Compelete max_sequence_len_op (#5913)

Author: JiayiFeng

paddle/operators/CMakeLists.txt

@@ -200,6 +200,7 @@ set(DEPS_OPS
     lod_rank_table_op
     lod_tensor_to_array_op
     array_to_lod_tensor_op
+    max_sequence_len_op
     lstm_op
     tensor_array_read_write_op
     gru_op
@@ -222,6 +223,7 @@ op_library(pool_with_index_op DEPS pooling)
 op_library(lod_rank_table_op SRCS lod_rank_table_op.cc DEPS lod_rank_table)
 op_library(lod_tensor_to_array_op SRCS lod_tensor_to_array_op.cc DEPS lod_rank_table_op)
 op_library(array_to_lod_tensor_op SRCS array_to_lod_tensor_op.cc DEPS lod_rank_table_op)
+op_library(max_sequence_len_op SRCS max_sequence_len_op.cc DEPS lod_rank_table)
 op_library(tensor_array_read_write_op SRCS tensor_array_read_write_op.cc)
 if(WITH_GPU)
 op_library(nccl_op DEPS nccl_common)

paddle/operators/max_sequence_len_op.cc

@@ -0,0 +1,66 @@
+/* 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/framework/lod_rank_table.h"
+#include "paddle/framework/op_registry.h"
+#include "paddle/framework/operator.h"
+
+namespace paddle {
+namespace operators {
+
+class MaxSeqenceLenOp : public framework::OperatorBase {
+ public:
+  MaxSeqenceLenOp(const std::string &type,
+                  const framework::VariableNameMap &inputs,
+                  const framework::VariableNameMap &outputs,
+                  const framework::AttributeMap &attrs)
+      : OperatorBase(type, inputs, outputs, attrs) {}
+
+  void Run(const framework::Scope &scope,
+           const platform::DeviceContext &dev_ctx) const override {
+    auto &rank_table =
+        scope.FindVar(Input("RankTable"))->Get<framework::LoDRankTable>();
+    auto *out =
+        scope.FindVar(Output("Out"))->GetMutable<framework::LoDTensor>();
+    int64_t *out_ptr = out->mutable_data<int64_t>({1}, platform::CPUPlace());
+    *out_ptr = rank_table.items()[0].length;
+  }
+};
+
+class MaxSeqenceLenOpProtoMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  MaxSeqenceLenOpProtoMaker(framework::OpProto *proto,
+                            framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("RankTable", "The lod_rank_table.");
+    AddOutput("Out", "The max sequence length.");
+    AddComment(
+        R"DOC(Calculate the max sequence length through lod_rank_table.)DOC");
+  }
+};
+
+class MaxSeqenceLenInferShape : public framework::InferShapeBase {
+ public:
+  void operator()(framework::InferShapeContext *context) const override {
+    PADDLE_ENFORCE(context->HasInput("RankTable"));
+    context->SetOutputDim("Out", {1});
+  }
+};
+}  // namespace operators
+}  // namespace paddle
+
+REGISTER_OPERATOR(max_sequence_len, paddle::operators::MaxSeqenceLenOp,
+                  paddle::operators::MaxSeqenceLenOpProtoMaker,
+                  paddle::operators::MaxSeqenceLenInferShape,
+                  paddle::framework::EmptyGradOpMaker);

python/paddle/v2/fluid/layers.py

@@ -1354,6 +1354,20 @@ def lod_rank_table(x, level=0, main_program=None):
     return table
 
 
+def max_sequence_len(rank_table, main_program=None):
+    """
+    This function creates an operator to calculate the length of 
+    max seqence through input rank_table(should be a lod_rank_table)
+    """
+    helper = LayerHelper("max_seqence_len", **locals())
+    res = helper.create_tmp_variable(dtype="int64")
+    helper.append_op(
+        type="max_sequence_len",
+        inputs={"RankTable": rank_table},
+        outputs={"Out": res})
+    return res
+
+
 def topk(input, k, main_program=None, startup_program=None):
     helper = LayerHelper('topk', **locals())
     topk_out = helper.create_tmp_variable(dtype=input.data_type)

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

@@ -18,7 +18,11 @@ def test_lod_tensor_to_array_level_0(self):
         tensor.set_lod([[0, 3, 9, 10]])
         expect = map(lambda x: numpy.array(x).astype('int32'),
                      [[3, 0, 9], [4, 1], [5, 2], [6], [7], [8]])
-        self.main(tensor=tensor, expect_array=expect, expect_lod=[] * 6)
+        self.main(
+            tensor=tensor,
+            expect_array=expect,
+            expect_lod=[] * 6,
+            expect_max_len=6)
 
     def test_lod_tensor_to_array_level_0_empty_seq(self):
         tensor = core.LoDTensor()
@@ -27,7 +31,11 @@ def test_lod_tensor_to_array_level_0_empty_seq(self):
         tensor.set_lod([[0, 3, 9, 9, 10]])
         expect = map(lambda x: numpy.array(x).astype('int32'),
                      [[3, 0, 9], [4, 1], [5, 2], [6], [7], [8]])
-        self.main(tensor=tensor, expect_array=expect, expect_lod=[] * 6)
+        self.main(
+            tensor=tensor,
+            expect_array=expect,
+            expect_lod=[] * 6,
+            expect_max_len=6)
 
     def test_lod_tensor_to_array_level_1(self):
         tensor = core.LoDTensor()
@@ -44,7 +52,11 @@ def test_lod_tensor_to_array_level_1(self):
         ]
 
         lod = [[[0, 2, 5]], [[0, 6, 12]], [[0, 3]]]
-        self.main(tensor=tensor, expect_array=expect, expect_lod=lod)
+        self.main(
+            tensor=tensor,
+            expect_array=expect,
+            expect_lod=lod,
+            expect_max_len=3)
 
     def test_lod_tensor_to_array_level_1_empty_seq(self):
         tensor = core.LoDTensor()
@@ -63,7 +75,11 @@ def test_lod_tensor_to_array_level_1_empty_seq(self):
         ]
 
         lod = [[[0, 5, 8, 8, 15]], [[0, 2, 6, 7, 8]], [[0, 2, 6]], [[0, 2]]]
-        self.main(tensor=tensor, expect_array=expect, expect_lod=lod)
+        self.main(
+            tensor=tensor,
+            expect_array=expect,
+            expect_lod=lod,
+            expect_max_len=4)
 
     def test_lod_tensor_to_array_level_2(self):
         tensor = core.LoDTensor()
@@ -80,22 +96,33 @@ def test_lod_tensor_to_array_level_2(self):
         ]
         lod = [[[0, 1, 3, 4], [0, 1, 4, 8, 12]],
                [[0, 4, 7], [0, 1, 5, 9, 17, 21, 27, 31]], [[0, 2], [0, 6, 7]]]
-        self.main(tensor=tensor, expect_array=expect, expect_lod=lod)
+        self.main(
+            tensor=tensor,
+            expect_array=expect,
+            expect_lod=lod,
+            expect_max_len=3)
 
     def test_lod_tensor_to_array_level_2_skip_level(self):
         tensor = core.LoDTensor()
         tensor.set(
             numpy.arange(50).reshape(50, 1).astype('int32'), self.place())
         tensor.set_lod([[0, 2, 5, 6], [0, 2, 5, 6, 10, 12, 13],
                         [0, 3, 7, 11, 17, 21, 22, 23, 27, 31, 39, 45, 46, 50]])
-        self.main(tensor=tensor, expect_array=None, expect_lod=None, level=1)
-
-    def main(self, tensor, expect_array, expect_lod, level=0):
+        self.main(
+            tensor=tensor,
+            expect_array=None,
+            expect_lod=None,
+            expect_max_len=4,
+            level=1)
+
+    def main(self, tensor, expect_array, expect_lod, expect_max_len, level=0):
         place = self.place()
         program = Program()
         x = layers.data(name='x', shape=[10], main_program=program)
         x.persistable = True
         table = layers.lod_rank_table(x, level=level, main_program=program)
+        max_len = layers.max_sequence_len(table, main_program=program)
+        max_len.persistable = True
         array = layers.lod_tensor_to_array(x, table, main_program=program)
         array.persistable = True
 
@@ -110,6 +137,10 @@ def main(self, tensor, expect_array, expect_lod, level=0):
             self.check_array_same(array, expect_array, expect_lod)
         self.check_tensor_same(scope.find_var(result.name).get_tensor(), tensor)
 
+        self.assertEqual(
+            numpy.array(scope.find_var(max_len.name).get_tensor())[0],
+            expect_max_len)
+
     def check_array_same(self, array, expect_tensor, expect_lod):
         self.assertEqual(len(expect_tensor), len(array))
         for i, exp in enumerate(zip(expect_tensor, expect_lod)):