mirror of
https://github.com/zebrajr/pytorch.git
synced 2025-12-07 12:21:27 +01:00
7f1a96d43c
Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/38574 Adding sparse L1 and L2 regularization operator to Caffe2. This doesn't work using run_on_loss, only run_after_optimize. Applying it to run_after_optimize rather than run_on_loss was easier to implement, particularly for the L1 norm which is preferable in some cases and is non-differentiable at zero. Test Plan: Wrote and ran unit tests in operator_test:sparse_lp_regularizer_test. Differential Revision: D21003029 fbshipit-source-id: 81070a621752560ce03e320d065ce27807a5d278
44 lines
1.1 KiB
C++
44 lines
1.1 KiB
C++
#pragma once
|
|
|
|
#include "caffe2/core/operator.h"
|
|
#include "caffe2/utils/math.h"
|
|
|
|
namespace caffe2 {
|
|
|
|
template <typename T, class Context>
|
|
class CAFFE2_API SparseLpRegularizerOp final : public Operator<Context> {
|
|
public:
|
|
USE_OPERATOR_CONTEXT_FUNCTIONS;
|
|
template <class... Args>
|
|
explicit SparseLpRegularizerOp(Args&&... args)
|
|
: Operator<Context>(std::forward<Args>(args)...),
|
|
p_(this->template GetSingleArgument<float>("p", 2.0)),
|
|
reg_lambda_(
|
|
this->template GetSingleArgument<float>("reg_lambda", 1e-5)) {
|
|
CAFFE_ENFORCE(
|
|
p_ == 1.0 || p_ == 2.0,
|
|
"Sparse Lp regularizer only implemented for p=1 or p=2.");
|
|
CAFFE_ENFORCE_GT(
|
|
reg_lambda_,
|
|
0.0,
|
|
"Lambda for sparse Lp regularizer must be greater than 0.");
|
|
CAFFE_ENFORCE_LT(
|
|
reg_lambda_,
|
|
1.0,
|
|
"Lambda for sparse Lp regularizer must be less than 1.");
|
|
}
|
|
|
|
bool RunOnDevice() override;
|
|
|
|
template <typename SIndex>
|
|
bool DoRunWithType();
|
|
|
|
protected:
|
|
float p_;
|
|
float reg_lambda_;
|
|
INPUT_TAGS(PARAM, INDICES);
|
|
OUTPUT_TAGS(OUTPUT_PARAM);
|
|
};
|
|
|
|
} // namespace caffe2
|