Merge pull request #14057 from velconia/continue_hash_op

[1.1] Add hash_op implementation

CMakeLists.txt

@@ -179,6 +179,7 @@ include(external/eigen)     # download eigen3
 include(external/pybind11)  # download pybind11
 include(external/cares)
 include(external/cub)
+include(external/xxhash)    # download xxhash
 
 if (NOT WIN32)
 # there is no official support of snappystream, warpctc, nccl, cupti in windows

cmake/external/xxhash.cmake

+INCLUDE(ExternalProject)
+
+set(XXHASH_SOURCE_DIR ${THIRD_PARTY_PATH}/xxhash)
+set(XXHASH_INSTALL_DIR ${THIRD_PARTY_PATH}/install/xxhash)
+set(XXHASH_INCLUDE_DIR "${XXHASH_INSTALL_DIR}/include")
+
+IF(WITH_STATIC_LIB)
+  SET(BUILD_CMD make lib)
+ELSE()
+  SET(BUILD_CMD sed -i "s/-Wstrict-prototypes -Wundef/-Wstrict-prototypes -Wundef -fPIC/g" ${XXHASH_SOURCE_DIR}/src/extern_xxhash/Makefile && make lib)
+ENDIF()
+
+ExternalProject_Add(
+    extern_xxhash
+    ${EXTERNAL_PROJECT_LOG_ARGS}
+    GIT_REPOSITORY  "https://github.com/Cyan4973/xxHash"
+    GIT_TAG         "v0.6.5"
+    PREFIX          ${XXHASH_SOURCE_DIR}
+    DOWNLOAD_NAME   "xxhash"
+    UPDATE_COMMAND  ""
+    CONFIGURE_COMMAND ""
+    BUILD_IN_SOURCE 1
+    PATCH_COMMAND
+    BUILD_COMMAND     ${BUILD_CMD}
+    INSTALL_COMMAND   export PREFIX=${XXHASH_INSTALL_DIR}/ && make install
+    TEST_COMMAND      ""
+)
+
+set(XXHASH_LIBRARIES "${XXHASH_INSTALL_DIR}/lib/libxxhash.a")
+INCLUDE_DIRECTORIES(${XXHASH_INCLUDE_DIR})
+
+add_library(xxhash STATIC IMPORTED GLOBAL)
+set_property(TARGET xxhash PROPERTY IMPORTED_LOCATION ${XXHASH_LIBRARIES})
+include_directories(${XXHASH_INCLUDE_DIR})
+add_dependencies(xxhash extern_xxhash)
+
+LIST(APPEND external_project_dependencies xxhash)
+
+IF(WITH_C_API)
+  INSTALL(DIRECTORY ${XXHASH_INCLUDE_DIR} DESTINATION third_party/xxhash)
+  IF(ANDROID)
+    INSTALL(FILES ${XXHASH_LIBRARIES} DESTINATION third_party/xxhash/lib/${ANDROID_ABI})
+  ELSE()
+    INSTALL(FILES ${XXHASH_LIBRARIES} DESTINATION third_party/xxhash/lib)
+  ENDIF()
+ENDIF()

cmake/inference_lib.cmake

@@ -31,7 +31,7 @@ function(copy TARGET)
     foreach(index RANGE ${len})
         list(GET copy_lib_SRCS ${index} src)
         list(GET copy_lib_DSTS ${index} dst)
-        add_custom_command(TARGET ${TARGET} PRE_BUILD 
+        add_custom_command(TARGET ${TARGET} PRE_BUILD
           COMMAND mkdir -p "${dst}"
           COMMAND cp -r "${src}" "${dst}"
           COMMENT "copying ${src} -> ${dst}")
@@ -67,6 +67,13 @@ copy(boost_lib
   DEPS boost
 )
 
+set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/xxhash")
+copy(xxhash_lib
+  SRCS ${XXHASH_INCLUDE_DIR} ${XXHASH_LIBRARIES}
+  DSTS ${dst_dir} ${dst_dir}/lib
+  DEPS xxhash
+)
+
 if(NOT PROTOBUF_FOUND)
     set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/protobuf")
     copy(protobuf_lib
@@ -186,7 +193,7 @@ copy(cmake_cache
   DSTS ${FLUID_INSTALL_DIR})
 
 # This command generates a complete fluid library for both train and inference
-add_custom_target(fluid_lib_dist DEPENDS ${fluid_lib_dist_dep}) 
+add_custom_target(fluid_lib_dist DEPENDS ${fluid_lib_dist_dep})
 
 # Following commands generate a inference-only fluid library
 # third_party, version.txt and CMakeCache.txt are the same position with ${FLUID_INSTALL_DIR}

paddle/fluid/API.spec

@@ -176,6 +176,7 @@ paddle.fluid.layers.sigmoid_cross_entropy_with_logits ArgSpec(args=['x', 'label'
 paddle.fluid.layers.maxout ArgSpec(args=['x', 'groups', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.sequence_reverse ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.affine_channel ArgSpec(args=['x', 'scale', 'bias', 'data_layout', 'name'], varargs=None, keywords=None, defaults=(None, None, 'NCHW', None))
+paddle.fluid.layers.hash ArgSpec(args=['input', 'hash_size', 'num_hash', 'name'], varargs=None, keywords=None, defaults=(1, None))
 paddle.fluid.layers.data ArgSpec(args=['name', 'shape', 'append_batch_size', 'dtype', 'lod_level', 'type', 'stop_gradient'], varargs=None, keywords=None, defaults=(True, 'float32', 0, VarType.LOD_TENSOR, True))
 paddle.fluid.layers.open_files ArgSpec(args=['filenames', 'shapes', 'lod_levels', 'dtypes', 'thread_num', 'buffer_size', 'pass_num', 'is_test'], varargs=None, keywords=None, defaults=(None, None, 1, None))
 paddle.fluid.layers.read_file ArgSpec(args=['reader'], varargs=None, keywords=None, defaults=None)

paddle/fluid/inference/api/demo_ci/CMakeLists.txt

@@ -52,6 +52,7 @@ include_directories("${PADDLE_LIB}")
 include_directories("${PADDLE_LIB}/third_party/install/protobuf/include")
 include_directories("${PADDLE_LIB}/third_party/install/glog/include")
 include_directories("${PADDLE_LIB}/third_party/install/gflags/include")
+include_directories("${PADDLE_LIB}/third_party/install/xxhash/include")
 if (NOT WIN32)
 include_directories("${PADDLE_LIB}/third_party/install/snappy/include")
 include_directories("${PADDLE_LIB}/third_party/install/snappystream/include")
@@ -61,8 +62,8 @@ endif(NOT WIN32)
 include_directories("${PADDLE_LIB}/third_party/boost")
 include_directories("${PADDLE_LIB}/third_party/eigen3")
 
-if (NOT WIN32) 
-  if (USE_TENSORRT AND WITH_GPU) 
+if (NOT WIN32)
+  if (USE_TENSORRT AND WITH_GPU)
       include_directories("${TENSORRT_INCLUDE_DIR}")
       link_directories("${TENSORRT_LIB_DIR}")
   endif()
@@ -77,13 +78,14 @@ endif(NOT WIN32)
 link_directories("${PADDLE_LIB}/third_party/install/protobuf/lib")
 link_directories("${PADDLE_LIB}/third_party/install/glog/lib")
 link_directories("${PADDLE_LIB}/third_party/install/gflags/lib")
+link_directories("${PADDLE_LIB}/third_party/install/xxhash/lib")
 link_directories("${PADDLE_LIB}/paddle/lib")
 
 add_executable(${DEMO_NAME} ${DEMO_NAME}.cc)
 
 if(WITH_MKL)
   include_directories("${PADDLE_LIB}/third_party/install/mklml/include")
-  set(MATH_LIB ${PADDLE_LIB}/third_party/install/mklml/lib/libmklml_intel${CMAKE_SHARED_LIBRARY_SUFFIX} 
+  set(MATH_LIB ${PADDLE_LIB}/third_party/install/mklml/lib/libmklml_intel${CMAKE_SHARED_LIBRARY_SUFFIX}
                ${PADDLE_LIB}/third_party/install/mklml/lib/libiomp5${CMAKE_SHARED_LIBRARY_SUFFIX})
   set(MKLDNN_PATH "${PADDLE_LIB}/third_party/install/mkldnn")
   if(EXISTS ${MKLDNN_PATH})
@@ -107,7 +109,7 @@ if (NOT WIN32)
 set(EXTERNAL_LIB "-lrt -ldl -lpthread")
 set(DEPS ${DEPS}
     ${MATH_LIB} ${MKLDNN_LIB}
-    glog gflags protobuf snappystream snappy z
+    glog gflags protobuf snappystream snappy z xxhash
     ${EXTERNAL_LIB})
 else()
 set(DEPS ${DEPS}
@@ -120,7 +122,7 @@ endif(NOT WIN32)
 
 if(WITH_GPU)
   if(NOT WIN32)
-    if (USE_TENSORRT) 
+    if (USE_TENSORRT)
       set(DEPS ${DEPS} ${TENSORRT_LIB_DIR}/libnvinfer${CMAKE_STATIC_LIBRARY_SUFFIX})
       set(DEPS ${DEPS} ${TENSORRT_LIB_DIR}/libnvinfer_plugin${CMAKE_STATIC_LIBRARY_SUFFIX})
     endif()

paddle/fluid/operators/CMakeLists.txt

@@ -268,6 +268,7 @@ if (WITH_GPU AND TENSORRT_FOUND)
 else()
     set(DEPS_OPS ${DEPS_OPS} tensorrt_engine_op)
 endif()
+op_library(hash_op DEPS xxhash)
 op_library(clip_by_norm_op DEPS selected_rows_functor selected_rows)
 op_library(sum_op DEPS selected_rows_functor)
 op_library(sgd_op DEPS selected_rows_functor)

paddle/fluid/operators/hash_op.cc

+/* 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 "paddle/fluid/operators/hash_op.h"
+#include <string>
+#include <vector>
+
+namespace paddle {
+namespace operators {
+
+class HashOp : public framework::OperatorWithKernel {
+ public:
+  HashOp(const std::string &type, const framework::VariableNameMap &inputs,
+         const framework::VariableNameMap &outputs,
+         const framework::AttributeMap &attrs)
+      : OperatorWithKernel(type, inputs, outputs, attrs) {}
+
+  void InferShape(framework::InferShapeContext *ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"),
+                   "Input(X) of HashOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Out"),
+                   "Output(Out) of HashOp should not be null.");
+
+    auto dims = ctx->GetInputDim("X");
+    PADDLE_ENFORCE_EQ(dims.size(), 2UL,
+                      "The input of hash_op's dimensions must be 2");
+    std::vector<int64_t> out_dims;
+    out_dims.reserve(dims.size() + 1);
+    // copy all dims except the last one
+    for (size_t i = 0u; i != dims.size() - 1; ++i) {
+      out_dims.emplace_back(dims[i]);
+    }
+    int num_hash = ctx->Attrs().Get<int>("num_hash");
+    out_dims.emplace_back(num_hash);
+    // keep the last dim to 1
+    out_dims.emplace_back(1);
+
+    ctx->SetOutputDim("Out", framework::make_ddim(out_dims));
+    ctx->ShareLoD("X", /*->*/ "Out");
+  }
+};
+
+class HashOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() override {
+    AddInput("X", "(Tensor) Input tensor of scale operator.");
+    AddOutput("Out", "(Tensor) Output tensor of scale operator.");
+    AddComment(R"DOC(
+**Hash Operator**
+$$Out = scale * X$$
+)DOC");
+    AddAttr<int>("num_hash", "").SetDefault(1);
+    AddAttr<int>("mod_by", "").SetDefault(100000);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+
+REGISTER_OP_WITHOUT_GRADIENT(hash, ops::HashOp, ops::HashOpMaker);
+REGISTER_OP_CPU_KERNEL(hash, ops::HashKerel<int>, ops::HashKerel<int64_t>);

paddle/fluid/operators/hash_op.h

+/* 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. */
+
+#pragma once
+
+extern "C" {
+#include <xxhash.h>
+}
+#include "paddle/fluid/framework/eigen.h"
+#include "paddle/fluid/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+// template <typename DeviceContext, typename T>
+template <typename T>
+class HashKerel : public framework::OpKernel<T> {
+ public:
+  virtual void Compute(const framework::ExecutionContext& context) const {
+    auto* out_t = context.Output<framework::LoDTensor>("Out");
+    auto* in_t = context.Input<framework::LoDTensor>("X");
+    int mod_by = context.Attr<int>("mod_by");
+    int num_hash = context.Attr<int>("num_hash");
+    auto* output = out_t->mutable_data<T>(context.GetPlace());
+
+    auto in_dims = in_t->dims();
+    auto in_lod = in_t->lod();
+    PADDLE_ENFORCE_EQ(
+        static_cast<uint64_t>(in_dims[0]), in_lod[0].back(),
+        "The actual input data's size mismatched with LoD information.");
+
+    auto seq_length = in_dims[0];
+    auto last_dim = in_dims[in_dims.size() - 1];
+    auto* input = in_t->data<T>();
+    for (int idx = 0; idx < seq_length; ++idx) {
+      for (int ihash = 0; ihash != num_hash; ++ihash) {
+        output[idx * num_hash + ihash] =
+            XXH64(input, sizeof(int) * last_dim, ihash) % mod_by;
+      }
+      input += last_dim;
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/train/demo/CMakeLists.txt

@@ -15,6 +15,7 @@ include_directories("${PADDLE_LIB}")
 include_directories("${PADDLE_LIB}/third_party/install/protobuf/include")
 include_directories("${PADDLE_LIB}/third_party/install/glog/include")
 include_directories("${PADDLE_LIB}/third_party/install/gflags/include")
+include_directories("${PADDLE_LIB}/third_party/install/xxhash/include")
 include_directories("${PADDLE_LIB}/third_party/install/snappy/include")
 include_directories("${PADDLE_LIB}/third_party/install/snappystream/include")
 include_directories("${PADDLE_LIB}/third_party/install/zlib/include")
@@ -27,6 +28,7 @@ link_directories("${PADDLE_LIB}/third_party/install/snappystream/lib")
 link_directories("${PADDLE_LIB}/third_party/install/protobuf/lib")
 link_directories("${PADDLE_LIB}/third_party/install/glog/lib")
 link_directories("${PADDLE_LIB}/third_party/install/gflags/lib")
+link_directories("${PADDLE_LIB}/third_party/install/xxhash/lib")
 link_directories("${PADDLE_LIB}/third_party/install/zlib/lib")
 
 add_executable(demo_trainer demo_trainer.cc)
@@ -62,5 +64,5 @@ target_link_libraries(demo_trainer
         ${ARCHIVE_END}
         ${MATH_LIB}
         ${MKLDNN_LIB}
-        glog gflags protobuf snappystream snappy z
+        glog gflags protobuf snappystream snappy z xxhash
         ${EXTERNAL_LIB})

paddle/scripts/paddle_build.sh

@@ -95,9 +95,9 @@ function cmake_gen() {
                 exit 1
             fi
         fi
-    else 
+    else
         if [ "$1" != "" ]; then
-            echo "using python abi: $1"     
+            echo "using python abi: $1"
             if [ "$1" == "cp27-cp27m" ]; then
                 export LD_LIBRARY_PATH=/opt/_internal/cpython-2.7.11-ucs2/lib:${LD_LIBRARY_PATH#/opt/_internal/cpython-2.7.11-ucs4/lib:}
                 export PATH=/opt/python/cp27-cp27m/bin/:${PATH}
@@ -119,7 +119,7 @@ function cmake_gen() {
            fi
         fi
     fi
-    
+
     if [ "$SYSTEM" == "Darwin" ]; then
         WITH_DISTRIBUTE=${WITH_DISTRIBUTE:-ON}
         WITH_AVX=${WITH_AVX:-ON}
@@ -127,7 +127,7 @@ function cmake_gen() {
     else
         INFERENCE_DEMO_INSTALL_DIR=${INFERENCE_DEMO_INSTALL_DIR:-/root/.cache/inference_demo}
     fi
-    
+
     cat <<EOF
     ========================================
     Configuring cmake in /paddle/build ...
@@ -394,8 +394,8 @@ EOF
         export http_proxy=
         export https_proxy=
         # TODO: jiabin need to refine this part when these tests fixed on mac
-        ctest --output-on-failure -j $1     
-        # make install should also be test when unittest 
+        ctest --output-on-failure -j $1
+        # make install should also be test when unittest
         make install -j 8
         pip install --user ${INSTALL_PREFIX:-/paddle/build}/opt/paddle/share/wheels/*.whl
         if [[ ${WITH_FLUID_ONLY:-OFF} == "OFF" ]] ; then

python/paddle/fluid/layers/nn.py

@@ -156,6 +156,7 @@ __all__ = [
     'maxout',
     'sequence_reverse',
     'affine_channel',
+    'hash',
 ]
 
 
@@ -1991,17 +1992,17 @@ def sequence_slice(input, offset, length, name=None):
     """
     **Sequence Slice Layer**
 
-    The layer crops a subsequence from given sequence with given start 
+    The layer crops a subsequence from given sequence with given start
     offset and subsequence length.
 
     It only supports sequence data (LoDTensor with lod_level equal to 1).
 
     .. code-block:: text
-    
+
 	- Case:
 
             Given the input Variable **input**:
-                
+
                 input.data = [[a1, a2], [b1, b2], [c1, c2], [d1, d2], [e1, e2]],
                 input.lod = [[3, 2]],
                 input.dims = (5, 2),
@@ -2009,16 +2010,16 @@ def sequence_slice(input, offset, length, name=None):
             with offset.data = [[0], [1]] and length.data = [[2], [1]],
 
             the output Variable will be
-                
+
                 out.data = [[a1, a2], [b1, b2], [e1, e2]],
                 out.lod = [[2, 1]],
                 out.dims = (3, 2).
-	
-    NOTE: The first dimension size of **input**, **offset** and **length** 
+
+    NOTE: The first dimension size of **input**, **offset** and **length**
           should be equal. The **offset** should start from 0.
-    
+
     Args:
-        input(Variable): The input Variable which consists of the complete 
+        input(Variable): The input Variable which consists of the complete
                          sequences.
         offset(Variable): The offset to slice each sequence.
         length(Variable): The length of each subsequence.
@@ -2037,7 +2038,7 @@ def sequence_slice(input, offset, length, name=None):
                               dtype='float32', lod_level=1)
              offset = fluid.layers.assign(input=np.array([[0, 1]]).astype("int32"))
              length = fluid.layers.assign(input=np.array([[2, 1]]).astype("int32"))
-             subseqs = fluid.layers.sequence_slice(input=seqs, offset=offset, 
+             subseqs = fluid.layers.sequence_slice(input=seqs, offset=offset,
                                                    length=length)
     """
     helper = LayerHelper("sequence_slice", **locals())
@@ -2420,12 +2421,12 @@ def layer_norm(input,
         param_attr(ParamAttr|None): The parameter attribute for the learnable
             gain :math:`g`. If :attr:`scale` is False, :attr:`param_attr` is
             omitted. If :attr:`scale` is True and :attr:`param_attr` is None,
-            a default :code:`ParamAttr` would be added as scale. The 
-            :attr:`param_attr` is initialized as 1 if it is added. Default None. 
+            a default :code:`ParamAttr` would be added as scale. The
+            :attr:`param_attr` is initialized as 1 if it is added. Default None.
         bias_attr(ParamAttr|None): The parameter attribute for the learnable
             bias :math:`b`. If :attr:`shift` is False, :attr:`bias_attr` is
             omitted. If :attr:`shift` is True and :attr:`param_attr` is None,
-            a default :code:`ParamAttr` would be added as bias. The 
+            a default :code:`ParamAttr` would be added as bias. The
             :attr:`bias_attr` is initialized as 0 if it is added. Default None.
         act(str): Activation to be applied to the output of layer normalizaiton.
                   Default None.
@@ -3043,8 +3044,8 @@ def sequence_unpad(x, length, name=None):
     """
     **Sequence Unpad Layer**
 
-    This layer removes the padding data in the input sequences and convert 
-    them into sequences with actual length as output, identitied by lod 
+    This layer removes the padding data in the input sequences and convert
+    them into sequences with actual length as output, identitied by lod
     information.
 
     .. code-block:: text
@@ -3054,9 +3055,9 @@ def sequence_unpad(x, length, name=None):
 	Given input Variable **x**:
 	    x.data = [[ 1.0,  2.0,  3.0,  4.0,  5.0],
 		      [ 6.0,  7.0,  8.0,  9.0, 10.0],
-		      [11.0, 12.0, 13.0, 14.0, 15.0]], 
-     
-	in which there are 3 sequences padded to length 5, and the acutal length 
+		      [11.0, 12.0, 13.0, 14.0, 15.0]],
+
+	in which there are 3 sequences padded to length 5, and the acutal length
 	specified by input Variable **length**:
 
 	    length.data = [[2], [3], [4]],
@@ -3064,7 +3065,7 @@ def sequence_unpad(x, length, name=None):
 	after unpadding, the output Variable will be:
 
 	    out.data = [[1.0, 2.0, 6.0, 7.0, 8.0, 11.0, 12.0, 13.0, 14.0]]
-	    out.lod = [[2, 3, 4]]      
+	    out.lod = [[2, 3, 4]]
 
     Args:
         x(Variable): Input Variable which contains the padded sequences with
@@ -5499,9 +5500,9 @@ def roi_align(input,
     Examples:
         .. code-block:: python
 
-            align_out = fluid.layers.roi_align(input=x, 
-                                               rois=rois, 
-                                               pooled_height=7, 
+            align_out = fluid.layers.roi_align(input=x,
+                                               rois=rois,
+                                               pooled_height=7,
                                                pooled_width=7,
                                                spatial_scale=0.5,
                                                sampling_ratio=-1)
@@ -7518,7 +7519,7 @@ def affine_channel(x, scale=None, bias=None, data_layout='NCHW', name=None):
     Useful for replacing spatial batch norm with its equivalent fixed
     transformation. The input also can be 2D tensor and applies a affine
     transformation in second dimension.
-    
+
     Args:
         x (Variable): Feature map input can be a 4D tensor with order NCHW
             or NHWC. It also can be a 2D tensor and the affine transformation
@@ -7551,3 +7552,31 @@ def affine_channel(x, scale=None, bias=None, data_layout='NCHW', name=None):
         attrs={"data_layout": data_layout},
         outputs={"Out": out})
     return out
+
+
+def hash(input, hash_size, num_hash=1, name=None):
+    """
+    hash the input
+     Args:
+        input (Variable): The input variable which is a one-hot word.
+        hash_size (int): The space size for hash algorithm.
+        num_hash (int): The times of hash, default 1.
+        name (str, default None): The name of this layer.
+     Returns:
+        Variable: The hash result variable which is a LoDTensor.
+     Examples:
+        .. code-block:: python
+            word_dict = paddle.dataset.imdb.word_dict()
+            x = fluid.layers.data(shape[1], dtype='int32', lod_level=1)
+            out = fluid.layers.hash(input=x, len(word_dict))
+    """
+    helper = LayerHelper('hash', **locals())
+    out = helper.create_variable_for_type_inference(
+        helper.input_dtype(), stop_gradient=True)
+    helper.append_op(
+        type='hash',
+        inputs={'X': input},
+        outputs={'Out': out},
+        attrs={'num_hash': num_hash,
+               'mod_by': hash_size})
+    return out

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

+#   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
+
+
+class TestScaleOp(OpTest):
+    def setUp(self):
+        self.op_type = "hash"
+        self.init_test_case()
+        self.inputs = {'X': (self.in_seq, self.lod)}
+        self.attrs = {'num_hash': 4, 'mod_by': 10000}
+        self.outputs = {'Out': (self.out_seq, self.lod)}
+
+    def init_test_case(self):
+        np.random.seed = 1
+        self.in_seq = np.random.randint(0, 10, (30, 1)).astype("int32")
+        self.lod = [[9, 4, 11, 6]]
+        #  self.out_seq = np.ones([30, 4, 1], dtype=np.int32)
+        self.out_seq = [
+            [[9662], [9217], [1129], [8487]], [[9662], [9217], [1129], [8487]],
+            [[8310], [1327], [1654], [4567]], [[6897], [3218], [2013], [1241]],
+            [[9407], [6715], [6949], [8094]], [[8473], [694], [5142], [2479]],
+            [[8310], [1327], [1654], [4567]], [[6897], [3218], [2013], [1241]],
+            [[4372], [9456], [8204], [6695]], [[6897], [3218], [2013], [1241]],
+            [[8473], [694], [5142], [2479]], [[4372], [9456], [8204], [6695]],
+            [[4372], [9456], [8204], [6695]], [[8473], [694], [5142], [2479]],
+            [[9407], [6715], [6949], [8094]], [[9369], [4525], [8935], [9210]],
+            [[4372], [9456], [8204], [6695]], [[4372], [9456], [8204], [6695]],
+            [[9369], [4525], [8935], [9210]], [[6897], [3218], [2013], [1241]],
+            [[9038], [7951], [5953], [8657]], [[9407], [6715], [6949], [8094]],
+            [[9662], [9217], [1129], [8487]], [[9369], [4525], [8935], [9210]],
+            [[9038], [7951], [5953], [8657]], [[9662], [9217], [1129], [8487]],
+            [[9369], [4525], [8935], [9210]], [[1719], [5986], [9919], [3421]],
+            [[4372], [9456], [8204], [6695]], [[9038], [7951], [5953], [8657]]
+        ]
+        self.out_seq = np.array(self.out_seq)
+
+    def test_check_output(self):
+        self.check_output()
+
+
+if __name__ == "__main__":
+    unittest.main()