Add nce op

paddle/operators/nce_op.cc

@@ -0,0 +1,159 @@
+/* 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/nce_op.h"
+
+namespace paddle {
+namespace operators {
+
+using framework::Tensor;
+
+class NCEOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("Input"));
+    PADDLE_ENFORCE(ctx->HasInput("Label"));
+    PADDLE_ENFORCE(ctx->HasInput("Weight"));
+    PADDLE_ENFORCE(ctx->HasOutput("Cost"));
+    PADDLE_ENFORCE(ctx->HasOutput("SampleLogits"));
+    PADDLE_ENFORCE(ctx->HasOutput("SampleLabels"));
+
+    auto x_dims = ctx->GetInputDim("Input");
+    auto label_dims = ctx->GetInputDim("Label");
+    PADDLE_ENFORCE_EQ(x_dims[0], label_dims[0]);
+    int num_true_classes = label_dims.size() == 2 ? label_dims[1] : 1;
+    if (ctx->HasInput("Bias")) {
+      PADDLE_ENFORCE_EQ(ctx->GetInputDim("Weight")[0],
+                        ctx->GetInputDim("Bias")[0]);
+    }
+    auto num_sampled_classes = ctx->Attrs().Get<int>("num_sampled_classes");
+    auto num_classes = ctx->Attrs().Get<int>("num_classes");
+    std::vector<int> sampled_labels =
+        ctx->Attrs().Get<std::vector<int>>("sampled_labels");
+    PADDLE_ENFORCE_EQ(num_classes, ctx->GetInputDim("Weight")[0]);
+    PADDLE_ENFORCE_LT(num_sampled_classes, num_classes);
+    if (sampled_labels.size() > 0) {
+      PADDLE_ENFORCE_EQ(sampled_labels.size(),
+                        static_cast<size_t>(num_sampled_classes));
+    }
+    // set dims of output(Out)
+    std::vector<int64_t> out_dims;
+    out_dims.push_back(x_dims[0]);
+    ctx->SetOutputDim("Cost", framework::make_ddim(out_dims));
+
+    // set dims of output(SampleOut)
+    std::vector<int64_t> sample_out_dims;
+    sample_out_dims.push_back(x_dims[0]);
+    sample_out_dims.push_back(num_sampled_classes + num_true_classes);
+    ctx->SetOutputDim("SampleLogits", framework::make_ddim(sample_out_dims));
+    ctx->SetOutputDim("SampleLabels", framework::make_ddim(sample_out_dims));
+  }
+
+ protected:
+  framework::OpKernelType GetKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    return framework::OpKernelType(
+        framework::ToDataType(ctx.Input<Tensor>("Input")->type()),
+        ctx.device_context());
+  }
+};
+
+class NCEOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  NCEOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("Input", "(Tensor) A tensor of shape [batch_size, dim].");
+    AddInput("Label",
+             "(Tensor) A tensor of shape [batch_size, num_true_class]. "
+             "'num_true_class' is the number of target class in each sample.");
+    AddInput("Weight",
+             "(Tensor) A tensor of shape [num_class, dim]. 'num_class' is the "
+             "total number of class.");
+    AddInput("Bias",
+             "(Tensor) A tensor of shape [num_class]. 'num_class' is the total "
+             "number of class. It is a dispensable input.")
+        .AsDispensable();
+    AddInput("SampleWeight",
+             "(Tensor) A tensor of shape [batch_size] storing a weight for "
+             "each sample. And it is a dispensable input. The default value of "
+             "sample is 1.")
+        .AsDispensable();
+    AddOutput("Cost",
+              "(Tensor) A tensor of shape [batch_size]. Cost of samples.");
+    AddOutput("SampleLogits", "An intermediate tensor.").AsIntermediate();
+    AddOutput("SampleLabels", "An intermediate tensor.").AsIntermediate();
+    AddAttr<int>("num_classes", "Total number of classes.");
+    AddAttr<int>("num_sampled_classes", "The number of negative classes.")
+        .SetDefault(10);
+    AddAttr<std::vector<int>>("sampled_labels", "");
+    AddComment(R"DOC(
+Computes and returns the noise-contrastive estimation training loss.
+See [Noise-contrastive estimation: A new estimation principle for unnormalized statistical models](http://www.jmlr.org/proceedings/papers/v9/gutmann10a/gutmann10a.pdf).
+By default this uses a uniform distribution for sampling.
+The number of target classes per example should be same. If you have a variable number of target classes, you can pad them out to a constant number by either repeating them or by padding with an otherwise unused class.
+)DOC");
+  }
+};
+
+class NCEOpGrad : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("Input"));
+    PADDLE_ENFORCE(ctx->HasInput("Weight"));
+    PADDLE_ENFORCE(ctx->HasInput("Cost"));
+    PADDLE_ENFORCE(ctx->HasInput("SampleLogits"));
+    PADDLE_ENFORCE(ctx->HasInput("SampleLabels"));
+    PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Cost")),
+                   "The input(Out@GRAD) should not be null");
+
+    auto x_dims = ctx->GetInputDim("Input");
+    auto x_grad_name = framework::GradVarName("Input");
+    if (ctx->HasOutput(x_grad_name)) {
+      ctx->SetOutputDim(x_grad_name, x_dims);
+    }
+
+    auto w_dims = ctx->GetInputDim("Weight");
+    auto w_grad_name = framework::GradVarName("Weight");
+    if (ctx->HasOutput(w_grad_name)) {
+      ctx->SetOutputDim(w_grad_name, w_dims);
+    }
+
+    auto bias_grad_name = framework::GradVarName("Bias");
+    if (ctx->HasOutput(bias_grad_name)) {
+      auto bias_dims = ctx->GetInputDim("Bias");
+      ctx->SetOutputDim(bias_grad_name, bias_dims);
+    }
+  }
+
+ protected:
+  framework::OpKernelType GetKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    return framework::OpKernelType(
+        framework::ToDataType(ctx.Input<Tensor>("Input")->type()),
+        ctx.device_context());
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP(nce, ops::NCEOp, ops::NCEOpMaker, nce_grad, ops::NCEOpGrad);
+REGISTER_OP_CPU_KERNEL(nce, ops::NCEKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(nce_grad,
+                       ops::NCEGradKernel<paddle::platform::CPUPlace, float>);

paddle/operators/nce_op.h

@@ -0,0 +1,209 @@
+/* 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 <math.h>
+#include <random>
+#include "paddle/framework/eigen.h"
+#include "paddle/framework/op_registry.h"
+#include "unsupported/Eigen/CXX11/Tensor"
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+
+template <typename T, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
+
+template <typename Place, typename T>
+void PrepareSamples(const framework::ExecutionContext& context) {
+  auto label = context.Input<Tensor>("Label");
+  const int64_t* label_data = label->data<int64_t>();
+  auto label_dims = label->dims();
+  int num_classes = context.Attr<int>("num_classes");
+  // for unitest
+  std::vector<int> sampled_labels =
+      context.Attr<std::vector<int>>("sampled_labels");
+  // random machine
+  std::random_device rd;
+  std::mt19937 rng(rd());
+  std::uniform_int_distribution<int> rand(0, num_classes - 1);
+
+  auto sample_labels = context.Output<Tensor>("SampleLabels");
+  auto sample_labels_dims = sample_labels->dims();
+  int64_t* sample_labels_data =
+      sample_labels->mutable_data<int64_t>(context.GetPlace());
+
+  int num_label = label_dims.size() == 2 ? label_dims[1] : 1;
+  int index = 0;
+  for (size_t i = 0; i < label_dims[0]; ++i) {
+    int j = 0;
+    for (; j < num_label; ++j) {
+      sample_labels_data[index++] = label_data[i * num_label + j];
+    }
+    if (sampled_labels.size() > 0) {
+      for (auto label : sampled_labels) {
+        sample_labels_data[index++] = label;
+      }
+    } else {
+      for (; j < sample_labels_dims[1]; ++j) {
+        sample_labels_data[index++] = rand(rng);
+      }
+    }
+  }
+}
+
+template <typename Place, typename T>
+class NCEKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    PrepareSamples<Place, T>(context);
+    auto sample_labels = context.Output<Tensor>("SampleLabels");
+    const int64_t* sample_labels_data = sample_labels->data<int64_t>();
+    auto sample_out = context.Output<Tensor>("SampleLogits");
+    T* sample_out_data = sample_out->mutable_data<T>(context.GetPlace());
+    auto label = context.Input<Tensor>("Label");
+    auto sample_weight = context.Input<Tensor>("SampleWeight");
+    const T* sample_weight_data = nullptr;
+    if (sample_weight != nullptr) {
+      sample_weight_data = sample_weight->data<T>();
+    }
+    auto out = context.Output<Tensor>("Cost");
+    T* out_data = out->mutable_data<T>(context.GetPlace());
+    int num_smalped_classes = context.Attr<int>("num_sampled_classes");
+    int num_classes = context.Attr<int>("num_classes");
+    int num_true_class = 1;
+    if (label != nullptr) {
+      num_true_class = label->dims()[1];
+    }
+    T b = 1. / num_classes * num_smalped_classes;
+    // forward bias
+    auto bias = context.Input<Tensor>("Bias");
+    if (bias != nullptr) {
+      const T* bias_data = bias->data<T>();
+      for (size_t i = 0; i < sample_labels->numel(); ++i) {
+        sample_out_data[i] = bias_data[sample_labels_data[i]];
+      }
+    } else {
+      for (size_t i = 0; i < sample_labels->numel(); ++i) {
+        sample_out_data[i] = 0;
+      }
+    }
+    // forward mul
+    auto input_mat = EigenMatrix<T>::From(*(context.Input<Tensor>("Input")));
+    auto weight_mat = EigenMatrix<T>::From(*(context.Input<Tensor>("Weight")));
+    for (size_t i = 0; i < sample_labels->numel(); ++i) {
+      Eigen::Tensor<float, 0, Eigen::RowMajor, Eigen::DenseIndex> result =
+          (input_mat.chip((int)(i / sample_labels->dims()[1]), 0) *
+           weight_mat.chip(sample_labels_data[i], 0))
+              .sum();
+      sample_out_data[i] += result(0);
+      // activation_->forward
+      sample_out_data[i] = (1. / (1. + exp(-sample_out_data[i])));
+    }
+    // forward cost
+    for (size_t i = 0; i < sample_labels->dims()[0]; ++i) {
+      size_t j = 0;
+      out_data[i] = 0;
+      T w = sample_weight == nullptr ? 1. : sample_weight_data[i];
+      // for true classes
+      for (; j < num_true_class; ++j) {
+        T o = sample_out_data[i * sample_out->dims()[1] + j];
+        T cost = -log(o / (o + b));
+        out_data[i] += w * cost;
+      }
+      // for sampled neg classes
+      for (; j < sample_labels->dims()[1]; ++j) {
+        T o = sample_out_data[i * sample_out->dims()[1] + j];
+        T cost = -log(b / (o + b));
+        out_data[i] += w * cost;
+      }
+    }
+  }
+};
+
+template <typename Place, typename T>
+class NCEGradKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto d_out = context.Input<Tensor>(framework::GradVarName("Cost"));
+    const T* d_out_data = d_out->data<T>();
+    auto label = context.Input<Tensor>("Label");
+    auto sample_out = context.Input<Tensor>("SampleLogits");
+    const T* sample_out_data = sample_out->data<T>();
+    auto sample_labels = context.Input<Tensor>("SampleLabels");
+    const int64_t* sample_labels_data = sample_labels->data<int64_t>();
+    auto sample_weight = context.Input<Tensor>("SampleWeight");
+    const T* sample_weight_data = nullptr;
+    if (sample_weight != nullptr) {
+      sample_weight_data = sample_weight->data<T>();
+    }
+    int num_smalped_classes = context.Attr<int>("num_sampled_classes");
+    int num_classes = context.Attr<int>("num_classes");
+    int num_true_class = 1;
+    if (label != nullptr) {
+      num_true_class = label->dims()[1];
+    }
+    T b = 1. / num_classes * num_smalped_classes;
+    Tensor sample_grad;  // tmp tensor
+    T* sample_grad_data =
+        sample_grad.mutable_data<T>(sample_labels->dims(), context.GetPlace());
+    // backward cost
+    for (size_t i = 0; i < sample_labels->numel(); ++i) {
+      T o = sample_out_data[i];
+      T w = sample_weight == nullptr
+                ? 1
+                : sample_weight_data[i / sample_labels->dims()[1]];
+      sample_grad_data[i] = (i % sample_labels->dims()[1]) < num_true_class
+                                ? w * (b / (o + b)) * (o - 1)
+                                : w * (o * (1 - o) / (o + b));
+      sample_grad_data[i] *= d_out_data[i / sample_labels->dims()[1]];
+    }
+    // get d_bias
+    auto d_bias = context.Output<Tensor>(framework::GradVarName("Bias"));
+    if (d_bias != nullptr) {
+      T* d_bias_data = d_bias->mutable_data<T>(context.GetPlace());
+      for (size_t i = 0; i < sample_labels->numel(); ++i) {
+        d_bias_data[sample_labels_data[i]] += sample_grad_data[i];
+      }
+    }
+    // get d_w
+    auto d_w = context.Output<Tensor>(framework::GradVarName("Weight"));
+    if (d_w != nullptr) {
+      d_w->mutable_data<T>(context.GetPlace());
+      auto d_w_matrix = EigenMatrix<T>::From(*d_w);
+      auto x_matrix = EigenMatrix<T>::From(*(context.Input<Tensor>("Input")));
+      for (size_t i = 0; i < sample_labels->numel(); ++i) {
+        d_w_matrix.chip(sample_labels_data[i], 0) +=
+            x_matrix.chip((int)(i / sample_labels->dims()[1]), 0) *
+            sample_grad_data[i];
+      }
+    }
+    // get d_x
+    auto d_x = context.Output<Tensor>(framework::GradVarName("Input"));
+    if (d_x != nullptr) {
+      d_x->mutable_data<T>(context.GetPlace());
+      auto d_x_matrix = EigenMatrix<T>::From(*d_x);
+      auto w_matrix = EigenMatrix<T>::From(*(context.Input<Tensor>("Weight")));
+      for (size_t i = 0; i < sample_labels->numel(); ++i) {
+        d_x_matrix.chip((int)(i / sample_labels->dims()[1]), 0) +=
+            w_matrix.chip(sample_labels_data[i], 0) * sample_grad_data[i];
+      }
+    }
+  }
+};
+}  // namespace operators
+}  // namespace paddle