Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into convert

paddle/api/test/testMatrix.py

@@ -68,7 +68,7 @@ class TestMatrix(unittest.TestCase):
 
     def test_numpyCpu(self):
         numpy_mat = np.matrix([[1, 2], [3, 4], [5, 6]], dtype="float32")
-        m = swig_paddle.Matrix.createCpuDenseFromNumpy(numpy_mat, copy=False)
+        m = swig_paddle.Matrix.createCpuDenseFromNumpy(numpy_mat, False)
         self.assertEqual((int(m.getHeight()), int(m.getWidth())),
                          numpy_mat.shape)
 

paddle/api/test/testVector.py

@@ -43,7 +43,7 @@ class TestIVector(unittest.TestCase):
 
     def test_cpu_numpy(self):
         vec = np.array([1, 3, 4, 65, 78, 1, 4], dtype="int32")
-        iv = swig_paddle.IVector.createCpuVectorFromNumpy(vec, copy=False)
+        iv = swig_paddle.IVector.createCpuVectorFromNumpy(vec, False)
         self.assertEqual(vec.shape[0], int(iv.__len__()))
         vec[4] = 832
         for i in xrange(len(iv)):
@@ -106,7 +106,7 @@ class TestVector(unittest.TestCase):
 
     def testCpuNumpy(self):
         numpy_arr = np.array([1.2, 2.3, 3.4, 4.5], dtype="float32")
-        vec = swig_paddle.Vector.createCpuVectorFromNumpy(numpy_arr, copy=False)
+        vec = swig_paddle.Vector.createCpuVectorFromNumpy(numpy_arr, False)
         assert isinstance(vec, swig_paddle.Vector)
         numpy_arr[0] = 0.1
         for n, v in zip(numpy_arr, vec):

paddle/cuda/include/hl_matrix.h

@@ -69,19 +69,6 @@ extern void hl_sequence_softmax_forward(real* A_d,
                                         const int* index,
                                         int numSequence);
 
-/**
- * @brief   Matrix classification error.
- *
- * @param[in]   A_d     input matrix (M x N).
- * @param[in]   B_d     input vector (M x 1).
- * @param[out]  C_d     output vector (M x 1).
- * @param[in]   dimM    matrix height.
- * @param[in]   dimN    matrix width.
- *
- */
-extern void hl_matrix_classification_error(
-    real* A_d, int* B_d, real* C_d, int dimM, int dimN);
-
 /**
  * @brief   Matrix cross entropy.
  *

paddle/cuda/include/hl_top_k.h

@@ -58,4 +58,30 @@ extern void hl_sparse_matrix_top_k(real* topVal,
                                    int beamSize,
                                    int numSamples);
 
-#endif /* HL_TOP_K_H_ */
+/**
+ * @brief   Matrix classification error.
+ *
+ * @param[out]  topVal         top k element.
+ * @param[in]   ldv            leading dimension of topVal.
+ * @param[out]  topIds         top k index.
+ * @param[in]   src            input value.
+ * @param[in]   lds            leading dimension of src.
+ * @param[in]   dim            width of input value.
+ * @param[in]   topkSize       size of top k element.
+ * @param[in]   numSamples     height of input value.
+ * @param[in]   label          ground truth label.
+ * @param[out]  recResult      top-k classification error.
+ *
+ */
+extern void hl_matrix_classification_error(real* topVal,
+                                           int ldv,
+                                           int* topIds,
+                                           real* src,
+                                           int lds,
+                                           int dim,
+                                           int topkSize,
+                                           int numSamples,
+                                           int* label,
+                                           real* recResult);
+
+#endif  // HL_TOP_K_H_

paddle/cuda/include/stub/hl_matrix_stub.h

@@ -35,8 +35,16 @@ inline void hl_sequence_softmax_forward(real* A_d,
 inline void hl_matrix_softmax_derivative(
     real* grad_d, real* output_d, real* sftmaxSum_d, int dimM, int dimN) {}
 
-inline void hl_matrix_classification_error(
-    real* A_d, int* B_d, real* C_d, int dimM, int dimN) {}
+inline void hl_matrix_classification_error(real* topVal,
+                                           int ldv,
+                                           int* topIds,
+                                           real* src,
+                                           int lds,
+                                           int dim,
+                                           int topkSize,
+                                           int numSamples,
+                                           int* label,
+                                           real* recResult) {}
 
 inline void hl_matrix_cross_entropy(
     real* A_d, real* C_d, int* label_d, int dimM, int dimN) {}

paddle/cuda/src/hl_cuda_matrix.cu

@@ -265,59 +265,6 @@ void hl_matrix_softmax_derivative(real *grad_d,
   CHECK_SYNC("hl_matrix_softmax_derivative failed");
 }
 
-template<int blockSize>
-__global__ void KeMatrixClassificationError(real* in_A,
-                                            int* in_B,
-                                            real* out_C,
-                                            int dimN) {
-  __shared__ real max_s[blockSize];
-  __shared__ int max_l[blockSize];
-  const int tid = threadIdx.x;
-  const int rowId = blockIdx.x;
-
-  max_s[tid] = -1e30f;
-  in_A += rowId * dimN;
-  real tmp;
-  for (int colId = tid; colId < dimN; colId += blockSize) {
-    tmp = in_A[colId];
-    if (max_s[tid] < tmp) {
-      max_s[tid] = tmp;
-      max_l[tid] = colId;
-    }
-  }
-  __syncthreads();
-
-  for (int stride = blockSize/2; stride > 0; stride = stride/2) {
-    if (tid < stride) {
-      if (max_s[tid] < max_s[tid + stride]) {
-        max_s[tid] = max_s[tid + stride];
-        max_l[tid] = max_l[tid + stride];
-      }
-    }
-    __syncthreads();
-  }
-  __syncthreads();
-
-  if (tid == 0) {
-    out_C[rowId] = (max_l[0] == in_B[rowId] ? 0 : 1.0f);
-  }
-}
-
-void hl_matrix_classification_error(real* A_d,
-                                    int* B_d,
-                                    real* C_d,
-                                    int dimM,
-                                    int dimN) {
-  CHECK_NOTNULL(A_d);
-  CHECK_NOTNULL(B_d);
-  CHECK_NOTNULL(C_d);
-
-  // each sample is calculated by one block
-  KeMatrixClassificationError<1024><<< dimM, 1024, 0, STREAM_DEFAULT >>>
-    (A_d, B_d, C_d, dimN);
-  CHECK_SYNC("hl_matrix_classification_error");
-}
-
 __global__ void KeMatrixMultiBinaryCrossEntropy(real* output,
                                                 real* entropy,
                                                 int* row,

paddle/cuda/src/hl_top_k.cu

@@ -384,3 +384,81 @@ void hl_sparse_matrix_top_k(real* topVal, int ldv,
   CHECK_SYNC("hl_sparse_matrix_top_k failed");
 }
 
+/**
+ * Each block compute one sample.
+ * In a block:
+ * 1. every thread get top maxLength value;
+ * 2. merge to shTopK, block reduce and get max value;
+ * 3. go to the second setp, until one thread's topK value is null;
+ * 4. go to the first setp, until get the topK value.
+ */
+template<int maxLength, int blockSize>
+__global__ void KeMatrixTopKClassificationError(real* topVal, int ldv,
+                                                int * topIds,
+                                                real* src, int lds,
+                                                int dim,
+                                                int beamSize,
+                                                int* label,
+                                                real* recResult) {
+  __shared__ Pair shTopK[blockSize];
+  __shared__ int maxId[blockSize / 2];
+  const int tid = threadIdx.x;
+  const int warp = threadIdx.x / 32;
+  src += blockIdx.x * lds;
+  topVal += blockIdx.x * ldv;
+  topIds += blockIdx.x * beamSize;
+
+  Pair topK[maxLength]; // NOLINT
+  int beam = maxLength;
+  Pair max;
+  bool isEmpty = false;
+  bool firstStep = true;
+  int topkSize = beamSize;
+
+  for (int k = 0; k < maxLength; k++) {
+    topK[k].set(-HL_FLOAT_MAX, -1);
+  }
+
+  while (beamSize) {
+    threadGetTopK<maxLength, blockSize>
+      (topK, beam, beamSize, src, firstStep, isEmpty, max, dim, tid);
+
+    shTopK[tid] = topK[0];
+    blockReduce<maxLength, blockSize>
+      (shTopK, maxId, topK, &topVal, &topIds, beam, beamSize, tid, warp);
+  }
+
+  __syncthreads();
+  if (tid == 0) {
+    for (int i = 0; i < topkSize; i++) {
+        if (*--topIds == label[blockIdx.x]) {
+            recResult[blockIdx.x] = 0;
+            break;
+        }
+        recResult[blockIdx.x] = 1.0f;
+    }
+  }
+}
+
+void hl_matrix_classification_error(real* topVal, int ldv,
+                                   int* topIds,
+                                   real* src, int lds,
+                                   int dim,
+                                   int topkSize,
+                                   int numSamples,
+                                   int* label,
+                                   real* recResult) {
+  CHECK_NOTNULL(topVal);
+  CHECK_NOTNULL(topIds);
+  CHECK_NOTNULL(src);
+
+  if (topkSize > dim) topkSize = dim;
+
+  dim3 threads(256, 1);
+  dim3 grid(numSamples, 1);
+  KeMatrixTopKClassificationError<5, 256>
+  <<< grid, threads, 0, STREAM_DEFAULT >>>
+  (topVal, ldv, topIds, src, lds, dim, topkSize, label, recResult);
+
+  CHECK_SYNC("hl_matrix_top_k classification error failed");
+}

paddle/gserver/evaluators/Evaluator.cpp

@@ -39,6 +39,14 @@ void Evaluator::eval(const NeuralNetwork& nn) {
  */
 class ClassificationErrorEvaluator : public Evaluator {
 public:
+  /*
+  ClassificationErrorEvaluator() : totalScore2_(0) {}
+
+  virtual void start() {
+    Evaluator::start();
+    totalScore2_ = 0;
+    } */
+
   virtual void updateSamplesNum(const std::vector<Argument>& arguments) {
     if (3 == arguments.size()) {
       numSamples_ += arguments[2].value->getSum();
@@ -76,9 +84,11 @@ public:
                                               1,
                                               /* trans= */ false,
                                               useGpu(arguments[0].deviceId));
+
     errorMat->zeroMem();
+
     if (label != nullptr) {
-      errorMat->classificationError(*output, *label);
+      errorMat->classificationError(*output, *label, config_.top_k());
     } else if (dynamic_cast<CpuSparseMatrix*>(multiBinaryLabel.get()) ||
                dynamic_cast<GpuSparseMatrix*>(multiBinaryLabel.get())) {
       errorMat->classificationErrorMulti(
@@ -94,6 +104,16 @@ public:
     return errorMat;
   }
 
+  void printStats(std::ostream& os) const {
+    if (config_.top_k() == 1) {
+      os << config_.name() << "="
+         << (numSamples_ ? totalScore_ / numSamples_ : 0);
+    } else {
+      os << " top_" << config_.top_k()
+         << "_error=" << (numSamples_ ? totalScore_ / numSamples_ : 0);
+    }
+  }
+
   virtual real evalImp(std::vector<Argument>& arguments) {
     MatrixPtr errorMat = calcError(arguments);
     return errorMat->getSum();

paddle/gserver/layers/Layer.h

@@ -311,6 +311,7 @@ public:
         return *output->second;
       } else {
         LOG(FATAL) << "No specific output " << str;
+        return *((Argument*)nullptr);
       }
     }
   }

paddle/gserver/tests/test_Evaluator.cpp

@@ -129,6 +129,7 @@ void testEvaluatorAll(TestConfig testConf,
 TEST(Evaluator, classification_error) {
   TestConfig config;
   config.evaluatorConfig.set_type("classification_error");
+  config.evaluatorConfig.set_top_k(5);
 
   config.inputDefs.push_back({INPUT_DATA, "output", 50});
   config.inputDefs.push_back({INPUT_LABEL, "label", 50});

paddle/math/Matrix.cpp

@@ -732,6 +732,7 @@ void GpuMatrix::rowMax(IVector& maxIds, Matrix& maxVal) {
   size_t beam = maxVal.getWidth();
   CHECK_EQ(maxIds.getSize(), numSamples * beam);
   CHECK_EQ(maxVal.getHeight(), numSamples);
+  CHECK_EQ(maxVal.getWidth(), beam);
 
   hl_matrix_top_k(maxVal.getData(),
                   maxVal.getStride(),
@@ -792,19 +793,32 @@ void GpuMatrix::maxoutBackward(Matrix& a,
 }
 
 /*calulate the error of classification */
-void GpuMatrix::classificationError(Matrix& output, IVector& label) {
-  auto output_ptr = dynamic_cast<const GpuMatrix*>(&output);
-  auto label_ptr = dynamic_cast<const GpuIVector*>(&label);
-  CHECK(output_ptr && label_ptr) << "Invalid argument pointer";
-
-  CHECK(height_ == output_ptr->height_ && width_ == 1)
+void GpuMatrix::classificationError(Matrix& output,
+                                    IVector& label,
+                                    size_t topkSize) {
+  auto gpuOutput = dynamic_cast<GpuMatrix*>(&output);
+  auto gpuLabel = dynamic_cast<GpuIVector*>(&label);
+  size_t numSamples = this->getHeight();
+  GpuMatrixPtr gpuTopVal = std::make_shared<GpuMatrix>(numSamples, topkSize);
+  GpuIVectorPtr gpuTopIds = std::make_shared<GpuIVector>(numSamples * topkSize);
+
+  CHECK(gpuOutput && gpuLabel) << "Invalid argument pointer";
+  CHECK(gpuTopVal && gpuTopIds) << "Allocate GPU memory failed";
+  CHECK(gpuLabel->getSize() == numSamples) << "Vector size is not equal";
+  CHECK(numSamples == gpuOutput->getHeight() && this->getWidth() == 1)
       << "Matrix dimensions are not equal";
 
-  hl_matrix_classification_error((real*)output_ptr->data_,
-                                 (int*)label_ptr->getData(),
-                                 data_,
-                                 height_,
-                                 output_ptr->width_);
+  size_t dim = gpuOutput->getWidth();
+  hl_matrix_classification_error(gpuTopVal->getData(),
+                                 gpuTopVal->getStride(),
+                                 gpuTopIds->getData(),
+                                 gpuOutput->getData(),
+                                 gpuOutput->getStride(),
+                                 dim,
+                                 topkSize,
+                                 numSamples,
+                                 gpuLabel->getData(),
+                                 this->getData());
 }
 
 /* copy -log(output[i * width + label]) to this->data[i] */
@@ -3039,7 +3053,7 @@ void CpuMatrix::rowMax(Matrix& max) {
   max.maxRows(*this);
 }
 
-/* get beam size of max ids and values */
+/* Get the top k elements of each row of this matrix */
 void CpuMatrix::rowMax(IVector& maxIds, Matrix& maxVal) {
   CHECK(isContiguous());
   CHECK(!maxIds.useGpu() && !maxVal.useGpu()) << "Matrix type are not equal";
@@ -3047,6 +3061,7 @@ void CpuMatrix::rowMax(IVector& maxIds, Matrix& maxVal) {
   size_t beam = maxVal.getWidth();
   CHECK_EQ(maxIds.getSize(), numSamples * beam);
   CHECK_EQ(maxVal.getHeight(), numSamples);
+  CHECK_EQ(maxVal.getWidth(), beam);
 
   real* a = getData();
   int* s = maxIds.getData();
@@ -3198,32 +3213,39 @@ void CpuMatrix::rowNormalizeL1(Matrix& out) {
 }
 
 /* calulate classification error */
-void CpuMatrix::classificationError(Matrix& output, IVector& label) {
-  CHECK(dynamic_cast<const CpuMatrix*>(&output));
-  CHECK(dynamic_cast<const CpuIVector*>(&label));
+void CpuMatrix::classificationError(Matrix& output,
+                                    IVector& label,
+                                    size_t topkSize) {
+  size_t numSamples = this->getHeight();
+  auto cpuOutput = dynamic_cast<CpuMatrix*>(&output);
+  auto cpuLabel = dynamic_cast<CpuIVector*>(&label);
+  IVectorPtr cpuTopIds = std::make_shared<CpuIVector>(numSamples * topkSize);
+  MatrixPtr cpuTopVal = std::make_shared<CpuMatrix>(numSamples, topkSize);
+
+  CHECK(cpuOutput && cpuLabel) << "Invalid argument pointer";
+  CHECK(cpuTopIds && cpuTopVal) << "Allocate cpu memory failed";
+  CHECK(cpuLabel->getSize() == numSamples) << "Vector size is not equal";
+  CHECK(cpuOutput->getHeight() == numSamples && this->getWidth() == 1)
+      << "Matrix dimensions are not equal";
 
-  CHECK_EQ(getWidth(), (size_t)1);
-  size_t numSamples = getHeight();
-  CHECK_EQ(label.getSize(), numSamples);
-  CHECK_EQ(output.getHeight(), numSamples);
+  // top k matrix classification
+  cpuOutput->rowMax(*cpuTopIds, *cpuTopVal);
 
-  size_t dim = output.getWidth();
-  real* out = output.getData();
-  int* lbl = label.getData();
-  real maxData = 0.0;
-  int maxIndex = -1;
+  size_t dim = cpuOutput->getWidth();
+  real* result = this->getData();
+  int* ids = cpuTopIds->getData();
+  int* lbl = cpuLabel->getData();
   for (size_t i = 0; i < numSamples; ++i) {
     CHECK_GE(lbl[i], 0);
     CHECK_LT((size_t)lbl[i], dim);
-    maxData = out[i * dim];
-    maxIndex = 0;
-    for (size_t j = 0; j < dim; ++j) {
-      if (maxData < out[i * dim + j]) {
-        maxIndex = j;
-        maxData = out[i * dim + j];
+
+    for (size_t j = 0; j < topkSize; ++j) {
+      if (ids[j + i * topkSize] == lbl[i]) {
+        result[i] = 0;
+        break;
       }
+      result[i] = 1.0f;
     }
-    getData()[i] = (maxIndex != lbl[i]);
   }
 }
 

paddle/math/Matrix.h

@@ -836,8 +836,11 @@ public:
    * output[i] = 1 if row i is an error.
    *
    * output[i] = 0 if row i is correct.
+   *
    */
-  virtual void classificationError(Matrix& output, IVector& label) {
+  virtual void classificationError(Matrix& output,
+                                   IVector& label,
+                                   size_t topkSize = 1) {
     LOG(FATAL) << "Not implemented";
   }
 
@@ -1314,7 +1317,7 @@ public:
   void check(std::ostream& os, Matrix& refMat, bool printDiff = true);
   void randomizeUniform();
 
-  void classificationError(Matrix& output, IVector& label);
+  void classificationError(Matrix& output, IVector& label, size_t topkSize = 1);
 
   void convExpand(Matrix& feature,
                   int feaImgHeight,
@@ -1739,7 +1742,7 @@ public:
 
   void randomizeUniform();
 
-  void classificationError(Matrix& output, IVector& label);
+  void classificationError(Matrix& output, IVector& label, size_t topkSize = 1);
 
   void addByBitCode(size_t numClasses, const IVector& codes, const Matrix& vec);
 

paddle/math/tests/test_matrixCompare.cpp

@@ -764,7 +764,7 @@ TEST(Matrix, paramReluBackwardDiff) {
   }
 }
 
-void testClassificationError(int numSamples, int dim) {
+void testClassificationError(int numSamples, int dim, int topkSize) {
   MatrixPtr cpuError = std::make_shared<CpuMatrix>(numSamples, 1);
   MatrixPtr gpuError = std::make_shared<GpuMatrix>(numSamples, 1);
   MatrixPtr cpuOutput = std::make_shared<CpuMatrix>(numSamples, dim);
@@ -777,17 +777,22 @@ void testClassificationError(int numSamples, int dim) {
   gpuOutput->copyFrom(*cpuOutput);
   gpuLabel->copyFrom(*cpuLabel);
 
-  cpuError->classificationError(*cpuOutput, *cpuLabel);
-  gpuError->classificationError(*gpuOutput, *gpuLabel);
+  cpuError->classificationError(*cpuOutput, *cpuLabel, topkSize);
+  gpuError->classificationError(*gpuOutput, *gpuLabel, topkSize);
 
   TensorCheckEqual(*cpuError, *gpuError);
 }
 
 TEST(Matrix, classificationError) {
-  for (auto numSamples : {1, 10, 100, 1000, 70000}) {
-    for (auto dim : {1, 10, 100, 1000}) {
-      VLOG(3) << " numSamples=" << numSamples << " dim=" << dim;
-      testClassificationError(numSamples, dim);
+  for (auto numSamples : {1, 5, 31, 90, 150, 300}) {
+    for (auto dim :
+         {1, 5, 8, 10, 15, 64, 80, 120, 256, 300, 1280, 5120, 50000}) {
+      for (auto topkSize : {1, 5, 10, 20, 40, (int)rand() % dim + 1}) {
+        if (topkSize > dim) continue;
+        VLOG(3) << " sample= " << numSamples << " topkSize= " << topkSize
+                << " dim= " << dim;
+        testClassificationError(numSamples, dim, topkSize);
+      }
     }
   }
 }

paddle/parameter/Parameter.cpp

@@ -375,10 +375,6 @@ bool Parameter::load(const std::string& filename) {
   std::ifstream fs(filename, std::ios_base::binary);
   if (!fs) {
     LOG(INFO) << "missing parameters [" << filename << "] while loading model.";
-    if (isStatic()) {
-      LOG(FATAL) << getName() << " is static but missing, not allowed.";
-      return false;
-    }
     if (kMissParameterFail == FLAGS_load_missing_parameter_strategy) {
       LOG(FATAL) << getName() << " missing, not allowed.";
       return false;

paddle/setup.py.in

@@ -55,6 +55,9 @@ elif is_osx == True:
 
 include_dirs = [np.get_include(), "../"]    # include numpy and paddle.
 
+os.environ["CC"] = "@CMAKE_C_COMPILER@"
+os.environ["CXX"] = "@CMAKE_CXX_COMPILER@"
+
 setup(name="py_paddle",
   version="@PADDLE_VERSION@",
   ext_modules=[

proto/ModelConfig.proto

@@ -475,6 +475,10 @@ message EvaluatorConfig {
   // Used by ChunkEvaluator
   // chunk of these types are not counted
   repeated int32 excluded_chunk_types = 12;
+
+  // Used by ClassificationErrorEvaluator
+  // top # classification error
+  optional int32 top_k = 13 [default = 1];
 }
 
 message LinkConfig {

python/paddle/reader/decorator.py

@@ -12,18 +12,135 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-__all__ = ['buffered']
+__all__ = ['buffered', 'compose', 'chain', 'shuffle', 'ComposeNotAligned']
 
 from Queue import Queue
 from threading import Thread
+import itertools
+import random
+
+
+def shuffle(reader, buf_size):
+    """Creates a data reader whose data output is suffled.
+
+    Output from the iterator that created by original reader will be
+    buffered into shuffle buffer, and then shuffled. The size of shuffle buffer
+    is determined by argument buf_size.
+
+    Args:
+        reader: the original reader whose output will be
+            shuffled.
+        buf_size: shuffle buffer size.
+
+    Returns:
+        the new reader whose output is shuffled.
+    """
+
+    def data_reader():
+        buf = []
+        for e in reader():
+            buf.append(e)
+            if len(buf) >= buf_size:
+                random.shuffle(buf)
+                for b in buf:
+                    yield b
+                buf = []
+
+        if len(buf) > 0:
+            random.shuffle(buf)
+            for b in buf:
+                yield b
+
+    return data_reader
+
+
+def chain(*readers):
+    """Creates a data reader whose output is the outputs of input data
+       readers chained together.
+
+    If input readers output following data entries:
+    [0, 0, 0]
+    [1, 1, 1]
+    [2, 2, 2]
+    The chained reader will output:
+    [0, 0, 0, 1, 1, 1, 2, 2, 2]
+
+    Args:
+        readers: input readers.
+
+    Returns:
+        the new data reader.
+    """
+
+    def reader():
+        rs = []
+        for r in readers:
+            rs.append(r())
+
+        for e in itertools.chain(*rs):
+            yield e
+
+    return reader
+
+
+class ComposeNotAligned(ValueError):
+    pass
+
+
+def compose(*readers, **kwargs):
+    """Creates a data reader whose output is the combination of input readers.
+
+    If input readers output following data entries:
+    (1, 2)    3    (4, 5)
+    The composed reader will output:
+    (1, 2, 3, 4, 5)
+
+    Args:
+        *readers: readers that will be composed together.
+        check_alignment: If True, will check if input readers are aligned
+            correctly. If False, will not check alignment and trailing outputs
+            will be discarded. Defaults to True.
+
+    Returns:
+        the new data reader.
+
+    Raises:
+        ComposeNotAligned: outputs of readers are not aligned.
+            Will not raise when check_alignment is set to False.
+    """
+    check_alignment = kwargs.pop('check_alignment', True)
+
+    def make_tuple(x):
+        if isinstance(x, tuple):
+            return x
+        else:
+            return (x, )
+
+    def reader():
+        rs = []
+        for r in readers:
+            rs.append(r())
+        if not check_alignment:
+            for outputs in itertools.izip(*rs):
+                yield sum(map(make_tuple, outputs), ())
+        else:
+            for outputs in itertools.izip_longest(*rs):
+                for o in outputs:
+                    if o is None:
+                        # None will be not be present if compose is aligned
+                        raise ComposeNotAligned(
+                            "outputs of readers are not aligned.")
+                yield sum(map(make_tuple, outputs), ())
+
+    return reader
 
 
 def buffered(reader, size):
     """Creates a buffered data reader.
 
-    The buffered data reader will read and save data entries into a buffer.
-    Reading from the buffered data reader will proceed as long as the buffer
-    is not empty.
+    The buffered data reader will read and save data entries into a
+    buffer. Reading from the buffered data reader will proceed as long
+    as the buffer is not empty.
     
     Args:
         reader: the data reader to read from.
@@ -43,7 +160,7 @@ def buffered(reader, size):
             q.put(d)
         q.put(end)
 
-    def create_reader():
+    def data_reader():
         r = reader()
         q = Queue(maxsize=size)
         t = Thread(
@@ -57,4 +174,4 @@ def buffered(reader, size):
             yield e
             e = q.get()
 
-    return create_reader
+    return data_reader

python/paddle/reader/tests/decorator_test.py

@@ -16,16 +16,20 @@ import paddle.reader
 import time
 
 
-def reader_10(dur):
-    for i in range(10):
-        time.sleep(dur)
-        yield i
+def reader_creator_10(dur):
+    def reader():
+        for i in range(10):
+            # this invocation helps testing paddle.reader.buffer
+            time.sleep(dur)
+            yield i
+
+    return reader
 
 
 class TestBuffered(unittest.TestCase):
     def test_read(self):
         for size in range(20):
-            b = paddle.reader.buffered(lambda: reader_10(0), size)
+            b = paddle.reader.buffered(reader_creator_10(0), size)
             c = 0
             for i in b():
                 self.assertEqual(i, c)
@@ -34,7 +38,7 @@ class TestBuffered(unittest.TestCase):
 
     def test_buffering(self):
         # read have 30ms delay.
-        b = paddle.reader.buffered(lambda: reader_10(0.03), 10)
+        b = paddle.reader.buffered(reader_creator_10(0.03), 10)
         last_time = time.time()
         for idx, i in enumerate(b()):
             elapsed_time = time.time() - last_time
@@ -42,9 +46,63 @@ class TestBuffered(unittest.TestCase):
                 time.sleep(0.3)
             else:
                 # read time should be short, meaning already buffered.
-                self.assertLess(elapsed_time, 0.01)
+                self.assertLess(elapsed_time, 0.05)
             last_time = time.time()
 
 
+class TestCompose(unittest.TestCase):
+    def test_compse(self):
+        reader = paddle.reader.compose(
+            reader_creator_10(0), reader_creator_10(0))
+        for idx, e in enumerate(reader()):
+            self.assertEqual(e, (idx, idx))
+
+    def test_compose_not_aligned(self):
+        total = 0
+        reader = paddle.reader.compose(
+            paddle.reader.chain(reader_creator_10(0), reader_creator_10(0)),
+            reader_creator_10(0))
+        with self.assertRaises(paddle.reader.ComposeNotAligned):
+            for e in reader():
+                total += 1
+        # expecting 10, not 20
+        self.assertEqual(total, 10)
+
+    def test_compose_not_aligned_no_check(self):
+        total = 0
+        reader = paddle.reader.compose(
+            paddle.reader.chain(reader_creator_10(0), reader_creator_10(0)),
+            reader_creator_10(0),
+            check_alignment=False)
+        for e in reader():
+            total += 1
+        # expecting 10, not 20
+        self.assertEqual(total, 10)
+
+
+class TestChain(unittest.TestCase):
+    def test_chain(self):
+        c = paddle.reader.chain(reader_creator_10(0), reader_creator_10(0))
+        idx = 0
+        for e in c():
+            self.assertEqual(e, idx % 10)
+            idx += 1
+        self.assertEqual(idx, 20)
+
+
+class TestShuffle(unittest.TestCase):
+    def test_shuffle(self):
+        case = [(0, True), (1, True), (10, False), (100, False)]
+        a = reader_creator_10(0)
+        for size, checkEq in case:
+            s = paddle.reader.shuffle(a, size)
+            total = 0
+            for idx, e in enumerate(s()):
+                if checkEq:
+                    self.assertEqual(idx, e)
+                total += 1
+            self.assertEqual(total, 10)
+
+
 if __name__ == '__main__':
     unittest.main()

python/paddle/trainer/config_parser.py

@@ -1253,6 +1253,7 @@ def Evaluator(
         dict_file=None,
         result_file=None,
         num_results=None,
+        top_k=None,
         delimited=None,
         excluded_chunk_types=None, ):
     evaluator = g_config.model_config.evaluators.add()
@@ -1280,6 +1281,8 @@ def Evaluator(
         evaluator.result_file = result_file
     if num_results is not None:
         evaluator.num_results = num_results
+    if top_k is not None:
+        evaluator.top_k = top_k
     if delimited is not None:
         evaluator.delimited = delimited
 

python/paddle/trainer_config_helpers/evaluators.py

@@ -71,6 +71,7 @@ def evaluator_base(
         result_file=None,
         num_results=None,
         delimited=None,
+        top_k=None,
         excluded_chunk_types=None, ):
     """
     Evaluator will evaluate the network status while training/testing.
@@ -104,12 +105,15 @@ def evaluator_base(
     :param weight: An input layer which is a weight for each sample.
                    Each evaluator may calculate differently to use this weight.
     :type weight: LayerOutput.
+    :param top_k: number k in top-k error rate
+    :type top_k: int
     """
     # inputs type assertions.
     assert classification_threshold is None or isinstance(
         classification_threshold, float)
     assert positive_label is None or isinstance(positive_label, int)
     assert num_results is None or isinstance(num_results, int)
+    assert top_k is None or isinstance(top_k, int)
 
     if not isinstance(input, list):
         input = [input]
@@ -130,6 +134,8 @@ def evaluator_base(
         dict_file=dict_file,
         result_file=result_file,
         delimited=delimited,
+        num_results=num_results,
+        top_k=top_k,
         excluded_chunk_types=excluded_chunk_types, )
 
 
@@ -139,6 +145,7 @@ def classification_error_evaluator(input,
                                    label,
                                    name=None,
                                    weight=None,
+                                   top_k=None,
                                    threshold=None):
     """
     Classification Error Evaluator. It will print error rate for classification.
@@ -167,6 +174,8 @@ def classification_error_evaluator(input,
                   then means not set weight. The larger weight it is, the more
                   important this sample is.
     :type weight: LayerOutput
+    :param top_k: number k in top-k error rate
+    :type top_k: int
     :param threshold: The classification threshold.
     :type threshold: float
     :return: None.
@@ -178,6 +187,7 @@ def classification_error_evaluator(input,
         input=input,
         label=label,
         weight=weight,
+        top_k=top_k,
         classification_threshold=threshold, )
 
 

python/paddle/trainer_config_helpers/layers.py

@@ -2870,8 +2870,8 @@ def gru_step_layer(input,
     :param name:
     :param gate_act:
     :param bias_attr:
-    :param param_attr: the parameter_attribute for transforming the output_mem 
-                       from previous step. 
+    :param param_attr: the parameter_attribute for transforming the output_mem
+                       from previous step.
     :param layer_attr:
     :return: LayerOutput object.
     :rtype: LayerOutput
@@ -2882,10 +2882,10 @@ def gru_step_layer(input,
     Layer(
         name=name,
         type=LayerType.GRU_STEP_LAYER,
-        # The parameter here is for transforming the output_mem. The input has 
-        # already been transformed outside this module so it does not need 
-        # parameter associated with it. 
-        # The parameter here is instead grouped with input is due to 
+        # The parameter here is for transforming the output_mem. The input has
+        # already been transformed outside this module so it does not need
+        # parameter associated with it.
+        # The parameter here is instead grouped with input is due to
         # backward model compatibility.
         inputs=[Input(input.name, **param_attr.attr), output_mem.name],
         bias=ParamAttr.to_bias(bias_attr),
@@ -3536,6 +3536,7 @@ def classification_cost(input,
                         label,
                         weight=None,
                         name=None,
+                        top_k=None,
                         evaluator=classification_error_evaluator,
                         layer_attr=None):
     """
@@ -3550,6 +3551,8 @@ def classification_cost(input,
     :param weight: The weight affects the cost, namely the scale of cost.
                    It is an optional argument.
     :type weight: LayerOutput
+    :param top_k: number k in top-k error rate
+    :type top_k: int
     :param evaluator: Evaluator method.
     :param layer_attr: layer's extra attribute.
     :type layer_attr: ExtraLayerAttribute
@@ -3577,7 +3580,7 @@ def classification_cost(input,
         assert isinstance(e.for_classification, bool)
         assert e.for_classification
 
-        e(name=e.__name__, input=input, label=label, weight=weight)
+        e(name=e.__name__, input=input, label=label, weight=weight, top_k=top_k)
 
     if not isinstance(evaluator, collections.Sequence):
         evaluator = [evaluator]

python/paddle/v2/layer.py

@@ -77,7 +77,9 @@ import data_type
 
 __all__ = [
     'parse_network', 'data', 'fc', 'max_id', 'classification_cost',
-    'cross_entropy_cost'
+    'cross_entropy_cost', 'cross_entropy_with_selfnorm_cost', 'regression_cost',
+    'multi_binary_label_cross_entropy_cost', 'rank_cost', 'lambda_cost',
+    'sum_cost', 'huber_cost'
 ]
 
 
@@ -137,7 +139,8 @@ def __convert_to_v2__(method_name, name_prefix, parent_names):
             parent_layers = dict()
             other_kwargs = dict()
             for pname in parent_names:
-                parent_layers[pname] = kwargs[pname]
+                if kwargs.has_key(pname):
+                    parent_layers[pname] = kwargs[pname]
 
             for key in kwargs.keys():
                 if key not in parent_names:
@@ -189,27 +192,61 @@ class DataLayerV2(Layer):
 data = DataLayerV2
 fc = __convert_to_v2__('fc_layer', name_prefix='fc', parent_names=['input'])
 max_id = __convert_to_v2__(
-    'maxid_layer', name_prefix='maxid_layer', parent_names=['input'])
+    'maxid_layer', name_prefix='maxid', parent_names=['input'])
 classification_cost = __convert_to_v2__(
     'classification_cost',
     name_prefix='classification_cost',
-    parent_names=['input', 'label'])
+    parent_names=['input', 'label', 'weight'])
+regression_cost = __convert_to_v2__(
+    'regression_cost',
+    name_prefix='regression_cost',
+    parent_names=['input', 'label', 'weight'])
 cross_entropy_cost = __convert_to_v2__(
     'cross_entropy',
     name_prefix='cross_entropy',
     parent_names=['input', 'label'])
+cross_entropy_with_selfnorm_cost = __convert_to_v2__(
+    'cross_entropy_with_selfnorm',
+    name_prefix='cross_entropy_with_selfnorm',
+    parent_names=['input', 'label'])
+multi_binary_label_cross_entropy_cost = __convert_to_v2__(
+    'multi_binary_label_cross_entropy',
+    name_prefix='multi_binary_label_cross_entropy',
+    parent_names=['input', 'label'])
+rank_cost = __convert_to_v2__(
+    'rank_cost',
+    name_prefix='rank_cost',
+    parent_names=['left', 'right', 'label', 'weight'])
+lambda_cost = __convert_to_v2__(
+    'lambda_cost', name_prefix='lambda_cost', parent_names=['input', 'score'])
+sum_cost = __convert_to_v2__(
+    'sum_cost', name_prefix='sum_cost', parent_names=['input'])
+huber_cost = __convert_to_v2__(
+    'huber_cost', name_prefix='huber_cost', parent_names=['input', 'label'])
 
 if __name__ == '__main__':
     pixel = data(name='pixel', type=data_type.dense_vector(784))
     label = data(name='label', type=data_type.integer_value(10))
+    weight = data(name='weight', type=data_type.dense_vector(10))
+    score = data(name='score', type=data_type.dense_vector(1))
+
     hidden = fc(input=pixel, size=100, act=conf_helps.SigmoidActivation())
     inference = fc(input=hidden, size=10, act=conf_helps.SoftmaxActivation())
     maxid = max_id(input=inference)
     cost1 = classification_cost(input=inference, label=label)
-    cost2 = cross_entropy_cost(input=inference, label=label)
+    cost2 = classification_cost(input=inference, label=label, weight=weight)
+    cost3 = cross_entropy_cost(input=inference, label=label)
+    cost4 = cross_entropy_with_selfnorm_cost(input=inference, label=label)
+    cost5 = regression_cost(input=inference, label=label)
+    cost6 = regression_cost(input=inference, label=label, weight=weight)
+    cost7 = multi_binary_label_cross_entropy_cost(input=inference, label=label)
+    cost8 = rank_cost(left=score, right=score, label=score)
+    cost9 = lambda_cost(input=inference, score=score)
+    cost10 = sum_cost(input=inference)
+    cost11 = huber_cost(input=score, label=label)
 
-    print parse_network(cost1)
-    print parse_network(cost2)
     print parse_network(cost1, cost2)
-    print parse_network(cost2)
+    print parse_network(cost3, cost4)
+    print parse_network(cost5, cost6)
+    print parse_network(cost7, cost8, cost9, cost10, cost11)
     print parse_network(inference, maxid)