GPU implementation of row conv.

paddle/function/CMakeLists.txt

@@ -28,6 +28,7 @@ if(WITH_TESTING)
     add_simple_unittest(PadOpTest)
     add_simple_unittest(MulOpTest)
     add_simple_unittest(CosSimOpTest)
+    add_simple_unittest(RowConvOpTest)
 endif()
 endif()
 

paddle/function/RowConvOp.cpp

@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "RowConvOp.h"
+#include <iostream>
 #include "paddle/math/Vector.h"
 
 namespace paddle {
@@ -127,10 +128,8 @@ public:
     RowConv<Device>(outMat, inMat, wMat, seqId);
   }
 };
-
 /**
- * \brief The backward propagation of padding Function. Remove the elements
+ * \brief TODO(qingqing)
- *        in the padding positions of forward.
  *
  * Argument in this Function:
  */
@@ -158,7 +157,37 @@ public:
                      : typename Tensor<real, Device>::Matrix(nullptr, 0, 0);
     const auto seqId = in.getSequenceId().vector<int, Device>();
 
+    std::cout << "in:" << std::endl;
+    for (int i = 0; i < inMat.getHeight(); ++i) {
+      for (int j = 0; j < inMat.getWidth(); ++j) {
+        std::cout << outGMat.getElement(i, j) << " ";
+      }
+      std::cout << std::endl;
+    }
+
+    std::cout << "w:" << std::endl;
+    for (int i = 0; i < wMat.getHeight(); ++i) {
+      for (int j = 0; j < wMat.getWidth(); ++j) {
+        std::cout << wMat.getElement(i, j) << " ";
+      }
+      std::cout << std::endl;
+    }
+
+    std::cout << "w:" << std::endl;
+    for (int i = 0; i < seqId.getSize(); ++i) {
+      std::cout << seqId.getElement(i) << " ";
+    }
+    std::cout << std::endl;
+
     RowConvGrad<Device>(outGMat, inMat, wMat, inGMat, wGMat, seqId);
+
+    std::cout << std::endl << "out:" << std::endl;
+    for (int i = 0; i < inGMat.getHeight(); ++i) {
+      for (int j = 0; j < inGMat.getWidth(); ++j) {
+        std::cout << inGMat.getElement(i, j) << " ";
+      }
+      std::cout << std::endl;
+    }
   }
 };
 
@@ -166,7 +195,7 @@ REGISTER_TYPED_FUNC(RowConv, CPU, RowConvFunc);
 REGISTER_TYPED_FUNC(RowConvGrad, CPU, RowConvGradFunc);
 #ifndef PADDLE_ONLY_CPU
 REGISTER_TYPED_FUNC(RowConv, GPU, RowConvFunc);
-REGISTER_TYPED_FUNC(RowConvGrad, GPU, PadGradFunc);
+REGISTER_TYPED_FUNC(RowConvGrad, GPU, RowConvGradFunc);
 #endif
 
 }  // namespace paddle

paddle/function/RowConvOpTest.cpp

@@ -0,0 +1,69 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+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 <gtest/gtest.h>
+#include "FunctionTest.h"
+
+namespace paddle {
+
+void testRowConvFw(size_t batchSize, size_t dim, size_t contextLength) {
+  FunctionCompare test("RowConv", FuncConfig());
+
+  test.addSequence(SequenceIdArg(TensorShape{batchSize}));
+  test.addInputs(SequenceArg(VALUE_TYPE_FLOAT, TensorShape{batchSize, dim}));
+  test.addInputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{contextLength, dim}));
+
+  test.addOutputs(SequenceArg(VALUE_TYPE_FLOAT, TensorShape{batchSize, dim}),
+                  ADD_TO);
+
+  test.run();
+}
+
+void testRowConvBw(size_t batchSize, size_t dim, size_t contextLength) {
+  FunctionCompare test("RowConvGrad", FuncConfig());
+
+  test.addSequence(SequenceIdArg(TensorShape{batchSize}));
+  test.addInputs(SequenceArg(VALUE_TYPE_FLOAT, TensorShape{batchSize, dim}));
+  test.addInputs(SequenceArg(VALUE_TYPE_FLOAT, TensorShape{batchSize, dim}));
+  test.addInputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{contextLength, dim}));
+
+  test.addOutputs(SequenceArg(VALUE_TYPE_FLOAT, TensorShape{batchSize, dim}),
+                  ADD_TO);
+  test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{contextLength, dim}),
+                  ADD_TO);
+
+  test.run();
+}
+
+TEST(RowConv, real) {
+  // for (size_t numSamples : {17, 129}) {
+  //   for (size_t dim : {16, 248}) {
+  //     for (size_t context: {3, 7, 65}) {
+  LOG(INFO) << "===========";
+  // for (size_t numSamples : {17}) {
+  //  for (size_t dim : {16}) {
+  //    for (size_t context: {3}) {
+  size_t numSamples = 17;
+  size_t dim = 16;
+  size_t context = 3;
+  LOG(INFO) << " numSamples=" << numSamples << " dim=" << dim
+            << " context length=" << context;
+  testRowConvFw(numSamples, dim, context);
+  // testRowConvBw(numSamples, dim, context);
+  //     }
+  //   }
+  // }
+}
+
+}  // namespace paddle