Merge pull request #7176 from pkuyym/fix-7173

Enhence shrink_rnn_memory_op.

paddle/operators/shrink_rnn_memory_op.cc

@@ -12,6 +12,7 @@ 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/lod_tensor.h"
 #include "paddle/operators/array_operator.h"
 #include "paddle/operators/math/math_function.h"
 
@@ -46,8 +47,21 @@ class ShrinkRNNMemoryOp : public ArrayOp {
     auto *out_var = scope.FindVar(Output("Out"));
     PADDLE_ENFORCE(out_var != nullptr, "Output Out must be set");
     auto &out_tensor = *out_var->GetMutable<framework::LoDTensor>();
+
+    size_t height = dst_num_rows;
+
+    // do shrink for the top level LoD
+    if (x_tensor.lod().size() > 0 &&
+        x_tensor.lod()[0].size() > static_cast<size_t>(dst_num_rows)) {
+      auto lod_offset = framework::GetSubLoDAndAbsoluteOffset(x_tensor.lod(), 0,
+                                                              dst_num_rows, 0);
+      height = lod_offset.second.second;
+      auto out_lod = out_tensor.mutable_lod();
+      framework::AppendLoD(out_lod, lod_offset.first);
+    }
+
     if (dst_num_rows != 0) {
-      out_tensor.ShareDataWith(x_tensor.Slice(0, dst_num_rows));
+      out_tensor.ShareDataWith(x_tensor.Slice(0, height));
     }
   }
 };
@@ -64,11 +78,11 @@ class ShrinkRNNMemoryOpProtoMaker : public framework::OpProtoAndCheckerMaker {
     AddOutput("Out", "(LoDTensor) The shrinked RNN step memory.");
     AddComment(
         R"DOC(
-        In dynamic RNN, we are able to handle sequences of different lengths. 
-        Because of the multiple lengths, the size of each step input can be 
+        In dynamic RNN, we are able to handle sequences of different lengths.
+        Because of the multiple lengths, the size of each step input can be
         different, which may lead to a mismatching between the input of
-        the current step and the memory generated by the previous one. This 
-        operator shrinks memory according to the size of the next step input, 
+        the current step and the memory generated by the previous one. This
+        operator shrinks memory according to the size of the next step input,
         to make sure that they can match each other.
         )DOC");
   }
@@ -132,6 +146,7 @@ class ShrinkRNNMemoryGradInferShape : public framework::InferShapeBase {
     PADDLE_ENFORCE(context->HasOutput(framework::GradVarName("X")));
     context->SetOutputDim(framework::GradVarName("X"),
                           context->GetInputDim("X"));
+    context->ShareLoD("X", framework::GradVarName("X"));
   }
 };
 

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

@@ -3,43 +3,86 @@ import paddle.v2.fluid.core as core
 from paddle.v2.fluid.executor import Executor
 import paddle.v2.fluid.layers as layers
 from paddle.v2.fluid.backward import append_backward
-from paddle.v2.fluid.framework import default_main_program
-import numpy
+from paddle.v2.fluid.framework import default_main_program, switch_main_program
+from paddle.v2.fluid.framework import Program
+import numpy as np
 
-main_program = default_main_program()
 
-
-class TestShrinkRNNMemory(unittest.TestCase):
-    def test_shrink_rnn_memory(self):
+class TestShrinkRNNMemoryBase(unittest.TestCase):
+    def setUp(self):
+        self.main_program = Program()
+        switch_main_program(self.main_program)
         x = layers.data('x', shape=[100], dtype='float32')
         x.stop_gradient = False
-        table = layers.lod_rank_table(x=x)
+        rank_table_tensor = layers.data(
+            'rank_table_tensor', shape=[1], dtype='float32', lod_level=1)
+        table = layers.lod_rank_table(x=rank_table_tensor)
         i = layers.zeros(dtype='int64', shape=[1])
-        mem1 = layers.shrink_memory(x=x, i=i, table=table)
+        self.mem1 = layers.shrink_memory(x=x, i=i, table=table)
         i = layers.increment(x=i)
         i.stop_gradient = True
-        mem2 = layers.shrink_memory(x=mem1, i=i, table=table)
+        self.mem2 = layers.shrink_memory(x=self.mem1, i=i, table=table)
         i = layers.increment(x=i)
         i.stop_gradient = True
-        mem3 = layers.shrink_memory(x=mem2, i=i, table=table)
+        self.mem3 = layers.shrink_memory(x=self.mem2, i=i, table=table)
+        mem3_mean = layers.mean(x=self.mem3)
+        append_backward(loss=mem3_mean)
+        self.x_grad = self.main_program.global_block().var('x@GRAD')
+
+    def sum_lodtensor(self, tensor):
+        sum_res = 0.0
+        for i in xrange(np.product(tensor.get_dims())):
+            sum_res += tensor.get_float_element(i)
+        return sum_res
 
+
+class TestShrinkRNNMemoryReferLoD(TestShrinkRNNMemoryBase):
+    def test_refer_lod(self):
         cpu = core.CPUPlace()
-        tensor = core.LoDTensor()
-        tensor.set_lod([[0, 2, 5, 6]])
-        tensor_np = numpy.random.random(size=(3, 100)).astype('float32')
-        tensor.set(tensor_np, cpu)
+        x_tensor = core.LoDTensor()
+        x_tensor.set_lod([[0, 2, 5, 6]])
+        tensor_np = np.random.random(size=(6, 100)).astype('float32')
+        x_tensor.set(tensor_np, cpu)
+
+        rank_table_tensor = core.LoDTensor()
+        rank_table_tensor.set_lod([[0, 1, 3, 6]])
+        rank_table_tensor.set(np.random.random(size=(6, 1)).astype('float32'),
+                              cpu)
+
         exe = Executor(cpu)
-        outs = exe.run(feed={'x': tensor}, fetch_list=[mem1, mem2, mem3])
-        self.assertTrue(numpy.allclose(tensor_np[0:3], outs[0]))
-        self.assertTrue(numpy.allclose(tensor_np[0:2], outs[1]))
-        self.assertTrue(numpy.allclose(tensor_np[0:1], outs[2]))
+        outs = exe.run(
+            feed={'x': x_tensor,
+                  'rank_table_tensor': rank_table_tensor},
+            fetch_list=[self.mem1, self.mem2, self.mem3, self.x_grad],
+            return_numpy=False)
+        self.assertTrue(np.allclose(tensor_np[0:6], outs[0]))
+        self.assertTrue(np.allclose(tensor_np[0:5], outs[1]))
+        self.assertTrue(np.allclose(tensor_np[0:2], outs[2]))
+        self.assertAlmostEqual(1.0, self.sum_lodtensor(outs[3]), delta=0.01)
 
-        mem3_mean = layers.mean(x=mem3)
-        append_backward(loss=mem3_mean)
-        x_grad = exe.run(
-            feed={'x': tensor},
-            fetch_list=[main_program.global_block().var('x@GRAD')])[0]
-        self.assertAlmostEqual(1.0, x_grad.sum(), delta=0.1)
+
+class TestShrinkRNNMemoryNoLoD(TestShrinkRNNMemoryBase):
+    def test_no_lod(self):
+        cpu = core.CPUPlace()
+        x_tensor = core.LoDTensor()
+        tensor_np = np.random.random(size=(3, 100)).astype('float32')
+        x_tensor.set(tensor_np, cpu)
+
+        rank_table_tensor = core.LoDTensor()
+        rank_table_tensor.set_lod([[0, 1, 3, 6]])
+        rank_table_tensor.set(np.random.random(size=(6, 1)).astype('float32'),
+                              cpu)
+
+        exe = Executor(cpu)
+        outs = exe.run(
+            feed={'x': x_tensor,
+                  'rank_table_tensor': rank_table_tensor},
+            fetch_list=[self.mem1, self.mem2, self.mem3, self.x_grad],
+            return_numpy=False)
+        self.assertTrue(np.allclose(tensor_np[0:3], outs[0]))
+        self.assertTrue(np.allclose(tensor_np[0:2], outs[1]))
+        self.assertTrue(np.allclose(tensor_np[0:1], outs[2]))
+        self.assertAlmostEqual(1.0, self.sum_lodtensor(outs[3]), delta=0.01)
 
 
 if __name__ == '__main__':