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

demo/image_classification/prediction.py

@@ -126,7 +126,7 @@ class ImageClassifier():
         # For oversampling, average predictions across crops.
         # If not, the shape of output[name]: (1, class_number),
         # the mean is also applicable.
-        return output[output_layer].mean(0)
+        return output[output_layer]['value'].mean(0)
 
     def predict(self, image=None, output_layer=None):
         assert isinstance(image, basestring)

demo/mnist/api_train_v2.py

-import numpy
 import paddle.v2 as paddle
 
 import mnist_util
@@ -27,19 +26,14 @@ def main():
     cost = paddle.layer.classification_cost(input=inference, label=label)
 
     parameters = paddle.parameters.create(cost)
-    for param_name in parameters.keys():
-        array = parameters.get(param_name)
-        array[:] = numpy.random.uniform(low=-1.0, high=1.0, size=array.shape)
-        parameters.set(parameter_name=param_name, value=array)
 
     adam_optimizer = paddle.optimizer.Adam(learning_rate=0.01)
 
     def event_handler(event):
         if isinstance(event, paddle.event.EndIteration):
-            para = parameters.get('___fc_2__.w0')
-            print "Pass %d, Batch %d, Cost %f, Weight Mean Of Fc 2 is %f" % (
-                event.pass_id, event.batch_id, event.cost, para.mean())
-
+            if event.batch_id % 100 == 0:
+                print "Pass %d, Batch %d, Cost %f, %s" % (
+                    event.pass_id, event.batch_id, event.cost, event.metrics)
         else:
             pass
 
@@ -50,11 +44,12 @@ def main():
                   parameters=parameters,
                   event_handler=event_handler,
                   batch_size=32,  # batch size should be refactor in Data reader
-                  data_types={  # data_types will be removed, It should be in
+                  data_types=[  # data_types will be removed, It should be in
                       # network topology
-                      'pixel': images.type,
-                      'label': label.type
-                  })
+                      ('pixel', images.type),
+                      ('label', label.type)],
+                  reader_dict={'pixel':0, 'label':1}
+                  )
 
 
 if __name__ == '__main__':

demo/model_zoo/resnet/classify.py

@@ -156,7 +156,7 @@ class ImageClassifier():
             # For oversampling, average predictions across crops.
             # If not, the shape of output[name]: (1, class_number),
             # the mean is also applicable.
-            res[name] = output[name].mean(0)
+            res[name] = output[name]['value'].mean(0)
 
         return res
 

doc/api/trainer_config_helpers/layers.rst

@@ -139,24 +139,12 @@ lstmemory
     :members: lstmemory
     :noindex:
 
-lstm_step_layer
----------------
-..  automodule:: paddle.trainer_config_helpers.layers
-    :members: lstm_step_layer
-    :noindex:
-
 grumemory
 ---------
 ..  automodule:: paddle.trainer_config_helpers.layers
     :members: grumemory
     :noindex:
 
-gru_step_layer
----------------
-..  automodule:: paddle.trainer_config_helpers.layers
-    :members: gru_step_layer
-    :noindex:
-
 Recurrent Layer Group
 =====================
 
@@ -172,6 +160,18 @@ recurrent_group
     :members: recurrent_group
     :noindex:
     
+lstm_step_layer
+---------------
+..  automodule:: paddle.trainer_config_helpers.layers
+    :members: lstm_step_layer
+    :noindex:
+
+gru_step_layer
+---------------
+..  automodule:: paddle.trainer_config_helpers.layers
+    :members: gru_step_layer
+    :noindex:
+
 beam_search
 ------------
 ..  automodule:: paddle.trainer_config_helpers.layers
@@ -308,6 +308,12 @@ repeat_layer
     :members: repeat_layer
     :noindex:
 
+rotate_layer
+------------
+..  automodule:: paddle.trainer_config_helpers.layers
+    :members: rotate_layer
+    :noindex:
+
 seq_reshape_layer
 -----------------
 ..  automodule:: paddle.trainer_config_helpers.layers
@@ -462,6 +468,12 @@ ctc_layer
     :members: ctc_layer
     :noindex:
 
+warp_ctc_layer
+--------------
+..  automodule:: paddle.trainer_config_helpers.layers
+    :members: warp_ctc_layer
+    :noindex:
+
 nce_layer
 -----------
 ..  automodule:: paddle.trainer_config_helpers.layers

paddle/api/Arguments.cpp

@@ -38,6 +38,13 @@ Arguments* Arguments::createByPaddleArgumentVector(void* ptr) {
   return args;
 }
 
+Arguments* Arguments::createByPaddleArgument(const void* ptr) {
+  auto p = (paddle::Argument*)(ptr);
+  auto args = new Arguments();
+  args->m->outputs.push_back(*p);
+  return args;
+}
+
 Matrix* Arguments::getSlotValue(size_t idx) const throw(RangeError) {
   auto& a = m->getArg(idx);
   return Matrix::createByPaddleMatrixPtr(&a.value);

paddle/api/GradientMachine.cpp

@@ -144,12 +144,12 @@ Parameter* GradientMachine::getParameter(size_t i) throw(RangeError) {
 
 void GradientMachine::randParameters() { m->machine->randParameters(); }
 
-Matrix* GradientMachine::getLayerOutput(const std::string& layerName) const
+Arguments* GradientMachine::getLayerOutput(const std::string& layerName) const
     throw(UnsupportError) {
-  auto nn = std::dynamic_pointer_cast<paddle::NeuralNetwork>(m->machine);
+  auto nn = m->machine;
   if (nn) {
-    auto mat = nn->getLayerOutput(layerName);
-    return Matrix::createByPaddleMatrixPtr(&mat);
+    auto arg = nn->getLayerOutput(layerName);
+    return Arguments::createByPaddleArgument(&arg);
   } else {
     throw UnsupportError();
   }

paddle/api/PaddleAPI.h

@@ -454,6 +454,7 @@ public:
 
 private:
   static Arguments* createByPaddleArgumentVector(void* ptr);
+  static Arguments* createByPaddleArgument(const void* ptr);
   void* getInternalArgumentsPtr() const;
 
 private:
@@ -769,7 +770,7 @@ public:
 
   void randParameters();
 
-  Matrix* getLayerOutput(const std::string& layerName) const
+  Arguments* getLayerOutput(const std::string& layerName) const
       throw(UnsupportError);
 
   /**
@@ -956,7 +957,7 @@ public:
 
   Arguments* getForwardOutput();
 
-  Matrix* getLayerOutput(const std::string& layerName);
+  Arguments* getLayerOutput(const std::string& layerName) const;
 };
 
 /// the N-Best results generated from one input sequence.

paddle/api/Trainer.cpp

@@ -131,12 +131,11 @@ void Trainer::testOneDataBatch(size_t batchSize, const Arguments& args) {
 void TrainerPrivate::finishTestPeriod() { tester_->finishTestPeriod(); }
 void Trainer::finishTestPeriod() { m->finishTestPeriod(); }
 
-Matrix* Trainer::getLayerOutput(const std::string& layerName) {
-  auto nn = std::dynamic_pointer_cast<paddle::NeuralNetwork>(
-      this->m->getGradientMachine());
+Arguments* Trainer::getLayerOutput(const std::string& layerName) const {
+  auto nn = this->m->getGradientMachine();
   CHECK(nn) << "trainerInternal_.getGradientMachine() is not NeuralNetwork";
-  auto m = nn->getLayerOutput(layerName);
-  return Matrix::createByPaddleMatrixPtr(&m);
+  auto arg = nn->getLayerOutput(layerName);
+  return Arguments::createByPaddleArgument(&arg);
 }
 
 void Trainer::forwardOneBatch(size_t batchSize) {

paddle/gserver/gradientmachines/GradientMachine.h

@@ -134,6 +134,10 @@ public:
     backward(callback);
   }
 
+  virtual Argument getLayerOutput(const std::string& layerName) {
+    return *((Argument*)nullptr);
+  }
+
   // see comment in Layer.h for the function with the same name
   virtual void resetState() {}
 

paddle/gserver/gradientmachines/MultiGradientMachine.cpp

@@ -282,6 +282,18 @@ void MultiGradientMachine::forwardBackward(const std::vector<Argument>& inArgs,
   backwardImp(callback);
 }
 
+Argument MultiGradientMachine::getLayerOutput(const std::string& layerName) {
+  std::vector<Argument> args;
+  args.reserve(threads_.size());
+
+  for (auto& thread : threads_) {
+    args.push_back(thread->getGradientMachine()->getLayerOutput(layerName));
+  }
+  outLayerArgs_.concat(args, false /* use_gpu */, outArgStream_, passType_);
+
+  return outLayerArgs_;
+}
+
 void MultiGradientMachine::backwardImp(const UpdateCallback& callback) {
   for (size_t i = 0; i < parameters_.size(); i++) {
     if (!parameters_[i]->useGpu() || parameters_[i]->isStatic()) continue;

paddle/gserver/gradientmachines/MultiGradientMachine.h

@@ -189,6 +189,8 @@ public:
                        PassType passType,
                        const UpdateCallback& callback);
 
+  virtual Argument getLayerOutput(const std::string& layerName);
+
   virtual void onPassEnd();
 
   virtual void finish();
@@ -314,6 +316,8 @@ protected:
   std::vector<Argument> outArgs_;
   hl_stream_t outArgStream_;
 
+  Argument outLayerArgs_;
+
   /// ParameterType which needs to be merged from each GPU
   std::vector<ParameterType> mergeTypes_;
   int numDevices_;         /* number of gpu devices */

paddle/gserver/gradientmachines/NeuralNetwork.cpp

@@ -293,11 +293,10 @@ void NeuralNetwork::backward(const UpdateCallback& callback) {
   }
 }
 
-MatrixPtr NeuralNetwork::getLayerOutput(const std::string& layerName) {
-  auto it = layerMap_.find(layerName);
-  CHECK(it != layerMap_.end()) << "Cannot find layer: " << layerName;
-  return it->second->getOutputValue();
+Argument NeuralNetwork::getLayerOutput(const std::string& layerName) {
+  return getLayer(layerName)->getOutput();
 }
+
 void NeuralNetwork::onPassEnd() {
   for (auto& layer : layers_) {
     layer->onPassEnd();

paddle/gserver/gradientmachines/NeuralNetwork.h

@@ -87,7 +87,8 @@ public:
 
   virtual void backward(const UpdateCallback& callback = nullptr);
 
-  MatrixPtr getLayerOutput(const std::string& layerName);
+  virtual Argument getLayerOutput(const std::string& layerName);
+
   const LayerPtr& getLayer(const std::string& layerName) const {
     auto it = layerMap_.find(layerName);
     CHECK(it != layerMap_.end()) << "Unknown layer " << layerName;

paddle/gserver/layers/CosSimLayer.cpp

@@ -42,7 +42,7 @@ void CosSimLayer::forward(PassType passType) {
   /* malloc memory for the output_ if necessary */
   int batchSize = getInputValue(0)->getHeight();
   int size = getSize();
-  CHECK_EQ(forward_.size(), 1) << "Only one forward function needed";
+  CHECK_EQ(forward_.size(), 1UL) << "Only one forward function needed";
 
   {
     REGISTER_TIMER_INFO("CosFwResetTimer", getName().c_str());
@@ -68,7 +68,7 @@ void CosSimLayer::forward(PassType passType) {
 void CosSimLayer::backward(const UpdateCallback& callback) {
   /* activation */ {
     REGISTER_TIMER_INFO("CosBpAtvTimer", getName().c_str());
-    CHECK_EQ(backward_.size(), 1) << "Only one backward function needed";
+    CHECK_EQ(backward_.size(), 1UL) << "Only one backward function needed";
 
     const auto outG = this->getOutputGrad();
     const auto outV = this->getOutputValue();

paddle/gserver/layers/CosSimVecMatLayer.cpp

@@ -112,7 +112,7 @@ bool CosSimVecMatLayer::init(const LayerMap& layerMap,
 
 void CosSimVecMatLayer::forward(PassType passType) {
   Layer::forward(passType);
-  CHECK_EQ(forward_.size(), 1) << "Only one forward function needed";
+  CHECK_EQ(forward_.size(), 1UL) << "Only one forward function needed";
 
   MatrixPtr inV0 = getInputValue(0);
   MatrixPtr inV1 = getInputValue(1);
@@ -145,7 +145,7 @@ void CosSimVecMatLayer::forward(PassType passType) {
 }
 
 void CosSimVecMatLayer::backward(const UpdateCallback& callback) {
-  CHECK_EQ(backward_.size(), 1) << "Only one forward function needed";
+  CHECK_EQ(backward_.size(), 1UL) << "Only one forward function needed";
 
   MatrixPtr inV0 = getInputValue(0);
   MatrixPtr inV1 = getInputValue(1);

paddle/math/tests/test_RowBuffer.cpp

@@ -17,10 +17,10 @@ limitations under the License. */
 
 TEST(RowBuffer, testAutoGrow) {
   paddle::RowBuffer buf(128);
-  ASSERT_EQ(128, buf.getWidth());
+  ASSERT_EQ(128UL, buf.getWidth());
   ASSERT_TRUE(buf.isAutoGrowth());
   buf.resize(2);
-  ASSERT_EQ(2, buf.getRowCount());
+  ASSERT_EQ(2UL, buf.getRowCount());
   for (size_t i = 0; i < buf.getWidth() * 2; ++i) {
     buf.data()[i] = i;
   }
@@ -35,7 +35,7 @@ TEST(RowBuffer, testAutoGrow) {
     data[i] = i;
   }
 
-  ASSERT_EQ(3, buf.getRowCount());
+  ASSERT_EQ(3UL, buf.getRowCount());
   for (size_t i = 0; i < buf.getRowCount() - 1; ++i) {
     for (size_t j = 0; j < buf.getWidth(); ++j) {
       ASSERT_NEAR(i * buf.getWidth() + j, buf.get(i)[j], 1e-5);
@@ -51,7 +51,7 @@ TEST(RowBuffer, testWithMemBuf) {
       std::make_shared<paddle::CpuMemoryHandle>(128 * 2 * sizeof(real));
   paddle::RowBuffer buf(mem, 128);
   ASSERT_TRUE(!buf.isAutoGrowth());
-  ASSERT_EQ(2, buf.getRowCount());
+  ASSERT_EQ(2UL, buf.getRowCount());
   for (size_t i = 0; i < buf.getWidth() * 2; ++i) {
     buf.data()[i] = i;
   }

paddle/py_paddle/dataprovider_converter.py

@@ -23,7 +23,8 @@ __all__ = ['DataProviderConverter']
 class IScanner(object):
     def __init__(self, input_type, pos):
         self.input_type = input_type
-        assert isinstance(self.input_type, dp2.InputType)
+        if not isinstance(self.input_type, dp2.InputType):
+            raise ValueError("input type should be dataprovider2.InputType")
         self.pos = pos
 
     def scan(self, dat):
@@ -50,7 +51,6 @@ class DenseScanner(IScanner):
 
     def finish_scan(self, argument):
         assert isinstance(argument, swig_paddle.Arguments)
-        assert isinstance(self.input_type, dp2.InputType)
         if self.__mat__.dtype != numpy.float32:
             self.__mat__ = self.__mat__.astype(numpy.float32)
         m = swig_paddle.Matrix.createDenseFromNumpy(self.__mat__, True, False)
@@ -63,7 +63,6 @@ class SparseBinaryScanner(IScanner):
         self.__rows__ = [0]
         self.__cols__ = []
         self.__height__ = 0
-        self.__nnz__ = 0
         self.__value__ = []
 
     def scan(self, dat):
@@ -76,7 +75,6 @@ class SparseBinaryScanner(IScanner):
 
     def finish_scan(self, argument):
         assert isinstance(argument, swig_paddle.Arguments)
-        assert isinstance(self.input_type, dp2.InputType)
         m = swig_paddle.Matrix.createSparse(self.__height__,
                                             self.input_type.dim,
                                             len(self.__cols__),

paddle/py_paddle/util.py

@@ -208,7 +208,7 @@ def __monkeypatch_gradient_machine__():
 
         output = dict()
         for name in layerNames:
-            output[name] = __matrix_to_numpy__(self.getLayerOutput(name))
+            output[name] = __arguments_to_numpy__(0, self.getLayerOutput(name))
         return output
 
     swig_paddle.GradientMachine.getLayerOutputs = getLayerOutputs

python/paddle/trainer_config_helpers/layers.py

@@ -112,6 +112,8 @@ __all__ = [
     'priorbox_layer',
     'spp_layer',
     'pad_layer',
+    'eos_layer',
+    'layer_support',
 ]
 
 
@@ -708,6 +710,7 @@ class MixedLayerType(LayerOutput):
         # update the size which might be computed inside MixedLayer
         # according to the operator's output size
         self.size = ml.config.size
+        self.finalized = True
 
 
 @wrap_name_default("mixed")
@@ -1287,6 +1290,12 @@ def last_seq(input,
     """
     Get Last Timestamp Activation of a sequence.
 
+    The simple usage is:
+
+    .. code-block:: python
+
+       seq = last_seq(input=layer)
+
     :param agg_level: Aggregated level
     :param name: Layer name.
     :type name: basestring
@@ -1325,6 +1334,12 @@ def first_seq(input,
     """
     Get First Timestamp Activation of a sequence.
 
+    The simple usage is:
+
+    .. code-block:: python
+
+       seq = first_seq(input=layer)
+
     :param agg_level: aggregation level
     :param name: Layer name.
     :type name: basestring
@@ -1425,7 +1440,7 @@ def repeat_layer(input, num_repeats, name=None, layer_attr=None):
 
     .. code-block:: python
 
-       expand = repeat_layer(layer, 4)
+       expand = repeat_layer(input=layer, num_repeats=4)
 
     :param input: Input layer
     :type input: LayerOutput
@@ -1797,6 +1812,12 @@ def cos_sim(a, b, scale=1, size=1, name=None, layer_attr=None):
     Note that the above computation is for one sample. Multiple samples are
     processed in one batch.
 
+    The example usage is:
+
+    .. code-block:: python
+
+       cos = cos_sim(a=layer1, b=layer2, size=3)
+
     :param name: layer name
     :type name: basestring
     :param a: input layer a
@@ -1958,6 +1979,16 @@ def img_conv_layer(input,
     pieces. First 256/4 = 64 channels will process by first 32 filters. The
     rest channels will be processed by rest group of filters.
 
+    The example usage is:
+
+    ..  code-block:: python
+
+        conv = img_conv_layer(input=data, filter_size=1, filter_size_y=1,
+                              num_channels=8,
+                              num_filters=16, stride=1,
+                              bias_attr=False,
+                              act=ReluActivation())
+
     :param name: Layer name.
     :type name: basestring
     :param input: Layer Input.
@@ -2097,6 +2128,34 @@ def img_pool_layer(input,
 
     .. _pooling: http://ufldl.stanford.edu/tutorial/supervised/Pooling/
 
+    - ceil_mode=True:
+
+    ..  math::
+
+        w = 1 + int(ceil(input\_width + 2 * padding - pool\_size) / float(stride))
+        h = 1 + int(ceil(input\_height + 2 * padding\_y - pool\_size\_y) / float(stride\_y))
+
+    - ceil_mode=False:
+
+    ..  math::
+
+        w = 1 + int(floor(input\_width + 2 * padding - pool\_size) / float(stride))
+        h = 1 + int(floor(input\_height + 2 * padding\_y - pool\_size\_y) / float(stride\_y))
+
+    The example usage is:
+
+    ..  code-block:: python
+
+        maxpool = img_pool_layer(input=conv,
+                                 pool_size=3,
+                                 pool_size_y=5,
+                                 num_channels=8,
+                                 stride=1,
+                                 stride_y=2,
+                                 padding=1,
+                                 padding_y=2,
+                                 pool_type=MaxPooling())
+
     :param padding: pooling padding width.
     :type padding: int
     :param padding_y: pooling padding height. It's equal to padding by default.
@@ -2123,19 +2182,6 @@ def img_pool_layer(input,
     :param ceil_mode: Wether to use ceil mode to calculate output height and with.
                       Defalut is True. If set false, Otherwise use floor.
 
-                      - ceil_mode=True:
-
-                      ..  math::
-
-                          w = 1 + int(ceil(input_width + 2 * padding - pool_size) / float(stride))
-                          h = 1 + int(ceil(input_height + 2 * padding_y - pool_size_y) / float(stride_y))
-
-                      - ceil_mode=False:
-
-                      ..  math::
-
-                          w = 1 + int(floor(input_width + 2 * padding - pool_size) / float(stride))
-                          h = 1 + int(floor(input_height + 2 * padding_y - pool_size_y) / float(stride_y))
     :type ceil_mode: bool
     :return: LayerOutput object.
     :rtype: LayerOutput
@@ -2197,6 +2243,15 @@ def spp_layer(input,
     The details please refer to
     `Kaiming He's paper <https://arxiv.org/abs/1406.4729>`_.
 
+    The example usage is:
+
+    ..  code-block:: python
+
+        spp = spp_layer(input=data, 
+                        pyramid_height=2, 
+                        num_channels=16, 
+                        pool_type=MaxPooling())
+
     :param name: layer name.
     :type name: basestring
     :param input: layer's input.
@@ -2285,6 +2340,12 @@ def img_cmrnorm_layer(input,
     The details please refer to
     `Alex's paper <http://www.cs.toronto.edu/~fritz/absps/imagenet.pdf>`_.
 
+    The example usage is:
+
+    ..  code-block:: python
+    
+        norm = img_cmrnorm_layer(input=net, size=5)
+
     :param name: layer name.
     :type name: None|basestring
     :param input: layer's input.
@@ -2340,6 +2401,12 @@ def batch_norm_layer(input,
     The details of batch normalization please refer to this
     `paper <http://arxiv.org/abs/1502.03167>`_.
 
+    The example usage is:
+
+    ..  code-block:: python
+    
+        norm = batch_norm_layer(input=net, act=ReluActivation())
+
     :param name: layer name.
     :type name: basestring
     :param input: batch normalization input. Better be linear activation.
@@ -3907,13 +3974,13 @@ def conv_shift_layer(a, b, name=None, layer_attr=None):
 
     .. code-block:: python
 
-       conv_shift = conv_shift_layer(input=[layer1, layer2])
+       conv_shift = conv_shift_layer(a=layer1, b=layer2)
 
     :param name: layer name
     :type name: basestring
     :param a: Input layer a.
     :type a: LayerOutput
-    :param b: input layer b
+    :param b: input layer b.
     :type b: LayerOutput
     :param layer_attr: layer's extra attribute.
     :type layer_attr: ExtraLayerAttribute
@@ -4005,8 +4072,8 @@ def tensor_layer(a,
 @wrap_act_default()
 @layer_support()
 def selective_fc_layer(input,
-                       select,
                        size,
+                       select=None,
                        act=None,
                        name=None,
                        pass_generation=False,
@@ -4033,6 +4100,7 @@ def selective_fc_layer(input,
     :type input: LayerOutput|list|tuple
     :param select: The select layer. The output of select layer should be a
                    sparse binary matrix, and treat as the mask of selective fc.
+                   If is None, acts exactly like fc_layer.
     :type select: LayerOutput
     :param size: The layer dimension.
     :type size: int
@@ -4261,7 +4329,7 @@ def block_expand_layer(input,
 
     .. code-block:: python
 
-       block_expand = block_expand_layer(input,
+       block_expand = block_expand_layer(input=layer,
                                          num_channels=128,
                                          stride_x=1,
                                          stride_y=1,
@@ -4465,7 +4533,7 @@ def warp_ctc_layer(input,
         - You can set 'blank' to any value ranged in [0, num_classes], which
           should be consistent as that used in your labels.
         - As a native 'softmax' activation is interated to the warp-ctc library,
-         'linear' activation is expected instead in the 'input' layer.
+          'linear' activation is expected instead in the 'input' layer.
 
     The simple usage:
 
@@ -4598,6 +4666,13 @@ def crf_decoding_layer(input,
     this layer will also calculate error. output.value[i] is 1 for incorrect
     decoding or 0 for correct decoding.
 
+    The simple usage:
+
+    .. code-block:: python
+
+      crf_decoding = crf_decoding_layer(input=input,
+                                        size=label_dim)
+
     :param input: The first input layer.
     :type input: LayerOutput
     :param size: size of this layer.

python/paddle/v2/__init__.py

@@ -18,12 +18,14 @@ import parameters
 import trainer
 import event
 import data_type
+import data_feeder
 import attr
+import pooling
 import py_paddle.swig_paddle as api
 
 __all__ = [
     'optimizer', 'layer', 'activation', 'parameters', 'init', 'trainer',
-    'event', 'data_type', 'attr'
+    'event', 'data_type', 'attr', 'pooling', 'data_feeder'
 ]
 
 

python/paddle/v2/data_feeder.py

+# 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.
+
+from py_paddle import swig_paddle
+from py_paddle import DataProviderConverter
+import data_type
+
+__all__ = ['DataFeeder']
+
+
+class DataFeeder(DataProviderConverter):
+    """
+    DataFeeder converts the data returned by paddle.reader into a data structure
+    of Arguments which is defined in the API. The paddle.reader usually returns
+    a list of mini-batch data entries. Each data entry in the list is one sampe.
+    Each sample is a list or a tuple with one feature or multiple features.
+    DataFeeder converts this mini-batch data entries into Arguments in order
+    to feed it to C++ interface.
+    
+    The example usage:
+    
+        data_types = [('image', paddle.data_type.dense_vector(784)),
+                      ('label', paddle.data_type.integer_value(10))]
+        reader_dict = {'image':0, 'label':1}
+        feeder = DataFeeder(data_types=data_types, reader_dict=reader_dict)
+        minibatch_data = [
+                           ( [1.0,2.0,3.0,4.0], 5, [6,7,8] ),  # first sample
+                           ( [1.0,2.0,3.0,4.0], 5, [6,7,8] )   # second sample
+                         ]
+        # or minibatch_data = [
+        #                       [ [1.0,2.0,3.0,4.0], 5, [6,7,8] ],  # first sample
+        #                       [ [1.0,2.0,3.0,4.0], 5, [6,7,8] ]   # second sample
+        #                     ]
+        arg = feeder(minibatch_data)
+    """
+
+    def __init__(self, data_types, reader_dict):
+        """
+        :param data_types: A list to specify data name and type. Each item is
+                           a tuple of (data_name, data_type). For example:
+                           [('image', paddle.data_type.dense_vector(784)),
+                            ('label', paddle.data_type.integer_value(10))]
+
+        :type data_types: A list of tuple
+        :param reader_dict: A dictionary to specify the position of each data
+                            in the input data.
+        :type reader_dict: dict()
+        """
+        self.input_names = []
+        input_types = []
+        self.reader_dict = reader_dict
+        for each in data_types:
+            self.input_names.append(each[0])
+            assert isinstance(each[1], data_type.InputType)
+            input_types.append(each[1])
+        DataProviderConverter.__init__(self, input_types)
+
+    def convert(self, dat, argument=None):
+        """
+        :param dat: A list of mini-batch data. Each sample is a list or tuple
+                    one feature or multiple features.
+                    for example:
+                    [ 
+                      ([0.2, 0.2], ), # first sample
+                      ([0.8, 0.3], ), # second sample
+                    ]
+                    or,
+                    [ 
+                      [[0.2, 0.2], ], # first sample
+                      [[0.8, 0.3], ], # second sample
+                    ]
+
+        :type dat: List
+        :param argument: An Arguments object contains this mini-batch data with
+                         one or multiple features. The Arguments definition is
+                         in the API.
+        :type argument: swig_paddle.Arguments
+        """
+
+        def reorder_data(data):
+            retv = []
+            for each in data:
+                reorder = []
+                for name in self.input_names:
+                    reorder.append(each[self.reader_dict[name]])
+                retv.append(reorder)
+            return retv
+
+        return DataProviderConverter.convert(self, reorder_data(dat), argument)

python/paddle/v2/data_type.py

@@ -14,9 +14,9 @@
 
 from paddle.trainer.PyDataProvider2 import \
     InputType, dense_vector, sparse_binary_vector,\
-    sparse_vector, integer_value
+    sparse_vector, integer_value, integer_value_sequence
 
 __all__ = [
     'InputType', 'dense_vector', 'sparse_binary_vector', 'sparse_vector',
-    'integer_value'
+    'integer_value', 'integer_value_sequence'
 ]

python/paddle/v2/dataset/config.py

+import os
+
+__all__ = ['DATA_HOME']
+
+DATA_HOME = os.path.expanduser('~/.cache/paddle_data_set')
+
+if not os.path.exists(DATA_HOME):
+    os.makedirs(DATA_HOME)

python/paddle/v2/dataset/mnist.py

+import sklearn.datasets.mldata
+import sklearn.model_selection
+import numpy
+from config import DATA_HOME
+
+__all__ = ['train_creator', 'test_creator']
+
+
+def __mnist_reader_creator__(data, target):
+    def reader():
+        n_samples = data.shape[0]
+        for i in xrange(n_samples):
+            yield (data[i] / 255.0).astype(numpy.float32), int(target[i])
+
+    return reader
+
+
+TEST_SIZE = 10000
+
+data = sklearn.datasets.mldata.fetch_mldata(
+    "MNIST original", data_home=DATA_HOME)
+X_train, X_test, y_train, y_test = sklearn.model_selection.train_test_split(
+    data.data, data.target, test_size=TEST_SIZE, random_state=0)
+
+
+def train_creator():
+    return __mnist_reader_creator__(X_train, y_train)
+
+
+def test_creator():
+    return __mnist_reader_creator__(X_test, y_test)
+
+
+def unittest():
+    assert len(list(test_creator()())) == TEST_SIZE
+
+
+if __name__ == '__main__':
+    unittest()

python/paddle/v2/event.py

@@ -3,8 +3,6 @@ All training events.
 
 There are:
 
-* BeginTraining
-* EndTraining
 * BeginIteration
 * EndIteration
 * BeginPass
@@ -12,15 +10,62 @@ There are:
 
 TODO(yuyang18): Complete it!
 """
-__all__ = ['EndIteration']
+import py_paddle.swig_paddle as api
+__all__ = ['EndIteration', 'BeginIteration', 'BeginPass', 'EndPass']
 
 
-class EndIteration(object):
+class WithMetric(object):
+    def __init__(self, evaluator):
+        if not isinstance(evaluator, api.Evaluator):
+            raise TypeError("Evaluator should be api.Evaluator type")
+        self.__evaluator__ = evaluator
+
+    @property
+    def metrics(self):
+        names = self.__evaluator__.getNames()
+        retv = dict()
+        for each_name in names:
+            val = self.__evaluator__.getValue(each_name)
+            retv[each_name] = val
+        return retv
+
+
+class BeginPass(object):
+    """
+    Event On One Pass Training Start.
+    """
+
+    def __init__(self, pass_id):
+        self.pass_id = pass_id
+
+
+class EndPass(WithMetric):
+    """
+    Event On One Pass Training Complete.
+    """
+
+    def __init__(self, pass_id, evaluator):
+        self.pass_id = pass_id
+        WithMetric.__init__(self, evaluator)
+
+
+class BeginIteration(object):
+    """
+    Event On One Batch Training Start.
+    """
+
+    def __init__(self, pass_id, batch_id):
+        self.pass_id = pass_id
+        self.batch_id = batch_id
+
+
+class EndIteration(WithMetric):
     """
     Event On One Batch Training Complete.
     """
 
-    def __init__(self, pass_id, batch_id, cost):
+    def __init__(self, pass_id, batch_id, cost, evaluator):
         self.pass_id = pass_id
         self.batch_id = batch_id
         self.cost = cost
+        WithMetric.__init__(self, evaluator)

python/paddle/v2/layer.py

@@ -71,17 +71,35 @@ import collections
 import paddle.trainer_config_helpers as conf_helps
 from paddle.trainer_config_helpers.config_parser_utils import \
     parse_network_config as __parse__
+from paddle.trainer_config_helpers.default_decorators import wrap_act_default
+from paddle.trainer_config_helpers.default_decorators import wrap_bias_attr_default
 from paddle.trainer_config_helpers.default_decorators import wrap_name_default
+from paddle.trainer_config_helpers.layers import layer_support
 
+import activation
 import data_type
 
 __all__ = [
-    'parse_network', 'data', 'fc', 'max_id', 'classification_cost',
-    'cross_entropy_cost', 'cross_entropy_with_selfnorm_cost', 'regression_cost',
+    'parse_network', 'data', 'fc', 'conv_shift', 'img_conv', 'img_pool', 'spp',
+    'maxout', 'img_cmrnorm', 'batch_norm', 'sum_to_one_norm', 'recurrent',
+    'lstmemory', 'grumemory', 'pool', 'last_seq', 'first_seq', 'concat',
+    'seq_concat', 'block_expand', 'expand', 'repeat', 'seq_reshape', 'addto',
+    'linear_comb', 'interpolation', 'bilinear_interp', 'power', 'scaling',
+    'slope_intercept', 'tensor', 'cos_sim', 'trans', 'max_id', 'sampling_id',
+    'pad', 'classification_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', 'memory', 'embedding', 'recurrent_group'
+    'sum_cost', 'huber_cost', 'crf', 'crf_decoding', 'ctc', 'warp_ctc', 'nce',
+    'hsigmoid', 'eos', 'memory', 'embedding', 'recurrent_group'
 ]
 
+__projection_names__ = filter(lambda x: x.endswith('_projection'),
+                              dir(conf_helps))
+__all__ += __projection_names__
+
+__operator_names__ = filter(lambda x: x.endswith('_operator'), dir(conf_helps))
+__all__ += __operator_names__
+
 
 def parse_network(*outputs):
     """
@@ -99,9 +117,8 @@ def parse_network(*outputs):
 
 
 class Layer(object):
-    def __init__(self, name, parent_layers):
+    def __init__(self, name=None, parent_layers=None):
         assert isinstance(parent_layers, dict)
-        assert isinstance(name, basestring)
         self.name = name
         self.__parent_layers__ = parent_layers
 
@@ -122,13 +139,7 @@ class Layer(object):
 
         if self.name is None:
             return self.to_proto_impl(**kwargs)
-
-        # memory may have the same name with some layer
-        if isinstance(self, MemoryV2):
-            return self.to_proto_impl(**kwargs)
-
-        # store v1 API's layer_output in context with the key of it's name.
-        if self.name not in context:
+        elif self.name not in context:
             context[self.name] = self.to_proto_impl(**kwargs)
 
         return context[self.name]
@@ -137,14 +148,14 @@ class Layer(object):
         raise NotImplementedError()
 
 
-def __convert_to_v2__(method_name, name_prefix, parent_names):
-    if name_prefix is not None:
-        wrapper = wrap_name_default(name_prefix=name_prefix)
+def __convert_to_v2__(method_name, parent_names, is_default_name=True):
+    if is_default_name:
+        wrapper = wrap_name_default(name_prefix=method_name)
     else:
         wrapper = None
 
     class V2LayerImpl(Layer):
-        def __init__(self, name=None, **kwargs):
+        def __init__(self, **kwargs):
             parent_layers = dict()
             other_kwargs = dict()
             for pname in parent_names:
@@ -155,6 +166,7 @@ def __convert_to_v2__(method_name, name_prefix, parent_names):
                 if key not in parent_names:
                     other_kwargs[key] = kwargs[key]
 
+            name = kwargs.get('name', None)
             super(V2LayerImpl, self).__init__(name, parent_layers)
             self.__other_kwargs__ = other_kwargs
 
@@ -167,7 +179,7 @@ def __convert_to_v2__(method_name, name_prefix, parent_names):
                 args[each] = kwargs[each]
             for each in self.__other_kwargs__:
                 args[each] = self.__other_kwargs__[each]
-            return getattr(conf_helps, method_name)(name=self.name, **args)
+            return getattr(conf_helps, method_name)(**args)
 
     return V2LayerImpl
 
@@ -276,44 +288,174 @@ class RecurrentGroupV2(Layer):
             name=self.name, in_args_converter=in_args_converter, **args)
 
 
+class MixedLayerV2(Layer):
+    """
+    This class is use to support `with` grammar. If not, the following code
+    could convert mixed_layer simply.
+
+        mixed = __convert_to_v2__(
+            'mixed_layer', name_prefix='mixed', parent_names=['input'])
+    """
+
+    class AddToSealedMixedLayerExceptionV2(Exception):
+        pass
+
+    def __init__(self,
+                 size=0,
+                 input=None,
+                 name=None,
+                 act=None,
+                 bias_attr=None,
+                 layer_attr=None):
+        self.__method_name__ = 'mixed_layer'
+        self.finalized = False
+        self.__inputs__ = []
+        if input is not None:
+            self.__inputs__ = input
+
+        other_kwargs = dict()
+        other_kwargs['name'] = name
+        other_kwargs['size'] = size
+        other_kwargs['act'] = act
+        other_kwargs['bias_attr'] = bias_attr
+        other_kwargs['layer_attr'] = layer_attr
+
+        parent_layers = {"input": self.__inputs__}
+        super(MixedLayerV2, self).__init__(name, parent_layers)
+        self.__other_kwargs__ = other_kwargs
+
+    def __iadd__(self, other):
+        if not self.finalized:
+            self.__inputs__.append(other)
+            return self
+        else:
+            raise MixedLayerTypeV2.AddToSealedMixedLayerExceptionV2()
+
+    def __enter__(self):
+        assert len(self.__inputs__) == 0
+        return self
+
+    def __exit__(self, *args, **kwargs):
+        self.finalized = True
+
+    def to_proto_impl(self, **kwargs):
+        args = dict()
+        for each in kwargs:
+            args[each] = kwargs[each]
+        for each in self.__other_kwargs__:
+            args[each] = self.__other_kwargs__[each]
+        return getattr(conf_helps, self.__method_name__)(**args)
+
+
+@wrap_name_default("mixed")
+@wrap_act_default(act=activation.Linear())
+@wrap_bias_attr_default(has_bias=False)
+@layer_support(conf_helps.layers.ERROR_CLIPPING, conf_helps.layers.DROPOUT)
+def mixed(size=0,
+          name=None,
+          input=None,
+          act=None,
+          bias_attr=False,
+          layer_attr=None):
+    return MixedLayerV2(size, input, name, act, bias_attr, layer_attr)
+
+
 data = DataLayerV2
-fc = __convert_to_v2__('fc_layer', name_prefix='fc', parent_names=['input'])
-max_id = __convert_to_v2__(
-    'maxid_layer', name_prefix='maxid', parent_names=['input'])
-classification_cost = __convert_to_v2__(
-    'classification_cost',
-    name_prefix='classification_cost',
-    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'])
-embedding = __convert_to_v2__(
-    'embedding_layer', name_prefix='embedding', parent_names=['input'])
-last_seq = __convert_to_v2__(
-    'last_seq', name_prefix='last_seq', parent_names=['input'])
+AggregateLevel = conf_helps.layers.AggregateLevel
+ExpandLevel = conf_helps.layers.ExpandLevel
 recurrent_group = RecurrentGroupV2
 memory = MemoryV2
 
-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'])
+layer_list = [
+    # [V2LayerImpl, V1_method_name, parent_names]
+    # fully connected layers
+    ['fc', 'fc_layer', ['input']],
+    ['embedding', 'embedding_layer', ['input']],
+    # conv layers
+    ['conv_shift', 'conv_shift_layer', ['a', 'b']],
+    ['img_conv', 'img_conv_layer', ['input']],
+    # image pooling layers
+    ['img_pool', 'img_pool_layer', ['input']],
+    ['spp', 'spp_layer', ['input']],
+    ['maxout', 'maxout_layer', ['input']],
+    # norm layers
+    ['img_cmrnorm', 'img_cmrnorm_layer', ['input']],
+    ['batch_norm', 'batch_norm_layer', ['input']],
+    ['sum_to_one_norm', 'sum_to_one_norm_layer', ['input']],
+    # recurrent layers
+    ['recurrent', 'recurrent_layer', ['input']],
+    ['lstmemory', 'lstmemory', ['input']],
+    ['grumemory', 'grumemory', ['input']],
+    # aggregate layers
+    ['pool', 'pooling_layer', ['input']],
+    ['last_seq', 'last_seq', ['input']],
+    ['first_seq', 'first_seq', ['input']],
+    ['concat', 'concat_layer', ['input']],
+    ['seq_concat', 'seq_concat_layer', ['a', 'b']],
+    # reshaping layers
+    ['block_expand', 'block_expand_layer', ['input']],
+    ['expand', 'expand_layer', ['input', 'expand_as']],
+    ['repeat', 'repeat_layer', ['input']],
+    ['rotate', 'rotate_layer', ['input']],
+    ['seq_reshape', 'seq_reshape_layer', ['input']],
+    # math layers
+    ['addto', 'addto_layer', ['input']],
+    ['linear_comb', 'linear_comb_layer', ['weights', 'vectors']],
+    ['interpolation', 'interpolation_layer', ['input', 'weight']],
+    ['bilinear_interp', 'bilinear_interp_layer', ['input']],
+    ['power', 'power_layer', ['input', 'weight']],
+    ['scaling', 'scaling_layer', ['input', 'weight']],
+    ['slope_intercept', 'slope_intercept_layer', ['input']],
+    ['tensor', 'tensor_layer', ['a', 'b']],
+    ['cos_sim', 'cos_sim', ['a', 'b']],
+    ['trans', 'trans_layer', ['input']],
+    # sampling layers
+    ['max_id', 'maxid_layer', ['input']],
+    ['sampling_id', 'sampling_id_layer', ['input']],
+    # slicing and joining layers
+    ['pad', 'pad_layer', ['input']],
+    # cost layers
+    [
+        'classification_cost', 'classification_cost',
+        ['input', 'label', 'weight']
+    ],
+    ['regression_cost', 'regression_cost', ['input', 'label', 'weight']],
+    ['cross_entropy_cost', 'cross_entropy', ['input', 'label']],
+    [
+        'cross_entropy_with_selfnorm_cost', 'cross_entropy_with_selfnorm',
+        ['input', 'label']
+    ],
+    [
+        'multi_binary_label_cross_entropy_cost',
+        'multi_binary_label_cross_entropy', ['input', 'label']
+    ],
+    ['rank_cost', 'rank_cost', ['left', 'right', 'label', 'weight']],
+    ['lambda_cost', 'lambda_cost', ['input', 'score']],
+    ['sum_cost', 'sum_cost', ['input']],
+    ['huber_cost', 'huber_cost', ['input', 'label']],
+    ['crf', 'crf_layer', ['input', 'label']],
+    ['crf_decoding', 'crf_decoding_layer', ['input']],
+    ['ctc', 'ctc_layer', ['input', 'label']],
+    ['warp_ctc', 'warp_ctc_layer', ['input', 'label']],
+    ['nce', 'nce_layer', ['input', 'label']],
+    ['hsigmoid', 'hsigmoid', ['input', 'label']],
+    # check layers
+    ['eos', 'eos_layer', ['input']]
+]
+for l in layer_list:
+    globals()[l[0]] = __convert_to_v2__(l[1], l[2])
+
+# convert projection
+for prj in __projection_names__:
+    globals()[prj] = __convert_to_v2__(
+        prj, parent_names=['input'], is_default_name=False)
+
+# convert operator
+operator_list = [
+    # [V1_method_name, parent_names],
+    ['dotmul_operator', ['a', 'b']],
+    ['conv_operator', ['img', 'filter']]
+]
+for op in operator_list:
+    globals()[op[0]] = __convert_to_v2__(
+        op[0], parent_names=op[1], is_default_name=False)

python/paddle/v2/optimizer.py

@@ -3,7 +3,10 @@ import paddle.trainer_config_helpers.optimizers as v1_optimizers
 import paddle.trainer_config_helpers.config_parser_utils as config_parser_utils
 import paddle.v2
 
-__all__ = ['Adam', 'Adamax']
+__all__ = [
+    'Momentum', 'Adam', 'Adamax', 'AdaGrad', 'DecayedAdaGrad', 'AdaDelta',
+    'RMSProp', 'ModelAverage', 'L2Regularization'
+]
 
 
 class Optimizer(object):
@@ -38,6 +41,14 @@ class Optimizer(object):
                                                              pass_num)
 
 
+class Momentum(Optimizer):
+    def __init__(self, momentum=None, sparse=False, **kwargs):
+        learning_method = v1_optimizers.MomentumOptimizer(
+            momentum=None, sparse=False)
+        super(Momentum, self).__init__(
+            learning_method=learning_method, **kwargs)
+
+
 class Adam(Optimizer):
     def __init__(self, beta1=0.9, beta2=0.999, epsilon=1e-8, **kwargs):
         learning_method = v1_optimizers.AdamOptimizer(
@@ -52,7 +63,45 @@ class Adamax(Optimizer):
         super(Adamax, self).__init__(learning_method=learning_method, **kwargs)
 
 
+class AdaGrad(Optimizer):
+    def __init__(self, **kwargs):
+        learning_method = v1_optimizers.AdaGradOptimizer()
+        super(AdaGrad, self).__init__(learning_method=learning_method, **kwargs)
+
+
+class DecayedAdaGrad(Optimizer):
+    def __init__(self, rho=0.95, epsilon=1e-06, **kwargs):
+        learning_method = v1_optimizers.DecayedAdaGradOptimizer(
+            rho=rho, epsilon=epsilon)
+        super(DecayedAdaGrad, self).__init__(
+            learning_method=learning_method, **kwargs)
+
+
+class AdaDelta(Optimizer):
+    def __init__(self, rho=0.95, epsilon=1e-06, **kwargs):
+        learning_method = v1_optimizers.AdaDeltaOptimizer(
+            rho=rho, epsilon=epsilon)
+        super(AdaDelta, self).__init__(
+            learning_method=learning_method, **kwargs)
+
+
+class RMSProp(Optimizer):
+    def __init__(self, rho=0.95, epsilon=1e-6, **kwargs):
+        learning_method = v1_optimizers.RMSPropOptimizer(
+            rho=rho, epsilon=epsilon)
+        super(RMSProp, self).__init__(learning_method=learning_method, **kwargs)
+
+
+ModelAverage = v1_optimizers.ModelAverage
+L2Regularization = v1_optimizers.L2Regularization
+
 if __name__ == '__main__':
     swig_api.initPaddle('--use_gpu=false')
-    opt = paddle.v2.optimizer.Adam()
-    print opt.enable_types()
+    for opt in [
+            Momentum(), Adam(), Adamax(), AdaGrad(), DecayedAdaGrad(),
+            AdaDelta(), RMSProp(), Adam(
+                model_average=ModelAverage(average_window=0.5),
+                regularization=L2Regularization(rate=0.5),
+                gradient_clipping_threshold=25)
+    ]:
+        print opt, opt.enable_types()

python/paddle/v2/pooling.py

+# 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.
+
+from paddle.trainer_config_helpers.poolings import *
+
+__all__ = ["Max", "CudnnMax", "Avg", "CudnnAvg", "Sum", "SquareRootN"]
+
+Max = MaxPooling
+CudnnMax = CudnnMaxPooling
+Avg = AvgPooling
+CudnnAvg = CudnnAvgPooling
+Sum = SumPooling
+SquareRootN = SquareRootNPooling

python/paddle/v2/tests/run_tests.sh

+#!/bin/bash
+# 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.
+
+
+pushd `dirname $0` > /dev/null
+SCRIPTPATH=$PWD
+popd > /dev/null
+
+cd $SCRIPTPATH
+
+$1 -m pip install ../../../../paddle/dist/*.whl
+
+test_list="test_data_feeder.py"
+
+export PYTHONPATH=$PWD/../../../../python/
+
+for fn in $test_list
+do
+  echo "test $fn"
+  $1 $fn
+  if [ $? -ne 0 ]; then
+    exit 1
+  fi
+done

python/paddle/v2/tests/test_data_feeder.py

+# 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.
+
+import unittest
+
+import py_paddle.swig_paddle as api
+import numpy as np
+
+from paddle.v2 import data_type
+from paddle.v2.data_feeder import DataFeeder
+
+
+class DataFeederTest(unittest.TestCase):
+    def dense_reader(self, size):
+        data = np.random.random(size)
+        return data
+
+    def sparse_binary_reader(self, high, size_limit, non_empty=False):
+        num = np.random.randint(size_limit)  # num could be 0
+        while non_empty and num == 0:
+            num = np.random.randint(size_limit)
+        return np.random.randint(high, size=num).tolist()
+
+    def test_dense(self):
+        def compare(input):
+            feeder = DataFeeder([('image', data_type.dense_vector(784))],
+                                {'image': 0})
+            arg = feeder(input)
+            output = arg.getSlotValue(0).copyToNumpyMat()
+            input = np.array(input, dtype='float32')
+            self.assertAlmostEqual(input.all(), output.all())
+
+        # test numpy array
+        batch_size = 32
+        dim = 784
+        data = []
+        for i in xrange(batch_size):
+            each_sample = []
+            each_sample.append(self.dense_reader(dim))
+            data.append(each_sample)
+        compare(data)
+
+        # each feature is a list
+        data = []
+        for i in xrange(batch_size):
+            each_sample = []
+            each_sample.append(self.dense_reader(dim).tolist())
+            data.append(each_sample)
+        compare(data)
+
+        # test tuple
+        data = []
+        for i in xrange(batch_size):
+            each_sample = (self.dense_reader(dim).tolist(), )
+            data.append(each_sample)
+        compare(data)
+
+    def test_sparse_binary(self):
+        dim = 10000
+        batch_size = 32
+        data = []
+        for i in xrange(batch_size):
+            each_sample = []
+            each_sample.append(self.sparse_binary_reader(dim, 50))
+            data.append(each_sample)
+        feeder = DataFeeder([('input', data_type.sparse_binary_vector(dim))],
+                            {'input': 0})
+        arg = feeder(data)
+        output = arg.getSlotValue(0)
+        assert isinstance(output, api.Matrix)
+        for i in xrange(batch_size):
+            self.assertEqual(output.getSparseRowCols(i), data[i][0])
+
+    def test_sparse(self):
+        dim = 10000
+        batch_size = 32
+        v = []
+        w = []
+        data = []
+        for dat in xrange(batch_size):
+            each_sample = []
+            a = self.sparse_binary_reader(dim, 40, non_empty=True)
+            b = self.dense_reader(len(a)).tolist()
+            v.append(a)
+            w.append(np.array(b, dtype="float32"))
+            each_sample.append(zip(a, b))
+            data.append(each_sample)
+
+        feeder = DataFeeder([('input', data_type.sparse_vector(dim))],
+                            {'input': 0})
+        arg = feeder(data)
+        output = arg.getSlotValue(0)
+        assert isinstance(output, api.Matrix)
+        for i in xrange(batch_size):
+            self.assertEqual(output.getSparseRowCols(i), v[i])
+            cols_value = output.getSparseRowColsVal(i)
+            value = [val[1] for val in cols_value]
+            value = np.array(value, dtype="float32")
+            self.assertAlmostEqual(value.all(), w[i].all())
+
+    def test_integer(self):
+        dim = 100
+        batch_size = 32
+        index = []
+        for i in xrange(batch_size):
+            each_sample = []
+            each_sample.append(np.random.randint(dim))
+            index.append(each_sample)
+        feeder = DataFeeder([('input', data_type.integer_value(dim))],
+                            {'input': 0})
+        arg = feeder(index)
+        output = arg.getSlotIds(0).copyToNumpyArray()
+        index = np.array(index, dtype='int')
+        self.assertEqual(output.all(), index.flatten().all())
+
+    def test_integer_sequence(self):
+        dim = 10000
+        batch_size = 32
+        start = [0]
+        data = []
+        for i in xrange(batch_size):
+            each_sample = []
+            each_sample.append(
+                self.sparse_binary_reader(
+                    dim, 30, non_empty=True))
+            data.append(each_sample)
+            start.append(len(each_sample[0]) + start[-1])
+        feeder = DataFeeder([('input', data_type.integer_value_sequence(dim))],
+                            {'input': 0})
+        arg = feeder(data)
+        output_data = arg.getSlotIds(0).copyToNumpyArray()
+        output_start = arg.getSlotSequenceStartPositions(0).copyToNumpyArray()
+
+        index = []
+        for dat in data:
+            index.extend(x for x in dat[0])  # only one feature, so dat[0]
+        index = np.array(index, dtype='int')
+        start = np.array(start, dtype='int')
+        self.assertEqual(output_data.all(), index.all())
+        self.assertEqual(output_start.all(), start.all())
+
+    def test_multiple_features(self):
+        batch_size = 2
+        data = []
+        for i in xrange(batch_size):
+            each_sample = []
+            each_sample.append(np.random.randint(10))
+            each_sample.append(
+                self.sparse_binary_reader(
+                    20000, 40, non_empty=True))
+            each_sample.append(self.dense_reader(100))
+            data.append(each_sample)
+
+        # test multiple features
+        data_types = [('fea0', data_type.dense_vector(100)),
+                      ('fea1', data_type.sparse_binary_vector(20000)),
+                      ('fea2', data_type.integer_value(10))]
+        feeder = DataFeeder(data_types, {'fea0': 2, 'fea1': 1, 'fea2': 0})
+        arg = feeder(data)
+        output_dense = arg.getSlotValue(0).copyToNumpyMat()
+        output_sparse = arg.getSlotValue(1)
+        output_index = arg.getSlotIds(2).copyToNumpyArray()
+        for i in xrange(batch_size):
+            self.assertEqual(output_dense[i].all(), data[i][2].all())
+            self.assertEqual(output_sparse.getSparseRowCols(i), data[i][1])
+            self.assertEqual(output_index[i], data[i][0])
+
+        # reader returns 3 features, but only use 2 features
+        data_types = [('fea0', data_type.dense_vector(100)),
+                      ('fea2', data_type.integer_value(10))]
+        feeder = DataFeeder(data_types, {'fea0': 2, 'fea2': 0})
+        arg = feeder(data)
+        output_dense = arg.getSlotValue(0).copyToNumpyMat()
+        output_index = arg.getSlotIds(1).copyToNumpyArray()
+        for i in xrange(batch_size):
+            self.assertEqual(output_dense[i].all(), data[i][2].all())
+            self.assertEqual(output_index[i], data[i][0])
+
+        # reader returns 3 featreus, one is duplicate data
+        data_types = [('fea0', data_type.dense_vector(100)),
+                      ('fea1', data_type.sparse_binary_vector(20000)),
+                      ('fea2', data_type.integer_value(10)),
+                      ('fea3', data_type.dense_vector(100))]
+        feeder = DataFeeder(data_types,
+                            {'fea0': 2,
+                             'fea1': 1,
+                             'fea2': 0,
+                             'fea3': 2})
+        arg = feeder(data)
+        fea0 = arg.getSlotValue(0).copyToNumpyMat()
+        fea1 = arg.getSlotValue(1)
+        fea2 = arg.getSlotIds(2).copyToNumpyArray()
+        fea3 = arg.getSlotValue(3).copyToNumpyMat()
+        for i in xrange(batch_size):
+            self.assertEqual(fea0[i].all(), data[i][2].all())
+            self.assertEqual(fea1.getSparseRowCols(i), data[i][1])
+            self.assertEqual(fea2[i], data[i][0])
+            self.assertEqual(fea3[i].all(), data[i][2].all())
+
+    def test_multiple_features_tuple(self):
+        batch_size = 2
+        data = []
+        for i in xrange(batch_size):
+            a = np.random.randint(10)
+            b = self.sparse_binary_reader(20000, 40, non_empty=True)
+            c = self.dense_reader(100)
+            each_sample = (a, b, c)
+            data.append(each_sample)
+
+        # test multiple features
+        data_types = [('fea0', data_type.dense_vector(100)),
+                      ('fea1', data_type.sparse_binary_vector(20000)),
+                      ('fea2', data_type.integer_value(10))]
+        feeder = DataFeeder(data_types, {'fea0': 2, 'fea1': 1, 'fea2': 0})
+        arg = feeder(data)
+        out_dense = arg.getSlotValue(0).copyToNumpyMat()
+        out_sparse = arg.getSlotValue(1)
+        out_index = arg.getSlotIds(2).copyToNumpyArray()
+        for i in xrange(batch_size):
+            self.assertEqual(out_dense[i].all(), data[i][2].all())
+            self.assertEqual(out_sparse.getSparseRowCols(i), data[i][1])
+            self.assertEqual(out_index[i], data[i][0])
+
+
+if __name__ == '__main__':
+    api.initPaddle("--use_gpu=0")
+    unittest.main()

python/paddle/v2/tests/test_layer.py

@@ -17,16 +17,104 @@ import paddle.v2.activation as activation
 import paddle.v2.attr as attr
 import paddle.v2.data_type as data_type
 import paddle.v2.layer as layer
+import paddle.v2.pooling as pooling
 
-pixel = layer.data(name='pixel', type=data_type.dense_vector(784))
+pixel = layer.data(name='pixel', type=data_type.dense_vector(128))
 label = layer.data(name='label', type=data_type.integer_value(10))
 weight = layer.data(name='weight', type=data_type.dense_vector(10))
 score = layer.data(name='score', type=data_type.dense_vector(1))
+
 hidden = layer.fc(input=pixel,
                   size=100,
                   act=activation.Sigmoid(),
                   param_attr=attr.Param(name='hidden'))
 inference = layer.fc(input=hidden, size=10, act=activation.Softmax())
+conv = layer.img_conv(
+    input=pixel,
+    filter_size=1,
+    filter_size_y=1,
+    num_channels=8,
+    num_filters=16,
+    act=activation.Linear())
+
+
+class ImageLayerTest(unittest.TestCase):
+    def test_conv_layer(self):
+        conv_shift = layer.conv_shift(a=pixel, b=score)
+        print layer.parse_network(conv, conv_shift)
+
+    def test_pooling_layer(self):
+        maxpool = layer.img_pool(
+            input=conv,
+            pool_size=2,
+            num_channels=16,
+            padding=1,
+            pool_type=pooling.Max())
+        spp = layer.spp(input=conv,
+                        pyramid_height=2,
+                        num_channels=16,
+                        pool_type=pooling.Max())
+        maxout = layer.maxout(input=conv, num_channels=16, groups=4)
+        print layer.parse_network(maxpool, spp, maxout)
+
+    def test_norm_layer(self):
+        norm1 = layer.img_cmrnorm(input=conv, size=5)
+        norm2 = layer.batch_norm(input=conv)
+        norm3 = layer.sum_to_one_norm(input=conv)
+        print layer.parse_network(norm1, norm2, norm3)
+
+
+class AggregateLayerTest(unittest.TestCase):
+    def test_aggregate_layer(self):
+        pool = layer.pool(
+            input=pixel,
+            pooling_type=pooling.Avg(),
+            agg_level=layer.AggregateLevel.EACH_SEQUENCE)
+        last_seq = layer.last_seq(input=pixel)
+        first_seq = layer.first_seq(input=pixel)
+        concat = layer.concat(input=[last_seq, first_seq])
+        seq_concat = layer.seq_concat(a=last_seq, b=first_seq)
+        print layer.parse_network(pool, last_seq, first_seq, concat, seq_concat)
+
+
+class MathLayerTest(unittest.TestCase):
+    def test_math_layer(self):
+        addto = layer.addto(input=[pixel, pixel])
+        linear_comb = layer.linear_comb(weights=weight, vectors=hidden, size=10)
+        interpolation = layer.interpolation(
+            input=[hidden, hidden], weight=score)
+        bilinear = layer.bilinear_interp(input=conv, out_size_x=4, out_size_y=4)
+        power = layer.power(input=pixel, weight=score)
+        scaling = layer.scaling(input=pixel, weight=score)
+        slope = layer.slope_intercept(input=pixel)
+        tensor = layer.tensor(a=pixel, b=pixel, size=1000)
+        cos_sim = layer.cos_sim(a=pixel, b=pixel)
+        trans = layer.trans(input=tensor)
+        print layer.parse_network(addto, linear_comb, interpolation, power,
+                                  scaling, slope, tensor, cos_sim, trans)
+
+
+class ReshapeLayerTest(unittest.TestCase):
+    def test_reshape_layer(self):
+        block_expand = layer.block_expand(
+            input=conv, num_channels=4, stride_x=1, block_x=1)
+        expand = layer.expand(
+            input=weight,
+            expand_as=pixel,
+            expand_level=layer.ExpandLevel.FROM_TIMESTEP)
+        repeat = layer.repeat(input=pixel, num_repeats=4)
+        reshape = layer.seq_reshape(input=pixel, reshape_size=4)
+        rotate = layer.rotate(input=pixel, height=16, width=49)
+        print layer.parse_network(block_expand, expand, repeat, reshape, rotate)
+
+
+class RecurrentLayerTest(unittest.TestCase):
+    def test_recurrent_layer(self):
+        word = layer.data(name='word', type=data_type.integer_value(12))
+        recurrent = layer.recurrent(input=word)
+        lstm = layer.lstmemory(input=word)
+        gru = layer.grumemory(input=word)
+        print layer.parse_network(recurrent, lstm, gru)
 
 
 class CostLayerTest(unittest.TestCase):
@@ -52,6 +140,116 @@ class CostLayerTest(unittest.TestCase):
         print layer.parse_network(cost5, cost6)
         print layer.parse_network(cost7, cost8, cost9, cost10, cost11)
 
+        crf = layer.crf(input=inference, label=label)
+        crf_decoding = layer.crf_decoding(input=inference, size=3)
+        ctc = layer.ctc(input=inference, label=label)
+        warp_ctc = layer.warp_ctc(input=pixel, label=label)
+        nce = layer.nce(input=inference, label=label, num_classes=3)
+        hsigmoid = layer.hsigmoid(input=inference, label=label, num_classes=3)
+
+        print layer.parse_network(crf, crf_decoding, ctc, warp_ctc, nce,
+                                  hsigmoid)
+
+
+class OtherLayerTest(unittest.TestCase):
+    def test_sampling_layer(self):
+        maxid = layer.max_id(input=inference)
+        sampling_id = layer.sampling_id(input=inference)
+        eos = layer.eos(input=maxid, eos_id=5)
+        print layer.parse_network(maxid, sampling_id, eos)
+
+    def test_slicing_joining_layer(self):
+        pad = layer.pad(input=conv, pad_c=[2, 3], pad_h=[1, 2], pad_w=[3, 1])
+        print layer.parse_network(pad)
+
+
+class ProjOpTest(unittest.TestCase):
+    def test_projection(self):
+        input = layer.data(name='data', type=data_type.dense_vector(784))
+        word = layer.data(
+            name='word', type=data_type.integer_value_sequence(10000))
+        fc0 = layer.fc(input=input, size=100, act=activation.Sigmoid())
+        fc1 = layer.fc(input=input, size=200, act=activation.Sigmoid())
+        mixed0 = layer.mixed(
+            size=256,
+            input=[
+                layer.full_matrix_projection(input=fc0),
+                layer.full_matrix_projection(input=fc1)
+            ])
+        with layer.mixed(size=200) as mixed1:
+            mixed1 += layer.full_matrix_projection(input=fc0)
+            mixed1 += layer.identity_projection(input=fc1)
+
+        table = layer.table_projection(input=word)
+        emb0 = layer.mixed(size=512, input=table)
+        with layer.mixed(size=512) as emb1:
+            emb1 += table
+
+        scale = layer.scaling_projection(input=fc0)
+        scale0 = layer.mixed(size=100, input=scale)
+        with layer.mixed(size=100) as scale1:
+            scale1 += scale
+
+        dotmul = layer.dotmul_projection(input=fc0)
+        dotmul0 = layer.mixed(size=100, input=dotmul)
+        with layer.mixed(size=100) as dotmul1:
+            dotmul1 += dotmul
+
+        context = layer.context_projection(input=fc0, context_len=5)
+        context0 = layer.mixed(size=100, input=context)
+        with layer.mixed(size=100) as context1:
+            context1 += context
+
+        conv = layer.conv_projection(
+            input=input,
+            filter_size=1,
+            num_channels=1,
+            num_filters=128,
+            stride=1,
+            padding=0)
+        conv0 = layer.mixed(input=conv, bias_attr=True)
+        with layer.mixed(bias_attr=True) as conv1:
+            conv1 += conv
+
+        print layer.parse_network(mixed0)
+        print layer.parse_network(mixed1)
+        print layer.parse_network(emb0)
+        print layer.parse_network(emb1)
+        print layer.parse_network(scale0)
+        print layer.parse_network(scale1)
+        print layer.parse_network(dotmul0)
+        print layer.parse_network(dotmul1)
+        print layer.parse_network(conv0)
+        print layer.parse_network(conv1)
+
+    def test_operator(self):
+        ipt0 = layer.data(name='data', type=data_type.dense_vector(784))
+        ipt1 = layer.data(name='word', type=data_type.dense_vector(128))
+        fc0 = layer.fc(input=ipt0, size=100, act=activation.Sigmoid())
+        fc1 = layer.fc(input=ipt0, size=100, act=activation.Sigmoid())
+
+        dotmul_op = layer.dotmul_operator(a=fc0, b=fc1)
+        dotmul0 = layer.mixed(input=dotmul_op)
+        with layer.mixed() as dotmul1:
+            dotmul1 += dotmul_op
+
+        conv = layer.conv_operator(
+            img=ipt0,
+            filter=ipt1,
+            filter_size=1,
+            num_channels=1,
+            num_filters=128,
+            stride=1,
+            padding=0)
+        conv0 = layer.mixed(input=conv)
+        with layer.mixed() as conv1:
+            conv1 += conv
+
+        print layer.parse_network(dotmul0)
+        print layer.parse_network(dotmul1)
+        print layer.parse_network(conv0)
+        print layer.parse_network(conv1)
+
 
 if __name__ == '__main__':
     unittest.main()

python/paddle/v2/trainer.py

@@ -2,7 +2,7 @@ import collections
 
 import py_paddle.swig_paddle as api
 from paddle.proto.ModelConfig_pb2 import ModelConfig
-from py_paddle import DataProviderConverter
+from data_feeder import DataFeeder
 
 from . import event as v2_event
 from . import layer as v2_layer
@@ -69,7 +69,8 @@ class SGD(ITrainer):
               test_data_reader=None,
               event_handler=None,
               batch_size=32,
-              data_types=None):
+              data_types=None,
+              reader_dict=None):
         """
         Training method. Will train num_passes of input data.
 
@@ -96,28 +97,34 @@ class SGD(ITrainer):
             topology, api.CREATE_MODE_NORMAL, self.__optimizer__.enable_types())
         assert isinstance(gm, api.GradientMachine)
         parameters.append_gradient_machine(gm)
-
+        gm.randParameters()
         updater = self.__optimizer__.create_local_updater()
         updater.init(gm)
 
         gm.start()
+        batch_evaluator = gm.makeEvaluator()
+        assert isinstance(batch_evaluator, api.Evaluator)
+        pass_evaluator = gm.makeEvaluator()
+        assert isinstance(pass_evaluator, api.Evaluator)
         out_args = api.Arguments.createArguments(0)
 
-        data_types_lists = []
-        for each in topology.input_layer_names:
-            if each not in data_types:
-                raise ValueError()
-            data_types_lists.append(data_types[each])
-
-        converter = DataProviderConverter(input_types=data_types_lists)
+        feeder = DataFeeder(data_types, reader_dict)
 
         for pass_id in xrange(num_passes):
+            event_handler(v2_event.BeginPass(pass_id))
+            pass_evaluator.start()
             updater.startPass()
             for batch_id, data_batch in enumerate(
                     __data_reader_to_batch__(train_data_reader, batch_size,
                                              topology)):
+                batch_evaluator.start()
+                event_handler(
+                    v2_event.BeginIteration(
+                        pass_id=pass_id, batch_id=batch_id))
                 pass_type = updater.startBatch(len(data_batch))
-                gm.forwardBackward(converter(data_batch), out_args, pass_type)
+                gm.forwardBackward(feeder(data_batch), out_args, pass_type)
+                gm.eval(pass_evaluator)
+                gm.eval(batch_evaluator)
                 for each_param in gm.getParameters():
                     updater.update(each_param)
                 # Get cost. We use numpy to calculate total cost for this batch.
@@ -125,11 +132,17 @@ class SGD(ITrainer):
                 cost_vec = cost_vec.copyToNumpyMat()
                 cost = cost_vec.sum() / len(data_batch)
                 updater.finishBatch(cost)
+                batch_evaluator.finish()
                 event_handler(
                     v2_event.EndIteration(
-                        pass_id=pass_id, batch_id=batch_id, cost=cost))
+                        pass_id=pass_id,
+                        batch_id=batch_id,
+                        cost=cost,
+                        evaluator=batch_evaluator))
 
             updater.finishPass()
+            pass_evaluator.finish()
+            event_handler(v2_event.EndPass(pass_id, evaluator=pass_evaluator))
         gm.finish()