Add ReorderLoDTensorByRank

paddle/operators/reorder_lod_tensor_by_rank_op.cc

@@ -0,0 +1,234 @@
+/* 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/operators/detail/safe_ref.h"
+
+namespace paddle {
+namespace operators {
+
+class ReorderLoDTensorByRankTableOpProtoMaker
+    : public framework::OpProtoAndCheckerMaker {
+ public:
+  ReorderLoDTensorByRankTableOpProtoMaker(OpProto *proto,
+                                          OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "(LoDTensor) the input lod tensor need to be reordered.");
+    AddInput("RankTable",
+             "(LoDRankTable) the rank table that input need follow");
+    AddOutput("Out", "(LoDTensor) reordered lod tensor");
+    AddComment(R"DOC(ReorderLoDTensorByRankTable
+
+Reorder the input X by the rank of `RankTable`. If `RankTable` is ordered by
+index [3, 0, 2, 1]. Input X will reorder its sequence, the third sequence of
+X will be the first sequence of Output.
+
+NOTE: The RankTable does not need to be calculated by X.
+
+For example:
+The X = [Seq0, Seq1, Seq2, Seq3]. The indices of RankTable are [3, 0, 2, 1].
+
+The Out =  [Seq3, Seq0, Seq2, Seq1] with correct LoD information.
+)DOC");
+  }
+};
+
+class ReorderLoDTensorByRankTableBase : public framework::OperatorBase {
+ public:
+  ReorderLoDTensorByRankTableBase(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 &x =
+        detail::Ref(scope.FindVar(Input("X")),
+                    "Cannot find input lod tensor variable %s", Input("X"))
+            .Get<framework::LoDTensor>();
+    auto &rank_table = detail::Ref(scope.FindVar(Input("RankTable")),
+                                   "Cannot find input rank table variable %s",
+                                   Input("RankTable"))
+                           .Get<framework::LoDRankTable>();
+    auto &out =
+        *detail::Ref(scope.FindVar(Output("Out")),
+                     "Cannot find output lod tensor variable %s", Output("Out"))
+             .GetMutable<framework::LoDTensor>();
+
+    out.Resize(x.dims());
+    out.mutable_data(x.place(), x.type());
+    this->process(dev_ctx, x, rank_table, &out);
+  }
+
+ protected:
+  virtual void process(const platform::DeviceContext &dev_ctx,
+                       const framework::LoDTensor &x,
+                       const framework::LoDRankTable &rank_table,
+                       framework::LoDTensor *out) const = 0;
+
+  struct AbsoluteRankTableItem {
+    size_t offset;  // the absolute/accumulated offset.
+    size_t length;  // the length
+    framework::LoD lod;
+  };
+
+  std::vector<AbsoluteRankTableItem> GetAbsoluteOffsetAndLengthByLoDRankTable(
+      const framework::LoDTensor &x) const {
+    std::vector<AbsoluteRankTableItem> absolute_table;
+    size_t level = 0;
+    size_t size = x.lod()[level].size();
+
+    for (size_t i = 0; i < size - 1; ++i) {
+      auto lod_offset =
+          framework::GetSubLoDAndAbsoluteOffset(x.lod(), i, i + 1, level);
+
+      auto &offset = lod_offset.second;
+
+      absolute_table.emplace_back();
+      absolute_table.back().length = offset.second - offset.first;
+      absolute_table.back().offset = offset.first;
+      absolute_table.back().lod = lod_offset.first;
+    }
+    return absolute_table;
+  }
+
+  size_t CopyTensorAndLod(const platform::DeviceContext &dev_ctx,
+                          const AbsoluteRankTableItem &item,
+                          const framework::LoDTensor &x,
+                          framework::LoDTensor *out, size_t out_offset) const {
+    auto &out_lod = *out->mutable_lod();
+    auto len = item.length;
+    auto x_offset = item.offset;
+
+    if (out_lod.empty()) {
+      for (size_t i = 0; i < item.lod.size(); ++i) {
+        out_lod.push_back(std::vector<size_t>({0}));
+      }
+    }
+
+    for (size_t i = 0; i < out_lod.size(); ++i) {
+      auto &out_v = out_lod[i];
+      auto &new_lod_v = item.lod[i];
+
+      for (auto &detail : new_lod_v) {
+        out_v.push_back(out_v.back() + detail);
+      }
+    }
+
+    auto x_sliced = x.Slice(x_offset, x_offset + len);
+    auto out_sliced = out->Slice(out_offset, out_offset + len);
+
+    framework::CopyFrom(x_sliced, out_sliced.place(), dev_ctx, &out_sliced);
+    out_offset += len;
+    return out_offset;
+  }
+};
+
+class ReorderLoDTensorByRankTableOp : public ReorderLoDTensorByRankTableBase {
+ public:
+  ReorderLoDTensorByRankTableOp(const std::string &type,
+                                const framework::VariableNameMap &inputs,
+                                const framework::VariableNameMap &outputs,
+                                const framework::AttributeMap &attrs)
+      : ReorderLoDTensorByRankTableBase(type, inputs, outputs, attrs) {}
+
+ protected:
+  void process(const platform::DeviceContext &dev_ctx,
+               const framework::LoDTensor &x,
+               const framework::LoDRankTable &rank_table,
+               framework::LoDTensor *out) const override {
+    auto absolute_table = GetAbsoluteOffsetAndLengthByLoDRankTable(x);
+    size_t out_offset = 0;
+    out->mutable_lod()->clear();
+    for (auto &item : rank_table.items()) {
+      PADDLE_ENFORCE_LT(item.index, absolute_table.size());
+      out_offset = CopyTensorAndLod(dev_ctx, absolute_table[item.index], x, out,
+                                    out_offset);
+    }
+  }
+};
+
+class IdentityInferShape : public framework::InferShapeBase {
+ public:
+  void operator()(framework::InferShapeContext *context) const override {
+    context->SetOutputDim("Out", context->GetInputDim("X"));
+  }
+};
+
+class ReorderLodTensorByRankGradOpMaker
+    : public framework::SingleGradOpDescMaker {
+ public:
+  using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
+
+ protected:
+  std::unique_ptr<framework::OpDescBind> Apply() const override {
+    auto *grad_op = new framework::OpDescBind();
+    grad_op->SetType("reorder_lod_tensor_by_rank_grad");
+    grad_op->SetInput("X", OutputGrad("Out"));
+    grad_op->SetOutput("Out", InputGrad("X"));
+    grad_op->SetInput("RankTable", Input("RankTable"));
+    return std::unique_ptr<framework::OpDescBind>(grad_op);
+  }
+};
+
+class ReorderLoDTensorByRankGradOp : public ReorderLoDTensorByRankTableBase {
+ public:
+  ReorderLoDTensorByRankGradOp(const std::string &type,
+                               const framework::VariableNameMap &inputs,
+                               const framework::VariableNameMap &outputs,
+                               const framework::AttributeMap &attrs)
+      : ReorderLoDTensorByRankTableBase(type, inputs, outputs, attrs) {}
+
+ protected:
+  void process(const platform::DeviceContext &dev_ctx,
+               const framework::LoDTensor &x,
+               const framework::LoDRankTable &rank_table,
+               framework::LoDTensor *out) const override {
+    auto absolute_table = GetAbsoluteOffsetAndLengthByLoDRankTable(x);
+
+    // offsets = enumerate([item.index for item in rank_table.items()])
+    std::vector<std::pair<size_t, size_t>> offsets;
+    offsets.reserve(rank_table.items().size());
+    for (size_t i = 0; i < rank_table.items().size(); ++i) {
+      offsets.push_back({i, rank_table.items()[i].index});
+    }
+
+    // offsets.sort(key=lambda x: x[1])
+    std::sort(
+        offsets.begin(), offsets.end(),
+        [](const std::pair<size_t, size_t> &a,
+           const std::pair<size_t, size_t> &b) { return a.second < b.second; });
+
+    // Copy TensorAndLod
+    size_t out_offset = 0;
+    for (auto &offset : offsets) {
+      out_offset = this->CopyTensorAndLod(dev_ctx, absolute_table[offset.first],
+                                          x, out, out_offset);
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+
+REGISTER_OPERATOR(reorder_lod_tensor_by_rank,
+                  ops::ReorderLoDTensorByRankTableOp,
+                  ops::ReorderLodTensorByRankGradOpMaker,
+                  ops::ReorderLoDTensorByRankTableOpProtoMaker,
+                  ops::IdentityInferShape);
+REGISTER_OPERATOR(reorder_lod_tensor_by_rank_grad,
+                  ops::ReorderLoDTensorByRankGradOp, ops::IdentityInferShape);

python/paddle/v2/fluid/framework.py

@@ -389,7 +389,10 @@ def find_name(var_list, name):
                             % (in_proto.name, len(in_args)))
                     in_arg_names = []
                     for arg in in_args:
-                        in_arg_names.append(arg.name)
+                        if isinstance(arg, basestring):
+                            in_arg_names.append(arg)
+                        else:
+                            in_arg_names.append(arg.name)
                     self.desc.set_input(in_proto.name, in_arg_names)
                 else:
                     self.desc.set_input(in_proto.name, [])

python/paddle/v2/fluid/layer_helper.py

@@ -194,3 +194,9 @@ def _get_default_initializer(self, dtype):
         else:
             # For integer and boolean types, initialize with all zeros
             return Constant()
+
+    def is_instance(self, param_name, cls):
+        param = self.kwargs.get(param_name, None)
+        if not isinstance(param, cls):
+            raise TypeError("The input {0} parameter of method {1} must be {2}",
+                            param_name, self.layer_type, cls.__name__)

python/paddle/v2/fluid/layers/control_flow.py

@@ -3,14 +3,15 @@
 from .. import core
 from tensor import assign, fill_constant
 import contextlib
+from ..registry import autodoc
 
 __all__ = [
     'split_lod_tensor', 'merge_lod_tensor', 'BlockGuard', 'StaticRNNGuard',
     'StaticRNNMemoryLink', 'WhileGuard', 'While', 'lod_rank_table',
     'max_sequence_len', 'topk', 'lod_tensor_to_array', 'array_to_lod_tensor',
     'increment', 'array_write', 'create_array', 'less_than', 'array_read',
     'shrink_memory', 'array_length', 'IfElse', 'DynamicRNN', 'ConditionalBlock',
-    'StaticRNN'
+    'StaticRNN', 'reorder_lod_tensor_by_rank'
 ]
 
 
@@ -981,3 +982,18 @@ def _assert_in_rnn_block_(self, method):
         if self.status != DynamicRNN.IN_RNN:
             raise ValueError("{0} can only be invoked inside rnn block.".format(
                 method))
+
+
+@autodoc
+def reorder_lod_tensor_by_rank(x, rank_table):
+    helper = LayerHelper('reorder_lod_tensor_by_rank', **locals())
+    helper.is_instance('x', Variable)
+    helper.is_instance('rank_table', Variable)
+
+    out = helper.create_tmp_variable(dtype=x.dtype)
+    helper.append_op(
+        type='reorder_lod_tensor_by_rank',
+        inputs={'X': [x],
+                'RankTable': [rank_table]},
+        outputs={'Out': [out]})
+    return out

python/paddle/v2/fluid/registry.py

@@ -8,7 +8,7 @@
 from framework import OpProtoHolder, Variable, Program, Operator
 from paddle.v2.fluid.layer_helper import LayerHelper, unique_name
 
-__all__ = ['deprecated', 'register_layer']
+__all__ = ['deprecated', 'register_layer', 'autodoc']
 
 
 def _convert_(name):
@@ -175,12 +175,18 @@ def func_wrapper(*args, **kwargs):
         """
         Wrap func with deprecated warning
         """
-        warnings.simplefilter('always', DeprecationWarning)  #turn off filter
+        warnings.simplefilter('always', DeprecationWarning)  # turn off filter
         warnings.warn(
             "Call to deprecated function {}.".format(func.__name__),
             category=DeprecationWarning,
             stacklevel=2)
-        warnings.simplefilter('default', DeprecationWarning)  #reset filter
+        warnings.simplefilter('default', DeprecationWarning)  # reset filter
         return func(*args, **kwargs)
 
     return func_wrapper
+
+
+def autodoc(func):
+    func.__doc__ = _generate_doc_string_(OpProtoHolder.instance().get_op_proto(
+        func.__name__))
+    return func

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

@@ -0,0 +1,47 @@
+import unittest
+import paddle.v2.fluid as fluid
+import numpy
+
+
+class TestReorderLoDTensor(unittest.TestCase):
+    def test_reorder(self):
+        dat = fluid.layers.data(name='input', shape=[1], lod_level=2)
+        dat.stop_gradient = False
+        rank_dat = fluid.layers.data(name='ref', shape=[1], lod_level=1)
+        table = fluid.layers.lod_rank_table(rank_dat)
+        new_dat = fluid.layers.reorder_lod_tensor_by_rank(
+            x=dat, rank_table=table)
+        loss = fluid.layers.mean(x=new_dat)
+        fluid.backward.append_backward_ops(loss=loss)
+
+        cpu = fluid.CPUPlace()
+        exe = fluid.Executor(cpu)
+        exe.run(fluid.default_startup_program())
+
+        ref = fluid.Tensor()
+        ref_lod = [0, 3, 4, 7, 8, 14]
+        ref.set_lod([ref_lod])
+
+        ref.set(numpy.random.random(size=[14, 1]).astype('float32'), cpu)
+        input = fluid.Tensor()
+        lod_level_0 = numpy.random.randint(low=1, high=5, size=5)
+        lod_level_0 = [0] + numpy.cumsum(lod_level_0).tolist()
+        lod_level_1 = numpy.random.randint(low=1, high=5, size=lod_level_0[-1])
+        lod_level_1 = [0] + numpy.cumsum(lod_level_1).tolist()
+
+        input.set_lod([lod_level_0, lod_level_1])
+        input.set(
+            numpy.random.random(size=[lod_level_1[-1], 1]).astype('float32'),
+            cpu)
+
+        ig = exe.run(fluid.default_main_program(),
+                     feed={'input': input,
+                           'ref': ref},
+                     fetch_list=['input@GRAD'],
+                     return_numpy=False)[0]
+        self.assertAlmostEqual(numpy.array(ig).sum(), 1.0, delta=0.001)
+        self.assertEqual(input.lod(), ig.lod())
+
+
+if __name__ == '__main__':
+    unittest.main()