Add hinge loss op (#5837)

* Add hinge loss op

* Update hinge-loss equation for proper latex

paddle/operators/hinge_loss_op.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed 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. */
+
+#include "paddle/operators/hinge_loss_op.h"
+
+namespace paddle {
+namespace operators {
+
+class HingeLossOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("Logits"),
+                   "Input(Logits) must be initialized.");
+    PADDLE_ENFORCE(ctx->HasInput("Labels"),
+                   "Input(Labels) must be initialized.");
+
+    auto pred_dims = ctx->GetInputDim("Logits");
+    auto label_dims = ctx->GetInputDim("Labels");
+
+    PADDLE_ENFORCE_EQ(pred_dims, label_dims);
+    PADDLE_ENFORCE_EQ(pred_dims.size(), 2,
+                      "The rank of Input(Logits) must be 2 and the shape is "
+                      "[batch_size, 1].");
+    PADDLE_ENFORCE_EQ(pred_dims[1], 1,
+                      "Each row of Input(Logits) contains a real value, "
+                      "so the 2nd dimension of Input(Logits) must be 1.");
+
+    ctx->SetOutputDim("Loss", {pred_dims[0], 1});
+    ctx->ShareLoD("Logits", "Loss");
+  }
+};
+
+template <typename AttrType>
+class HingeLossOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  HingeLossOpMaker(framework::OpProto* proto,
+                   framework::OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("Logits",
+             "The input value (Logits) of Hinge loss op."
+             "Logits is a 2-D tensor with shape [batch_size, 1].");
+    AddInput("Labels",
+             "The target value (Labels) of Hinge loss op."
+             "Labels is a 2-D tensor with shape [batch_size, 1].");
+    AddOutput("Loss",
+              "The output tensor with shape [batch_size, 1] "
+              "which represents the hinge loss.");
+    AddComment(R"DOC(
+HingeLoss Operator.
+
+Let x be a logit (prediction) and y be the actual label. The logit can
+take any values from (-inf, inf), but the labels should be either -1 or 1.
+Then, the hinge loss is computed as follows:
+
+$$
+L_(x, y) = max(1 - y.x, 0) 
+$$
+
+Note that the labels passed as input will have values as either 0 or 1.
+
+)DOC");
+  }
+};
+
+class HingeLossGradOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("Logits"),
+                   "Input(Logits) should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Labels"),
+                   "Input(Labels) should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Loss")),
+                   "Input(Loss@GRAD) should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("Logits")),
+                   "Input(Logits@GRAD) should not be null.");
+
+    auto pred_dims = ctx->GetInputDim("Logits");
+    auto lab_dims = ctx->GetInputDim("Labels");
+    auto loss_grad_dims = ctx->GetInputDim(framework::GradVarName("Loss"));
+
+    PADDLE_ENFORCE_EQ(loss_grad_dims, pred_dims);
+
+    auto pred_grad_name = framework::GradVarName("Logits");
+    ctx->SetOutputDim(pred_grad_name, pred_dims);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP(hinge_loss, ops::HingeLossOp, ops::HingeLossOpMaker<float>,
+            hinge_loss_grad, ops::HingeLossGradOp);
+REGISTER_OP_CPU_KERNEL(hinge_loss,
+                       ops::HingeLossKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(
+    hinge_loss_grad,
+    ops::HingeLossGradKernel<paddle::platform::CPUPlace, float>);

paddle/operators/hinge_loss_op.cu

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed 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. */
+
+#define EIGEN_USE_GPU
+#include "paddle/operators/hinge_loss_op.h"
+
+namespace ops = paddle::operators;
+REGISTER_OP_GPU_KERNEL(hinge_loss,
+                       ops::HingeLossKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_GPU_KERNEL(
+    hinge_loss_grad,
+    ops::HingeLossGradKernel<paddle::platform::GPUPlace, float>);

paddle/operators/hinge_loss_op.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed 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. */
+
+#pragma once
+#include "paddle/framework/eigen.h"
+#include "paddle/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename Place, typename T, typename AttrType = T>
+class HingeLossKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* pred = context.Input<framework::Tensor>("Logits");
+    auto* label = context.Input<framework::Tensor>("Labels");
+    auto* loss = context.Output<framework::Tensor>("Loss");
+    auto place = context.GetEigenDevice<Place>();
+
+    auto x = framework::EigenVector<T>::Flatten(*pred);
+    auto y = framework::EigenVector<T>::Flatten(*label);
+    loss->mutable_data<T>(context.GetPlace());
+    auto l = framework::EigenVector<T>::Flatten(*loss);
+    l.device(place) =
+        (static_cast<T>(1) - x * (static_cast<T>(2) * y - static_cast<T>(1)))
+            .cwiseMax(static_cast<T>(0));
+  }
+};
+
+template <typename Place, typename T, typename AttrType = T>
+class HingeLossGradKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* pred = context.Input<framework::Tensor>("Logits");
+    auto* label = context.Input<framework::Tensor>("Labels");
+    auto* dloss =
+        context.Input<framework::Tensor>(framework::GradVarName("Loss"));
+    auto* dpred =
+        context.Output<framework::Tensor>(framework::GradVarName("Logits"));
+    auto place = context.GetEigenDevice<Place>();
+
+    auto x = framework::EigenVector<T>::Flatten(*pred);
+    auto y = framework::EigenVector<T>::Flatten(*label);
+    auto dl = framework::EigenVector<T>::Flatten(*dloss);
+
+    if (dpred) {
+      dpred->mutable_data<T>(context.GetPlace());
+      auto dx = framework::EigenVector<T>::Flatten(*dpred);
+      auto alt_labels = static_cast<T>(2) * y - static_cast<T>(1);
+      dx.device(place) =
+          dl * ((x * alt_labels) < static_cast<T>(1)).template cast<T>() *
+          (-alt_labels);
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

python/paddle/v2/fluid/tests/test_hinge_loss_op.py

+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+class TestHingeLossOp(OpTest):
+    def setUp(self):
+        self.op_type = 'hinge_loss'
+        samples_num = 64
+        logits = np.random.uniform(-10, 10, (samples_num, 1)).astype('float32')
+        labels = np.random.randint(0, 2, (samples_num, 1)).astype('float32')
+
+        self.inputs = {
+            'Logits': logits,
+            'Labels': labels,
+        }
+        loss = np.maximum(1.0 - (2 * labels - 1) * logits, 0)
+        self.outputs = {'Loss': loss}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(['Logits'], 'Loss', max_relative_error=0.008)
+
+
+if __name__ == '__main__':
+    unittest.main()