
-template<>
-Operator *CreateRegressionOutputOp<cpu>(reg_enum::RegressionOutputType type,
-                                        RegressionOutputParam param) {
-  switch (type) {
-    case reg_enum::kLinear:
-      return new RegressionOutputOp<cpu, op::mshadow_op::identity, op::mshadow_op::minus>(param);
-    case reg_enum::kLogistic:
-      return new RegressionOutputOp<cpu, mshadow_op::sigmoid, op::mshadow_op::minus>(param);
-    case reg_enum::kMAE:
-      return new RegressionOutputOp<cpu, op::mshadow_op::identity, mshadow_op::minus_sign>(param);
-    default:
-      LOG(FATAL) << "unknown activation type " << type;
-  }
-  return nullptr;
-}
-
-// DO_BIND_DISPATCH comes from operator_common.h
-template<reg_enum::RegressionOutputType type>
-Operator *RegressionOutputProp<type>::CreateOperator(Context ctx) const {
-  DO_BIND_DISPATCH(CreateRegressionOutputOp, type, param_);
-}
 
 DMLC_REGISTER_PARAMETER(RegressionOutputParam);
 
-MXNET_REGISTER_OP_PROPERTY(LinearRegressionOutput, RegressionOutputProp<reg_enum::kLinear>)
+NNVM_REGISTER_OP(LinearRegressionOutput)
 .describe(R"code(Computes and optimizes for squared loss during backward propagation.
 Just outputs ``data`` during forward propagation.
 
 If :math:`\hat{y}_i` is the predicted value of the i-th sample, and :math:`y_i` is the corresponding target value,
 then the squared loss estimated over :math:`n` samples is defined as
 
-:math:`\text{SquaredLoss}(y, \hat{y} ) = \frac{1}{n} \sum_{i=0}^{n-1} \left( y_i - \hat{y}_i \right)^2`
+:math:`\text{SquaredLoss}(\textbf{Y}, \hat{\textbf{Y}} ) = \frac{1}{n} \sum_{i=0}^{n-1} \lVert  \textbf{y}_i - \hat{\textbf{y}}_i  \rVert_2`
 
 .. note::
    Use the LinearRegressionOutput as the final output layer of a net.
 
-By default, gradients of this loss function are scaled by factor `1/n`, where n is the number of training examples.
-The parameter `grad_scale` can be used to change this scale to `grad_scale/n`.
+By default, gradients of this loss function are scaled by factor `1/m`, where m is the number of dimensions of a training example.
 
 Review comment:
   `number of dimensions ` is a bit confusing to me. For shape `(3,4)`, my first thought of num dimension is two (2-D shape). Is "number of features" clearer?

----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
 
For queries about this service, please contact Infrastructure at:
users@infra.apache.org


With regards,
Apache Git Services