add lookup_sparse_table_op

paddle/fluid/framework/lod_tensor_test.cc

@@ -255,11 +255,11 @@ TEST(LoDTensor, RecordIO) {
     std::unique_ptr<std::istream> stream_ptr(stream);
     recordio::Scanner scanner(std::move(stream_ptr));
     auto tensors = ReadFromRecordIO(&scanner, ctx);
-    ASSERT_EQ(tensors.size(), 2);
+    ASSERT_EQ(tensors.size(), static_cast<size_t>(2));
     assert_tensor_ok(tensors[0]);
     assert_tensor_ok(tensors[1]);
     tensors = ReadFromRecordIO(&scanner, ctx);
-    ASSERT_EQ(tensors.size(), 2);
+    ASSERT_EQ(tensors.size(), static_cast<size_t>(2));
     assert_tensor_ok(tensors[0]);
     assert_tensor_ok(tensors[1]);
   }

paddle/fluid/framework/selected_rows.cc

@@ -120,11 +120,11 @@ bool SelectedRows::HasKey(int64_t key) const {
                                                                    : true;
 }
 
-std::vector<int64_t> SelectedRows::Get(std::vector<int64_t> keys,
+std::vector<std::pair<int64_t, int64_t>> SelectedRows::Get(
-                                       framework::Tensor* value) const {
+    std::vector<int64_t> keys, framework::Tensor* value) const {
   PADDLE_ENFORCE(value->IsInitialized(),
                  "The value tensor should be initialized.");
-  std::vector<int64_t> non_keys;
+  std::vector<std::pair<int64_t, int64_t>> non_keys_pair;
   int64_t value_width = value_->numel() / value_->dims()[0];
   PADDLE_ENFORCE_EQ(value_width, value->numel() / value->dims()[0],
                     "output tensor should have the same shape with table "
@@ -133,7 +133,7 @@ std::vector<int64_t> SelectedRows::Get(std::vector<int64_t> keys,
   for (size_t i = 0; i < keys.size(); ++i) {
     int64_t index = Index(keys[i]);
     if (index == -1) {
-      non_keys.push_back(keys[i]);
+      non_keys_pair.push_back(std::make_pair(keys[i], static_cast<int64_t>(i)));
     } else {
       framework::VisitDataType(
           framework::ToDataType(value_->type()),
@@ -141,7 +141,7 @@ std::vector<int64_t> SelectedRows::Get(std::vector<int64_t> keys,
                             index * value_width, value_width));
     }
   }
-  return non_keys;
+  return non_keys_pair;
 }
 
 bool SelectedRows::Set(int64_t key, const framework::Tensor& value) {

paddle/fluid/framework/selected_rows.h

@@ -15,6 +15,7 @@ limitations under the License. */
 #pragma once
 
 #include <algorithm>
+#include <utility>
 #include <vector>
 
 #include "paddle/fluid/framework/lod_tensor.h"
@@ -78,10 +79,11 @@ class SelectedRows {
   /*
    * @brief Get value by the key list, if the
    *
-   * @return a list of keys which does not exists in table
+   * @return a list of pair which contains the non-exists key and the index in
+   * the value
    */
-  std::vector<int64_t> Get(std::vector<int64_t> keys,
+  std::vector<std::pair<int64_t, int64_t>> Get(std::vector<int64_t> keys,
-                           framework::Tensor* tensor) const;
+                                               framework::Tensor* value) const;
 
   /*
    * @brief Set a key-value pair into the table.

paddle/fluid/framework/selected_rows_test.cc

@@ -59,7 +59,7 @@ TEST_F(SelectedRowsTester, SerializeAndDeseralize) {
   ASSERT_EQ(selected_rows_->GetCompleteDims(), dst_tensor.GetCompleteDims());
 }
 
-TEST_F(SelectedRowsTester, Table) {
+TEST_F(SelectedRowsTester, SparseTable) {
   platform::CPUPlace cpu;
   SelectedRows table;
   // initialize a sparse table
@@ -87,11 +87,11 @@ TEST_F(SelectedRowsTester, Table) {
   framework::Tensor get_value;
   get_value.mutable_data<float>(framework::make_ddim({2, 100}), cpu);
   std::vector<int64_t> keys({non_key, key});
-  auto non_keys = table.Get(keys, &get_value);
+  auto non_key_pairs = table.Get(keys, &get_value);
 
   ASSERT_EQ(get_value.data<float>()[100], static_cast<float>(10));
-  ASSERT_EQ(non_keys.size(), static_cast<size_t>(1));
+  ASSERT_EQ(non_key_pairs.size(), static_cast<size_t>(1));
-  ASSERT_EQ(non_keys[0], non_key);
+  ASSERT_EQ(non_key_pairs[0].first, non_key);
 }
 
 }  // namespace framework

paddle/fluid/operators/detail/serde_test.cc

@@ -108,7 +108,7 @@ void RunSerdeTestSelectedRows(platform::Place place) {
     EXPECT_FLOAT_EQ(tensor_data2[i], 32.7);
   }
   for (size_t i = 0; i < rows2->size(); ++i) {
-    EXPECT_EQ(rows_data2[i], i);
+    EXPECT_EQ(rows_data2[i], static_cast<int64_t>(i));
   }
   EXPECT_EQ(slr2->height(), 1000);
 }

paddle/fluid/operators/lookup_sparse_table_op.cc

@@ -0,0 +1,154 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+
+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 <algorithm>
+
+#include "paddle/fluid/framework/data_type.h"
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/operators/math/math_function.h"
+#include "paddle/fluid/platform/device_context.h"
+
+namespace paddle {
+namespace operators {
+
+constexpr int64_t kNoPadding = -1;
+
+class LookupSparseTableInferShape : public framework::InferShapeBase {
+ public:
+  void operator()(framework::InferShapeContext *ctx) const override {
+    PADDLE_ENFORCE(ctx->HasOutput("Out"),
+                   "Output(Out) of LookupSparseTableOp should not be null.");
+    auto shape_w = ctx->GetInputDim("W");
+    auto shape_ids = ctx->GetInputDim("Ids");
+    shape_w[0] = shape_ids.size();
+    ctx->SetOutputDim("Out", shape_w);
+  }
+};
+
+class LookupSparseTableOp : public framework::OperatorBase {
+ public:
+  using framework::OperatorBase::OperatorBase;
+
+ private:
+  void RunImpl(const framework::Scope &scope,
+               const platform::Place &dev_place) const override {
+    auto out_var = scope.FindVar(Output("Out"));
+    auto w_var = scope.FindVar(Input("W"));
+    auto ids_var = scope.FindVar(Input("Ids"));
+    unsigned int seed = static_cast<unsigned int>(Attr<int>("seed"));
+    float min = Attr<float>("min");
+    float max = Attr<float>("max");
+
+    PADDLE_ENFORCE(out_var->IsType<framework::LoDTensor>(),
+                   "The type of Out var should be LodTensor.");
+    PADDLE_ENFORCE(w_var->IsType<framework::SelectedRows>(),
+                   "The type of W var should be SelectedRows.");
+    PADDLE_ENFORCE(ids_var->IsType<framework::SelectedRows>(),
+                   "The type of Ids var should be SelectedRows.");
+    auto &ids_t = ids_var->Get<framework::SelectedRows>();
+    auto out_t = out_var->GetMutable<framework::LoDTensor>();
+    auto w_t = w_var->GetMutable<framework::SelectedRows>();
+    auto keys = ids_t.rows();
+
+    // TODO(Yancey1989): support CUDA Place for the sparse table
+    platform::CPUPlace cpu;
+    auto out_shape = w_t->value().dims();
+    out_shape[0] = keys.size();
+    out_t->Resize(out_shape);
+    out_t->mutable_data(cpu, w_t->value().type());
+
+    PADDLE_ENFORCE_EQ(framework::ToDataType(w_t->value().type()),
+                      framework::proto::VarType::FP32,
+                      "The sparse table only support FP32");
+
+    auto non_keys_pair = w_t->Get(keys, out_t);
+    auto value_shape = w_t->value().dims();
+    value_shape[0] = 1;
+    for (const auto &it : non_keys_pair) {
+      const auto key = it.first;
+      const auto index = it.second;
+      framework::Tensor value;
+      value.Resize(value_shape);
+      auto data = value.mutable_data<float>(cpu);
+
+      std::minstd_rand engine;
+      engine.seed(seed);
+      std::uniform_real_distribution<float> dist(min, max);
+      int64_t size = value.numel();
+      for (int64_t i = 0; i < size; ++i) {
+        data[i] = dist(engine);
+      }
+      w_t->Set(key, value);
+      memory::Copy(cpu, out_t->mutable_data<float>(cpu) + index * value.numel(),
+                   cpu, value.data<float>(), value.numel() * sizeof(float));
+    }
+  }
+};
+
+class LookupSparseTableOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  LookupSparseTableOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("W",
+             "(SelectedRows) The input represents embedding table, "
+             "which is a learnable parameter.");
+    AddInput("Ids",
+             "(SelectedRows) Ids's type should be SelectedRows "
+             "the rows of Ids contains the Ids to be looked up in W.");
+    AddOutput("Out",
+              "(SelectedRows) The lookup results, which have the "
+              "same type as W.");
+    AddAttr<int64_t>("padding_idx",
+                     "(int64, default -1) "
+                     "If the value is -1, it makes no effect to lookup. "
+                     "Otherwise the given value indicates padding the output "
+                     "with zeros whenever lookup encounters it in Ids.")
+        .SetDefault(kNoPadding);
+    AddAttr<float>("min",
+                   "(float, default -1.0) "
+                   "Minimum value of uniform random")
+        .SetDefault(-1.0f);
+    AddAttr<float>("max",
+                   "(float, default 1.0) "
+                   "Maximun value of uniform random")
+        .SetDefault(1.0f);
+    AddAttr<int>("seed",
+                 "(int, default 0) "
+                 "Random seed used for generating samples. "
+                 "0 means use a seed generated by the system."
+                 "Note that if seed is not 0, this operator will always "
+                 "generate the same random numbers every time.")
+        .SetDefault(0);
+    AddComment(R"DOC(
+Lookup Sprase Tablel Operator.
+
+This operator is used to perform lookup on parameter W,
+then concatenated into a sparse tensor.
+
+The type of Ids(Input) is SelectedRows, the rows of Ids contains
+the ids to be looked up in W;
+if the Id is not in the sparse table, this operator will return a
+random value and set the value into the table for the next looking up.
+
+)DOC");
+  }
+};
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OPERATOR(lookup_sparse_table, ops::LookupSparseTableOp,
+                  ops::LookupSparseTableInferShape,
+                  ops::LookupSparseTableOpMaker,
+                  paddle::framework::EmptyGradOpMaker);

python/paddle/fluid/tests/unittests/test_lookup_sparse_table_op.py

@@ -0,0 +1,86 @@
+#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# 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.
+
+import unittest
+import numpy as np
+from op_test import OpTest
+import paddle.fluid.core as core
+from paddle.fluid.op import Operator
+
+
+def output_hist(out):
+    hist, _ = np.histogram(out, range=(-5, 10))
+    hist = hist.astype("float32")
+    hist /= float(out.size)
+    prob = 0.1 * np.ones((10))
+    return hist, prob
+
+
+class TestLookupSpraseTable(OpTest):
+    def check_with_place(self, place):
+        scope = core.Scope()
+
+        # create and initialize Id Variable
+        ids = scope.var("Ids").get_selected_rows()
+        ids_array = [0, 2, 3, 5, 100]
+        ids.set_rows(ids_array)
+
+        # create and initialize W Variable
+        rows = [0, 1, 2, 3, 4, 5, 6]
+        row_numel = 10000
+
+        w_selected_rows = scope.var('W').get_selected_rows()
+        w_selected_rows.set_height(len(rows))
+        w_selected_rows.set_rows(rows)
+        w_array = np.ones((len(rows), row_numel)).astype("float32")
+        for i in range(len(rows)):
+            w_array[i] *= i
+        w_tensor = w_selected_rows.get_tensor()
+        w_tensor.set(w_array, place)
+
+        # create Out Variable
+        out_tensor = scope.var('Out').get_tensor()
+
+        # create and run lookup_table operator
+        lookup_table = Operator(
+            "lookup_sparse_table",
+            W='W',
+            Ids='Ids',
+            Out='Out',
+            min=-5.0,
+            max=10.0,
+            seed=10)
+        lookup_table.run(scope, place)
+
+        # get result from Out
+        result_array = np.array(out_tensor)
+        # all(): return True if all elements of the iterable are true (or if the iterable is empty)
+        for idx, row in enumerate(ids_array[:-2]):
+            assert (row == result_array[idx]).all()
+
+        # check the random value
+        hist, prob = output_hist(result_array[-1])
+        self.assertTrue(
+            np.allclose(
+                hist, prob, rtol=0, atol=0.01), "hist: " + str(hist))
+
+    def test_w_is_selected_rows(self):
+        places = [core.CPUPlace()]
+        # currently only support CPU
+        for place in places:
+            self.check_with_place(place)
+
+
+if __name__ == "__main__":
+    unittest.main()