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

.travis.yml

@@ -4,6 +4,7 @@ cache:
     - $HOME/third_party
     - $HOME/.ccache
     - $HOME/.cache/pip
+    - $HOME/Library/Caches/Homebrew
 sudo: required
 dist: trusty
 os:
@@ -54,7 +55,9 @@ before_install:
     fi
   - if [[ "$TRAVIS_OS_NAME" == "osx" ]]; then paddle/scripts/travis/before_install.osx.sh; fi
   - if [[ "$JOB" == "PRE_COMMIT" ]]; then sudo ln -s /usr/bin/clang-format-3.8 /usr/bin/clang-format; fi
-  - pip install numpy wheel protobuf sphinx recommonmark sphinx_rtd_theme virtualenv pre-commit requests==2.9.2 LinkChecker
+  # Paddle is using protobuf 3.1 currently. Protobuf 3.2 breaks the compatibility. So we specify the python 
+  # protobuf version.
+  - pip install numpy wheel 'protobuf==3.1' sphinx recommonmark sphinx_rtd_theme virtualenv pre-commit requests==2.9.2 LinkChecker
 script:
   - paddle/scripts/travis/main.sh
 notifications:

cmake/cblas.cmake

@@ -16,7 +16,8 @@
 set(CBLAS_FOUND OFF)
 
 ## Find MKL First.
-set(MKL_ROOT $ENV{MKLROOT} CACHE PATH "Folder contains MKL")
+set(INTEL_ROOT "/opt/intel" CACHE PATH "Folder contains intel libs")
+set(MKL_ROOT ${INTEL_ROOT}/mkl CACHE PATH "Folder contains MKL")
 
 find_path(MKL_INCLUDE_DIR mkl.h PATHS
   ${MKL_ROOT}/include)

demo/mnist/api_train.py

@@ -6,25 +6,15 @@ passed to C++ side of Paddle.
 
 The user api could be simpler and carefully designed.
 """
-import py_paddle.swig_paddle as api
-from py_paddle import DataProviderConverter
-import paddle.trainer.PyDataProvider2 as dp
-import numpy as np
 import random
-from mnist_util import read_from_mnist
-from paddle.trainer_config_helpers import *
-import paddle.v2
 
+import numpy as np
+import paddle.v2 as paddle_v2
+import py_paddle.swig_paddle as api
+from paddle.trainer_config_helpers import *
+from py_paddle import DataProviderConverter
 
-def network_config():
-    imgs = data_layer(name='pixel', size=784)
-    hidden1 = fc_layer(input=imgs, size=200)
-    hidden2 = fc_layer(input=hidden1, size=200)
-    inference = fc_layer(input=hidden2, size=10, act=SoftmaxActivation())
-    cost = classification_cost(
-        input=inference, label=data_layer(
-            name='label', size=10))
-    outputs(cost)
+from mnist_util import read_from_mnist
 
 
 def init_parameter(network):
@@ -67,7 +57,7 @@ def input_order_converter(generator):
 def main():
     api.initPaddle("-use_gpu=false", "-trainer_count=4")  # use 4 cpu cores
 
-    optimizer = paddle.v2.optimizer.Adam(
+    optimizer = paddle_v2.optimizer.Adam(
         learning_rate=1e-4,
         batch_size=1000,
         model_average=ModelAverage(average_window=0.5),
@@ -79,8 +69,20 @@ def main():
     updater = optimizer.create_local_updater()
     assert isinstance(updater, api.ParameterUpdater)
 
+    # define network
+    images = paddle_v2.layer.data(
+        name='pixel', type=paddle_v2.data_type.dense_vector(784))
+    label = paddle_v2.layer.data(
+        name='label', type=paddle_v2.data_type.integer_value(10))
+    hidden1 = paddle_v2.layer.fc(input=images, size=200)
+    hidden2 = paddle_v2.layer.fc(input=hidden1, size=200)
+    inference = paddle_v2.layer.fc(input=hidden2,
+                                   size=10,
+                                   act=paddle_v2.activation.Softmax())
+    cost = paddle_v2.layer.classification_cost(input=inference, label=label)
+
     # Create Simple Gradient Machine.
-    model_config = parse_network_config(network_config)
+    model_config = paddle_v2.layer.parse_network(cost)
     m = api.GradientMachine.createFromConfigProto(model_config,
                                                   api.CREATE_MODE_NORMAL,
                                                   optimizer.enable_types())
@@ -97,8 +99,7 @@ def main():
 
     # DataProvider Converter is a utility convert Python Object to Paddle C++
     # Input. The input format is as same as Paddle's DataProvider.
-    converter = DataProviderConverter(
-        input_types=[dp.dense_vector(784), dp.integer_value(10)])
+    converter = DataProviderConverter(input_types=[images.type, label.type])
 
     train_file = './data/raw_data/train'
     test_file = './data/raw_data/t10k'

demo/mnist/api_train_v2.py

+import numpy
+import paddle.v2 as paddle
+
+import mnist_util
+
+
+def train_reader():
+    train_file = './data/raw_data/train'
+    generator = mnist_util.read_from_mnist(train_file)
+    for item in generator:
+        yield item
+
+
+def main():
+    paddle.init(use_gpu=False, trainer_count=1)
+
+    # define network topology
+    images = paddle.layer.data(
+        name='pixel', type=paddle.data_type.dense_vector(784))
+    label = paddle.layer.data(
+        name='label', type=paddle.data_type.integer_value(10))
+    hidden1 = paddle.layer.fc(input=images, size=200)
+    hidden2 = paddle.layer.fc(input=hidden1, size=200)
+    inference = paddle.layer.fc(input=hidden2,
+                                size=10,
+                                act=paddle.activation.Softmax())
+    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())
+
+        else:
+            pass
+
+    trainer = paddle.trainer.SGD(update_equation=adam_optimizer)
+
+    trainer.train(train_data_reader=train_reader,
+                  topology=cost,
+                  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
+                      # network topology
+                      'pixel': images.type,
+                      'label': label.type
+                  })
+
+
+if __name__ == '__main__':
+    main()

demo/sentiment/dataprovider.py

@@ -32,4 +32,6 @@ def process(settings, file_name):
             word_slot = [
                 settings.word_dict[w] for w in words if w in settings.word_dict
             ]
+            if not word_slot:
+                continue
             yield word_slot, label

demo/sentiment/predict.py

@@ -138,7 +138,11 @@ def main():
 
     batch = []
     for line in sys.stdin:
-        batch.append([predict.get_index(line)])
+        words = predict.get_index(line)
+        if words:
+            batch.append([words])
+        else:
+            print('All the words in [%s] are not in the dictionary.' % line)
         if len(batch) == batch_size:
             predict.batch_predict(batch)
             batch = []

doc/api/trainer_config_helpers/layers.rst

@@ -279,6 +279,12 @@ concat_layer
     :members: concat_layer
     :noindex:
 
+seq_concat_layer
+----------------
+..  automodule:: paddle.trainer_config_helpers.layers
+    :members: seq_concat_layer
+    :noindex:
+
 Reshaping Layers
 ================
 
@@ -302,6 +308,12 @@ repeat_layer
     :members: repeat_layer
     :noindex:
 
+seq_reshape_layer
+-----------------
+..  automodule:: paddle.trainer_config_helpers.layers
+    :members: seq_reshape_layer
+    :noindex:
+
 Math Layers
 ===========
 

doc/design/api.md

+# PaddlePaddle Design Doc
+
+## Ingredients
+
+As our design principle is starting from the essence: how could we
+allow users to express and solve their problems at neural networks.
+Some essential concepts that our API have to provide include:
+
+1. A *topology* is an expression of *layers*.
+
+1. A layer could be any kind of computation, including *cost*.
+
+1. Some layers have parameters, some don't. Most costs don't have
+   parameters.
+
+1. In some topologies, layers share parameters.  For
+   example,
+   [the network for training a ranking model](https://github.com/PaddlePaddle/Paddle/issues/1311#issuecomment-279121850).
+
+1. At programming time, users specify topologies and possible sharing
+   of parameters.  PaddlePaddle can figure out and create parameters
+   required (and possibly shared) by one or more topologies.
+
+
+## Starting from Examples
+
+As a summarization
+of
+[our disucssion](https://github.com/PaddlePaddle/Paddle/issues/1315),
+let us present two examples here:
+
+
+### Example 1. Sharing Parameters between Layers
+
+We use
+the
+[3-branch ranking](https://github.com/PaddlePaddle/Paddle/issues/1311#issuecomment-279121850) model
+in this example.  For your convenience, I copy-a-paste the model's
+topology as follows:
+
+```
+A -> f -\
+Q -> f --> cost
+B -> f -/
+```
+
+The following program trains the topology including the cost, and then
+use the sub-network in the trained topology in inference:
+
+```python
+def f(in):
+    e = paddle.layer.embedding(in, parameter_name="embedding")
+    o = paddle.layer.softmax(e, parameter_name="semantic")
+    return o
+
+# Create 3 topologies (subnets), they share parameters because all
+# correspoinding layers have the same parameter names.
+fA = f(paddle.layer.data(input_name="A"))
+fB = f(paddle.layer.data(input_name="B"))
+fQ = f(paddle.layer.data(input_name="Q"))
+
+topology = paddle.layer.less_than(
+               paddle.layer.cross_entropy(fA, fQ),
+               paddle.layer.corss_entropy(fB, fQ))
+
+# Derive parameters required in topology and create them in model.
+parameters = paddle.parameters.create(topology)
+
+# Estimate parameters used in topology from data.
+paddle.train(topology, parameters, reader=read_ranking_model_data)
+
+# Inference using fA (or fB or fC, as they share their parameters).
+[testA, testB, testQ] = read_ranking_model_data()
+print "The sematic-vector of testA: ", paddle.infer(fA, parameters, testA)
+```
+
+
+### Example 2. Sharing Parameters between "Models"
+
+We use [GAN](https://github.com/PaddlePaddle/book/tree/develop/gan) in
+this example.  In the following example program, `d0` and `d1`
+correspond to the two networks in the following figure:
+
+<img src="https://github.com/wangyang59/book/raw/00036f4b0da5225041a6824587c1a01cf20159b1/gan/image/gan_ig.png" width=400 />
+
+```python
+def G(in):
+    # over-simplified example as G has only one layers:
+    return paddle.layer.fc(in, parameter_name="G")
+
+def D(in);
+    # again, over-simplified:
+    return paddle.layer.fc(in, parameter_name="D")
+
+# Construct the first topology, which contains both D and G.
+# By learning this topology, we update parameters of G.
+d0 = paddle.layer.should_be_false(D(G(paddle.layer.data())))
+
+# Construct a second topology d1, which contains only D. By
+# training this topology, we update parameters of D.  Note
+# that d1 share parameters with d0.
+d1 = paddle.layer.should_be_true(D(paddle.layer.data()))
+
+# Create parameters from a list of multiple topologies (models) for
+# the chance to share parameters between these topologies.
+parameters = paddle.parameters.create([d0, d1])
+
+# Iterative training of GAN.
+for ...:
+    train(d0, parameters, reader=read_from_rng, immutable_parameters={"D"})
+    train(d1, parameters, reader=read_from_realistic_images)
+
+# Use d1 for inference:
+print "D thinks a batch of images are realistic ", infer(d1, parameters, read_mnist_images)
+```
+
+
+### Summarization
+
+
+Above two programs reveal some important design concerns:
+
+1. Users describe a topology as an expression of layers.  Every layer
+   has a *parameter name*.  If the users don't specify it explicitly, it's automatically generated as a unique name.  By
+   specifying the parameter name, users can specify the sharing of
+   parameters between layers and even between topologies.
+
+1. `paddle.parameters.create` figures out parameters required by one
+   or more topologies from parameter names of layers.  It creates these
+   parameters and returns a `ParameterSet` object, which is in essence
+   a map from *parameter names* to *parameters*.
+
+1. At training and inference time, `paddle.train` and `paddle.infer`
+   requires both a topology and the parameter set that holds the parameters of that topology.  There are some reasons:
+
+   1. This prevents users from forgetting to call
+      `paddle.parameters.create`.
+   1. `paddle.train` needs to know which parameter set to update.
+   1. Users could load another (pre-trained) parameter set and use it
+      with a topology in `train.infer`.
+
+1. By specifying the `immutable_parameters` parameter of
+   `paddle.train`, we can forbid the update of these parameters.
+
+
+## Reader
+
+Not all programming frameworks allow users to define I/O functions.
+An example is Google MapReduce, which can only read from text,
+SSTable, and RecordIO files.  Hadoop MapReduce allows users to define
+readers and writers by deriving from base classes `Reader` and
+`Writer`.  The former is less flexible but also less error-prone.  We
+decide to provide the flexibility to users to define their readers.
+
+
+There are some open questions here:
+
+1. **Should a reader return a Python dictionary?**
+
+1. **How to map multiple outputs from a reader to multiple data layers?**
+
+1. **How to easily compose some existing readers to read more data and
+   feed a topology with more data layers?**
+
+
+## Training
+
+The recommended way to training a model is to call `paddle.train`,
+which simply calls `paddle.trainer.Default`, a global variable of
+type `paddle.trainer.SGD`.  Equivalently, we can do
+
+```python
+opt = paddle.trainer.SGD(..., paddle.updater.Adam(...))
+opt.train(topology, parameters, reader=read, ...)
+```
+
+### Updater
+
+Please be aware that a trainer can accept an updater as its data
+member, where an updater is a class derived from
+`paddle.trainer.Updater`.  This is to make it easier to customize
+trainers, as discussed
+[here](https://github.com/PaddlePaddle/Paddle/issues/1319).
+
+### Event Handler
+
+`paddle.train` and `paddle.trainer.XXX.train` take an optional
+parameter `event_handler`, which should be either `None` or a function
+that handle some events:
+
+1. BeginTraining
+1. EndTraining
+1. BeginIteration
+1. EndIteration
+1. BeginPass
+1. EndPass
+
+where EndPass is sent if and only if the reader yields
+`end_pass=True`.
+
+An example as follows:
+
+```python
+def event_handler(event):
+    if ininstance(event, paddle.event.EndIteration):
+        print paddle.test(...)
+
+paddle.train(topology, parameters, reader, event_handler)
+```
+
+If we are writing a PaddlePaddle program in and for iPython/Jypyter,
+we can use metaplotlib in the event handler to plot a curve of
+cost/error versus iterations, as shown
+[here](https://blog.dominodatalab.com/interactive-dashboards-in-jupyter/).
+
+### Distributed Training
+
+If users want to do distributed training on a cluster, s/he should
+call `paddle.dist_train` and provides access tokens to the cluster as
+a parameter.
+
+For example, if the user has a TLS certificate that allows him to
+access a Kubernetes cluster, s/he should be able to call
+
+```python
+paddle.dist_train(model,
+                  trainer=paddle.trainer.SGD(...,
+                                             paddle.updater.Adam(...)),
+                  reader=read,
+                  k8s_user="yi",
+                  k8s_token="kube_cluster_tls.pem",
+                  k8s_job="hello",
+                  num_parameter_servers=15)
+```
+
+The pseudo code if `paddle.dist_train` is as follows:
+
+```python
+def dist_train(topology, parameters, trainer, reader, ...):
+    if os.getenv("KUBERNETES_SERVICE_HOST") == None:
+        image_name = k8s_user + '/' + k8s_job
+        docker_build(image_name)
+        docker_push()
+        kube_ctrl_start_job(image_name, k8s_user, k8s_token)
+    else:
+        rank = kube_list_containers_in_job_and_return_current_containers_rank()
+        if rank == 0:
+            master()
+        elif rank < 15:
+            parameter_server()
+        else:
+            trainer.train(model, reader=read)
+```
+
+Please be aware that if a process is running on the Kubernetes
+cluster, it will have some environment variables pre-defined.
+
+If `dist_train` doesn't see these environment variables, it knows
+that it's running on users' personal computer, and it should work as a
+*launcher*.  Otherwise, it knows that it's running on the cluster and
+need to figure out its role as either the master, or a trainer, or a
+parameter server.

doc/design/reader/README.md

+# Python Data Reader Design Doc
+
+At training and testing time, PaddlePaddle programs need to read data. To ease the users' work to write data reading code, we define that
+
+- A *reader* is a function that reads data (from file, network, random number generator, etc) and yields data items.
+- A *reader creator* is a function that returns a reader function.
+- A *reader* decorator is a function, which accepts one or more readers, and returns a reader.
+
+and provide frequently used reader creators and reader decorators.
+
+## Data Reader Interface
+
+Indeed, *data reader* doesn't have to be a function that reads and yields data items. It can be any function with no parameter that creates a iterable (anything can be used in `for x in iterable`):
+
+```
+iterable = data_reader()
+```
+
+Element produced from the iterable should be a **single** entry of data, **not** a mini batch. That entry of data could be a single item, or a tuple of items. Item should be of [supported type](http://www.paddlepaddle.org/doc/ui/data_provider/pydataprovider2.html?highlight=dense_vector#input-types) (e.g., numpy 1d array of float32, int, list of int)
+
+An example implementation for single item data reader creator:
+
+```python
+def reader_creator_random_image(width, height):
+	def reader():
+		while True:
+			yield numpy.random.uniform(-1, 1, size=width*height)
+	return reader
+```
+
+An example implementation for multiple item data reader creator:
+```python
+def reader_creator_random_imageand_label(widht, height, label):
+	def reader():
+		while True:
+			yield numpy.random.uniform(-1, 1, size=width*height), label
+	return reader
+```
+
+## Usage
+
+data reader, mapping from item(s) read to data layer, batch size and number of total pass will be passed into `paddle.train`:
+
+```python
+# two data layer is created:
+image_layer = paddle.layer.data("image", ...)
+label_layer = paddle.layer.data("label", ...)
+
+# ...
+
+paddle.train(paddle.dataset.mnist, {"image":0, "label":1}, 128, 10, ...)
+```
+
+## Data Reader Decorator
+
+*Data reader decorator* takes a single or multiple data reader, returns a new data reader. It is similar to a [python decorator](https://wiki.python.org/moin/PythonDecorators), but it does not use `@` syntax.
+
+Since we have a strict interface for data readers (no parameter, return a single data item). Data reader can be used flexiable via data reader decorators. Following are a few examples:
+
+### Prefetch Data
+
+Since reading data may take time and training can not proceed without data. It is generally a good idea to prefetch data.
+
+Use `paddle.reader.buffered` to prefetch data:
+
+```python
+buffered_reader = paddle.reader.buffered(paddle.dataset.mnist, 100)
+```
+
+`buffered_reader` will try to buffer (prefetch) `100` data entries.
+
+### Compose Multiple Data Readers
+
+For example, we want to use a source of real images (reusing mnist dataset), and a source of random images as input for [Generative Adversarial Networks](https://arxiv.org/abs/1406.2661).
+
+We can do:
+
+```python
+def reader_creator_random_image(width, height):
+	def reader():
+		while True:
+			yield numpy.random.uniform(-1, 1, size=width*height)
+	return reader
+
+def reader_creator_bool(t):
+	def reader:
+		while True:
+			yield t
+	return reader
+
+true_reader = reader_creator_bool(True)
+false_reader = reader_creator_bool(False)
+
+reader = paddle.reader.compose(paddle.dataset.mnist, data_reader_creator_random_image(20, 20), true_reader, false_reader)
+# Skipped 1 because paddle.dataset.mnist produces two items per data entry.
+# And we don't care second item at this time.
+paddle.train(reader, {"true_image":0, "fake_image": 2, "true_label": 3, "false_label": 4}, ...)
+```
+
+### Shuffle
+
+Given shuffle buffer size `n`, `paddle.reader.shuffle` will return a data reader that buffers `n` data entries and shuffle them before a data entry is read.
+
+Example:
+```python
+reader = paddle.reader.shuffle(paddle.dataset.mnist, 512)
+```
+
+## Q & A
+
+### Why return only a single entry, but not a mini batch?
+
+If a mini batch is returned, data reader need to take care of batch size. But batch size is a concept for training, it makes more sense for user to specify batch size as a parameter for `train`.
+
+Practically, always return a single entry make reusing existing data readers much easier (e.g., if existing reader return not a single entry but 3 entries, training code will be more complex because it need to handle cases like batch size 2).
+
+### Why use a dictionary but not a list to provide mapping?
+
+We decided to use dictionary (`{"image":0, "label":1}`) instead of list (`["image", "label"]`) is because that user can easily resue item (e.g., using `{"image_a":0, "image_b":0, "label":1}`) or skip item (e.g., using `{"image_a":0, "label":2}`).
+
+### How to create custom data reader creator
+
+```python
+def image_reader_creator(image_path, label_path, n):
+	def reader():
+		f = open(image_path)
+		l = open(label_path)
+		images = numpy.fromfile(
+			f, 'ubyte', count=n * 28 * 28).reshape((n, 28 * 28)).astype('float32')
+		images = images / 255.0 * 2.0 - 1.0
+		labels = numpy.fromfile(l, 'ubyte', count=n).astype("int")
+		for i in xrange(n):
+			yield images[i, :], labels[i] # a single entry of data is created each time
+		f.close()
+		l.close()
+	return reader
+
+# images_reader_creator creates a reader
+reader = image_reader_creator("/path/to/image_file", "/path/to/label_file", 1024)
+paddle.train(reader, {"image":0, "label":1}, ...)
+```
+
+### How is `paddle.train` implemented
+
+An example implementation of paddle.train could be:
+
+```python
+def make_minibatch(reader, minibatch_size):
+	def ret():
+		r = reader()
+		buf = [r.next() for x in xrange(minibatch_size)]
+		while len(buf) > 0:
+			yield buf
+			buf = [r.next() for x in xrange(minibatch_size)]
+	return ret
+
+def train(reader, mapping, batch_size, total_pass):
+	for pass_idx in range(total_pass):
+		for mini_batch in make_minibatch(reader): # this loop will never end in online learning.
+			do_forward_backward(mini_batch, mapping)
+```

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.
  *
@@ -188,48 +175,6 @@ extern void hl_param_relu_backward_diff(real* grad_o,
                                         int width,
                                         int height,
                                         int partial_sum);
-/**
- * @brief cos sim forward
- *
- * @param[out]    output         output data
- * @param[in]     input1         input1 data(matrix)
- * @param[in]     input2         input2 data(matrix or vector)
- * @param[in]     width          matrix width
- * @param[in]     input1_height  input1_height
- * @param[in]     input2_height  input2_height
- * @param[in]     scale          scale factor
- */
-extern void hl_cossim(real* output,
-                      real* input1,
-                      real* input2,
-                      int width,
-                      int input1_height,
-                      int input2_height,
-                      real scale);
-/**
- * @brief cos sim derivate
- *
- * @param[in]     grad             output grad
- * @param[in]     output           output data
- * @param[in]     prevOutX         input1 data
- * @param[in]     prevOutY         input2 data
- * @param[out]    prevGradX        input1 grad
- * @param[out]    prevGradY        input2 grad
- * @param[in]     width            matrix width
- * @param[in]     input1_height    input1 height
- * @param[in]     input2_height    input2 height
- * @param[in]     scale            scale factor
- */
-extern void hl_cossim_derivative(real* grad,
-                                 real* output,
-                                 real* prevOutX,
-                                 real* prevOutY,
-                                 real* prevGradX,
-                                 real* prevGradY,
-                                 int width,
-                                 int input1_height,
-                                 int input2_height,
-                                 real scale);
 
 /**
  * @brief   Matrix addition: A_d[i][j] += scale * B_d[j/channel].

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) {}
@@ -74,25 +82,6 @@ inline void hl_param_relu_backward_diff(real* grad_o,
                                         int height,
                                         int partial_sum) {}
 
-inline void hl_cossim(real* output,
-                      real* input1,
-                      real* input2,
-                      int width,
-                      int input1_height,
-                      int input2_height,
-                      real scale) {}
-
-inline void hl_cossim_derivative(real* grad,
-                                 real* output,
-                                 real* prevOutX,
-                                 real* prevOutY,
-                                 real* prevGradX,
-                                 real* prevGradY,
-                                 int width,
-                                 int input1_height,
-                                 int input2_height,
-                                 real scale) {}
-
 inline void hl_matrix_add_shared_bias(real* A_d,
                                       real* B_d,
                                       const int channel,

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,
@@ -584,177 +531,6 @@ void hl_param_relu_backward_diff(real* grad_o,
   CHECK_SYNC("hl_param_relu_backward_diff failed");
 }
 
-template<int blockSize>
-__global__ void KeCosSim(real* output,
-                         real* input1,
-                         real* input2,
-                         int width,
-                         int input1_height,
-                         int input2_height,
-                         real scale) {
-  const int ty = blockIdx.y;
-  int tid = threadIdx.x;
-
-  __shared__ real xx[blockSize];
-  __shared__ real yy[blockSize];
-  __shared__ real xy[blockSize];
-
-  xx[tid] = 0.0;
-  yy[tid] = 0.0;
-  xy[tid] = 0.0;
-  __syncthreads();
-
-  input1 += ty * width;
-  if (input2_height > 1) {
-    input2 += ty * width;
-  }
-  for (int index = tid; index < width; index += blockSize) {
-    real x = input1[index];
-    real y = input2[index];
-    xx[tid] += x * x;
-    yy[tid] += y * y;
-    xy[tid] += x * y;
-  }
-  __syncthreads();
-
-  for (int s = blockSize / 2; s > 0; s >>= 1) {
-    if (tid < s) {
-      xx[tid] += xx[tid + s];
-      yy[tid] += yy[tid + s];
-      xy[tid] += xy[tid + s];
-    }
-    __syncthreads();
-  }
-  if (tid == 0) {
-    output[ty] = scale * xy[0] / (sqrt(xx[0]) * sqrt(yy[0]));
-  }
-}
-
-void hl_cossim(real* output,
-               real* input1,
-               real* input2,
-               int width,
-               int input1_height,
-               int input2_height,
-               real scale) {
-  CHECK_NOTNULL(output);
-  CHECK_NOTNULL(input1);
-  CHECK_NOTNULL(input2);
-  const int blockSize = 256;
-  dim3 threads(blockSize, 1);
-  dim3 grid(1, input1_height);
-
-  KeCosSim<blockSize><<<grid, threads, 0, STREAM_DEFAULT>>>
-    (output, input1, input2, width, input1_height, input2_height, scale);
-  CHECK_SYNC("hl_cossim failed");
-}
-
-template<int blockSize>
-__global__ void KeCosSimDerivative(real* grad,
-                                   real* output,
-                                   real* prevOutX,
-                                   real* prevOutY,
-                                   real* prevGradX,
-                                   real* prevGradY,
-                                   int width,
-                                   int input1_height,
-                                   int input2_height,
-                                   real scale) {
-  const int ty = blockIdx.y;
-  int tid = threadIdx.x;
-
-  __shared__ real xx[blockSize];
-  __shared__ real yy[blockSize];
-  __shared__ real xy[blockSize];
-
-  xx[tid] = 0.0;
-  yy[tid] = 0.0;
-  xy[tid] = 0.0;
-  __syncthreads();
-
-  prevOutX += ty * width;
-  prevGradX += ty * width;
-  if (input2_height > 1) {
-    prevOutY += ty * width;
-    prevGradY += ty * width;
-  }
-  for (int index = tid; index < width; index += blockSize) {
-    real x = prevOutX[index];
-    real y = prevOutY[index];
-    xx[tid] += x * x;
-    yy[tid] += y * y;
-    xy[tid] += x * y;
-  }
-  __syncthreads();
-
-  for (int s = blockSize / 2; s > 0; s >>= 1) {
-    if (tid < s) {
-      xx[tid] += xx[tid + s];
-      yy[tid] += yy[tid + s];
-      xy[tid] += xy[tid + s];
-    }
-    __syncthreads();
-  }
-  if (xy[0] == 0) {
-    real reciprocal = 1.0 / (sqrt(xx[0]) * sqrt(yy[0]));
-    for (int index = tid; index < width; index += blockSize) {
-      prevGradX[index] +=
-        scale * grad[ty] * prevOutY[index] * reciprocal;
-      if (input2_height > 1) {
-        prevGradY[index] +=
-          scale * grad[ty] * prevOutX[index] * reciprocal;
-      } else {
-        paddle::paddleAtomicAdd(prevGradY + index,
-          scale * grad[ty] * prevOutX[index] * reciprocal);
-      }
-    }
-  } else {
-    real reciprocalXY = 1.0 / xy[0];
-    real reciprocalSquareSumX = 1.0 / xx[0];
-    real reciprocalSquareSumY = 1.0 / yy[0];
-    for (int index = tid; index < width; index += blockSize) {
-      prevGradX[index] += output[ty] * grad[ty] *
-        (prevOutY[index] * reciprocalXY -
-         prevOutX[index] * reciprocalSquareSumX);
-      if (input2_height > 1) {
-        prevGradY[index] += output[ty] * grad[ty] *
-          (prevOutX[index] * reciprocalXY -
-           prevOutY[index] * reciprocalSquareSumY);
-      } else {
-        paddle::paddleAtomicAdd(prevGradY + index, output[ty] * grad[ty] *
-          (prevOutX[index] * reciprocalXY -
-           prevOutY[index] * reciprocalSquareSumY));
-      }
-    }
-  }
-}
-
-
-void hl_cossim_derivative(real* grad,
-                          real* output,
-                          real* prevOutX,
-                          real* prevOutY,
-                          real* prevGradX,
-                          real* prevGradY,
-                          int width,
-                          int input1_height,
-                          int input2_height,
-                          real scale) {
-  CHECK_NOTNULL(grad);
-  CHECK_NOTNULL(output);
-  CHECK_NOTNULL(prevOutX);
-  CHECK_NOTNULL(prevOutY);
-  CHECK_NOTNULL(prevGradX);
-  CHECK_NOTNULL(prevGradY);
-  const int blockSize = 256;
-  dim3 threads(blockSize, 1);
-  dim3 grid(1, input1_height);
-  KeCosSimDerivative<blockSize><<<grid, threads, 0, STREAM_DEFAULT>>>
-    (grad, output, prevOutX, prevOutY, prevGradX, prevGradY, width,
-        input1_height, input2_height, scale);
-  CHECK_SYNC("hl_cossim_derivate failed");
-}
-
 __global__ void KeMatrixAddSharedBias(real* A,
                                       real* B,
                                       const int channel,

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/cuda/src/hl_warpctc_wrap.cc

@@ -54,22 +54,26 @@ DYNAMIC_LOAD_WARPCTC_WRAP(get_workspace_size)
 #define WARPCTC_GET_VERSION dynload::get_warpctc_version
 #define WARPCTC_GET_STATUS_STRING dynload::ctcGetStatusString
 
+static int g_warpctcVersion = -1;
 #ifndef PADDLE_TYPE_DOUBLE
 #define WARPCTC_COMPUTE_LOSS dynload::compute_ctc_loss
 #define WARPCTC_GET_WORKSPACE_SIZE dynload::get_workspace_size
 #else
-#define WARPCTC_LOG_FATAL                                \
-  LOG(FATAL) << "warp-ctc [version " << g_warpctcVersion \
-             << "] Error: not support double precision."
-#define WARPCTC_COMPUTE_LOSS(...) WARPCTC_LOG_FATAL(__VA_ARGS__)
-#define WARPCTC_GET_WORKSPACE_SIZE(...) WARPCTC_LOG_FATAL(__VA_ARGS__)
+hl_warpctc_status_t fatal(...) {
+  LOG(FATAL) << "warp-ctc [version " << g_warpctcVersion
+             << "] Error: not support double precision.";
+  // both of get_warpctc_version() and get_workspace_size() return an ctcStatus
+  // type value
+  return CTC_STATUS_EXECUTION_FAILED;
+}
+#define WARPCTC_COMPUTE_LOSS fatal
+#define WARPCTC_GET_WORKSPACE_SIZE fatal
 #endif
 
 /**
  * Check build-in warp-ctc function using glog and it also
  * support << operator for more details error info.
  */
-static int g_warpctcVersion = -1;
 #define CHECK_WARPCTC(warpctcStat)                \
   CHECK_EQ(CTC_STATUS_SUCCESS, warpctcStat)       \
       << "warp-ctc [version " << g_warpctcVersion \

paddle/function/BufferArg.h

@@ -190,7 +190,7 @@ public:
       : BufferArg(VALUE_TYPE_INT32, shape, argType) {
     bufferType_ = TENSOR_SEQUENCE_ID;
     CHECK_EQ(shape_.ndims(), 1UL);
-    CHECK_GT(shape_[0], 1UL);
+    CHECK_GE(shape_[0], 1UL);
     numSeqs_ = shape_[0] - 1;
   }
 
@@ -226,7 +226,8 @@ public:
   SequenceArg(ValueType valueType,
               const TensorShape& shape,
               ArgType argType = UNSPECIFIED)
-      : BufferArg(valueType, shape, argType), startPositions_(TensorShape()) {
+      : BufferArg(valueType, shape, argType),
+        startPositions_(TensorShape({shape[0]})) {
     bufferType_ = TENSOR_SEQUENCE_DATA;
   }
 

paddle/function/CMakeLists.txt

@@ -27,6 +27,7 @@ if(WITH_TESTING)
     add_simple_unittest(ContextProjectionOpTest)
     add_simple_unittest(PadOpTest)
     add_simple_unittest(MulOpTest)
+    add_simple_unittest(CosSimOpTest)
 endif()
 endif()
 

paddle/function/ContextProjectionOp.cpp

@@ -108,26 +108,23 @@ public:
   }
 
   void calc(const BufferArgs& inputs, const BufferArgs& outputs) override {
-    CHECK(1 == inputs.size() || 2 == inputs.size());
-    CHECK_EQ((size_t)1, outputs.size());
+    CHECK(1UL == inputs.size() || 2UL == inputs.size());
+    CHECK_EQ(1UL, outputs.size());
     CHECK(inputs[0].isSequenceArg() && outputs[0].isSequenceArg())
         << "SequenceArg required here";
     const auto val_seqs = dynamic_cast<const SequenceArg&>(inputs[0]);
     auto out_seq = dynamic_cast<const SequenceArg&>(outputs[0]);
 
     CHECK(out_seq.data() && val_seqs.data() && val_seqs.getSequenceId().data());
-    CHECK_EQ(out_seq.shape().ndims(), (size_t)2);
-    CHECK_EQ(val_seqs.shape().ndims(), (size_t)2);
-    CHECK_EQ(val_seqs.getSequenceId().shape().ndims(), (size_t)1);
-    if (2 == inputs.size()) {
-      CHECK_EQ(inputs[1].shape().ndims(), (size_t)2);
-    }
+    CHECK_EQ(out_seq.shape().ndims(), 2UL);
+    CHECK_EQ(val_seqs.shape().ndims(), 2UL);
     /// dim of output = dim of input * context_length
     CHECK_EQ(out_seq.shape()[1], val_seqs.shape()[1] * context_length_);
     /// input and output has the same batch_size
     CHECK_EQ(val_seqs.shape()[0], out_seq.shape()[0]);
-    /// dim of input == dim of weight
-    if (2 == inputs.size()) {
+    if (2UL == inputs.size()) {
+      CHECK_EQ(inputs[1].shape().ndims(), 2UL);
+      /// dim of input == dim of weight
       CHECK_EQ(val_seqs.shape()[1], inputs[1].shape()[1]);
     }
 
@@ -135,10 +132,11 @@ public:
     auto out_mat = out_seq.matrix<Device>();
     const auto in_mat = val_seqs.matrix<Device>();
     const auto w_mat =
-        (2 == inputs.size())
+        (2UL == inputs.size() && inputs[1].data())
             ? inputs[1].matrix<Device>()
             : typename Tensor<real, Device>::Matrix(nullptr, 0, 0);
     const auto seq_vec = val_seqs.getSequenceId().vector<int, Device>();
+
     ContextProjectionForward<Device>(out_mat,
                                      in_mat,
                                      w_mat,
@@ -235,36 +233,40 @@ public:
   }
 
   void calc(const BufferArgs& inputs, const BufferArgs& outputs) override {
-    CHECK_EQ((size_t)1, inputs.size());
-    CHECK_EQ((size_t)2, outputs.size());
+    CHECK_EQ(1UL, inputs.size());
+    CHECK(1UL == outputs.size() || 2UL == outputs.size());
     CHECK(inputs[0].isSequenceArg() && outputs[0].isSequenceArg())
         << "SequenceArg required here";
     const auto in_seq = dynamic_cast<const SequenceArg&>(inputs[0]);
     auto out_seq = dynamic_cast<const SequenceArg&>(outputs[0]);
     CHECK(in_seq.data() && in_seq.getSequenceId().data());
-    CHECK_EQ(in_seq.shape().ndims(), (size_t)2);
-    CHECK_EQ(in_seq.getSequenceId().shape().ndims(), (size_t)1);
-    CHECK_EQ(out_seq.shape().ndims(), (size_t)2);
-    CHECK_EQ(out_seq.getSequenceId().shape().ndims(), (size_t)1);
-    CHECK_EQ(outputs[1].shape().ndims(), (size_t)2);
+    CHECK_EQ(in_seq.shape().ndims(), 2UL);
+    CHECK_EQ(out_seq.shape().ndims(), 2UL);
+    CHECK_EQ(out_seq.getSequenceId().shape().ndims(), 1UL);
 
-    /// dim of input grad == dim of weight
-    CHECK_EQ(out_seq.shape()[1], outputs[1].shape()[1]);
     /// input and output grad has the same batch_size
     CHECK_EQ(out_seq.shape()[0], in_seq.shape()[0]);
     /// dim of output grad = dim of input grad * context_length
     CHECK_EQ(in_seq.shape()[1], out_seq.shape()[1] * context_length_);
     CHECK_EQ(out_seq.getArgType(), ADD_TO);
-    CHECK_EQ(outputs[1].getArgType(), ADD_TO);
+
+    if (2UL == outputs.size()) {
+      CHECK_EQ(outputs[1].shape().ndims(), 2UL);
+      /// dim of input grad == dim of weight
+      CHECK_EQ(out_seq.shape()[1], outputs[1].shape()[1]);
+      CHECK_EQ(outputs[1].getArgType(), ADD_TO);
+    }
 
     const auto seq_vec = in_seq.getSequenceId().vector<int, Device>();
     const auto out_grad_mat = in_seq.matrix<Device>();
     auto in_grad_mat =
         !out_seq.data() ? typename Tensor<real, Device>::Matrix(nullptr, 0, 0)
                         : out_seq.matrix<Device>();
-    auto w_grad_mat = !outputs[1].data()
-                          ? typename Tensor<real, Device>::Matrix(nullptr, 0, 0)
-                          : outputs[1].matrix<Device>();
+    auto w_grad_mat =
+        (2UL == outputs.size() && outputs[1].data())
+            ? outputs[1].matrix<Device>()
+            : typename Tensor<real, Device>::Matrix(nullptr, 0, 0);
+
     ContextProjectionBackward<Device>(out_grad_mat,
                                       in_grad_mat,
                                       w_grad_mat,
@@ -304,17 +306,17 @@ public:
   }
 
   void calc(const BufferArgs& inputs, const BufferArgs& outputs) override {
-    CHECK_EQ(1, static_cast<int>(inputs.size()));
-    CHECK_EQ(1, static_cast<int>(outputs.size()));
+    CHECK_EQ(1UL, inputs.size());
+    CHECK_EQ(1UL, outputs.size());
     CHECK(inputs[0].isSequenceArg() && outputs[0].isSequenceArg())
         << "SequenceArg required here";
     const auto in_seq = dynamic_cast<const SequenceArg&>(inputs[0]);
     const auto out_seq = dynamic_cast<const SequenceArg&>(outputs[0]);
 
     CHECK(in_seq.data() && out_seq.data() && in_seq.getSequenceId().data());
-    CHECK_EQ(static_cast<int>(out_seq.shape().ndims()), 2);
-    CHECK_EQ(static_cast<int>(in_seq.shape().ndims()), 2);
-    CHECK_EQ(static_cast<int>(in_seq.getSequenceId().shape().ndims()), 1);
+    CHECK_EQ(out_seq.shape().ndims(), 2UL);
+    CHECK_EQ(in_seq.shape().ndims(), 2UL);
+    CHECK_EQ(in_seq.getSequenceId().shape().ndims(), 1UL);
     /// output layer grad dim == input layer grad dim * context_length_
     CHECK_EQ(in_seq.shape().ndims(), out_seq.shape().ndims() * context_length_);
     /// input and output has the same batch_size
@@ -355,14 +357,14 @@ public:
   }
 
   void calc(const BufferArgs& inputs, const BufferArgs& outputs) override {
-    CHECK_EQ(1, static_cast<int>(inputs.size()));
-    CHECK_EQ(1, static_cast<int>(outputs.size()));
+    CHECK_EQ(1UL, inputs.size());
+    CHECK_EQ(1UL, outputs.size());
     CHECK(inputs[0].isSequenceArg()) << "SequenceArg required here";
     const auto in_seq = dynamic_cast<const SequenceArg&>(inputs[0]);
     CHECK(in_seq.data() && in_seq.getSequenceId().data() && outputs[0].data());
-    CHECK_EQ(static_cast<int>(outputs[0].shape().ndims()), 2);
-    CHECK_EQ(static_cast<int>(in_seq.shape().ndims()), 2);
-    CHECK_EQ(static_cast<int>(in_seq.getSequenceId().shape().ndims()), 1);
+    CHECK_EQ(outputs[0].shape().ndims(), 2UL);
+    CHECK_EQ(in_seq.shape().ndims(), 2UL);
+    CHECK_EQ(in_seq.getSequenceId().shape().ndims(), 1UL);
     CHECK_EQ(in_seq.shape()[0], outputs[0].shape()[0]);
     /// output layer grad dim == weight dim * context_length_
     CHECK_EQ(in_seq.shape()[1], outputs[0].shape()[1] * context_length_);

paddle/function/ContextProjectionOp.h

@@ -13,7 +13,6 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
-
 #include "Function.h"
 
 namespace paddle {

paddle/function/ContextProjectionOpTest.cpp

@@ -28,55 +28,26 @@ void testMatrixProjectionForward(int context_start,
                std::max(0, (int)(context_start + context_length - 1));
   if (pad == 0) is_padding = false;
 
-  FunctionCompare compare("ContextProjectionForward",
-                          FuncConfig()
-                              .set("context_length", context_length)
-                              .set("context_start", context_start)
-                              .set("begin_pad", std::max(0, -context_start)));
-
-  CpuMatrix cpu_in(batch_size, input_dim);
-  cpu_in.randomizeUniform();
-  GpuMatrix gpu_in(batch_size, input_dim);
-  gpu_in.copyFrom(cpu_in);
-  auto cpu_weight =
-      is_padding ? std::make_shared<CpuMatrix>(pad, input_dim) : nullptr;
-  auto gpu_weight =
-      is_padding ? std::make_shared<GpuMatrix>(pad, input_dim) : nullptr;
-  if (is_padding) {
-    cpu_weight->randomizeUniform();
-    gpu_weight->copyFrom(*cpu_weight);
+  FunctionCompare test("ContextProjectionForward",
+                       FuncConfig()
+                           .set("context_length", context_length)
+                           .set("context_start", context_start)
+                           .set("begin_pad", std::max(0, -context_start)));
+
+  // prepare input arguments
+  test.addSequence(SequenceIdArg(TensorShape{batch_size}));
+  test.addInputs(
+      SequenceArg(VALUE_TYPE_FLOAT, TensorShape{batch_size, input_dim}));
+  if (is_padding) {  // weight
+    test.addInputs(SequenceArg(VALUE_TYPE_FLOAT, TensorShape{pad, input_dim}));
   }
-  IVectorPtr cpu_seq;
-  generateSequenceStartPositions(batch_size, cpu_seq);
-  IVectorPtr gpu_seq = IVector::create(cpu_seq->getSize(), true);
-  gpu_seq->copyFrom(*cpu_seq);
-
-  CpuMatrix cpu_out(batch_size, input_dim * context_length);
-  GpuMatrix gpu_out(batch_size, input_dim * context_length);
-  cpu_out.randomizeUniform();
-  gpu_out.copyFrom(cpu_out);
-
-  BufferArgs cpu_inputs;
-  BufferArgs cpu_outputs;
-  cpu_inputs.addArg(cpu_in, *cpu_seq);
-  if (cpu_weight) {
-    cpu_inputs.addArg(*cpu_weight, *cpu_seq);
-  }
-  cpu_outputs.addArg(cpu_out, *cpu_seq, ADD_TO);
-
-  compare.getCpuFunction()->calc(cpu_inputs, cpu_outputs);
+  test.addOutputs(
+      SequenceArg(VALUE_TYPE_FLOAT,
+                  TensorShape{batch_size, input_dim * context_length}),
+      ADD_TO);
 
-  BufferArgs gpu_inputs;
-  BufferArgs gpu_outputs;
-  gpu_inputs.addArg(gpu_in, *gpu_seq);
-  if (gpu_weight) {
-    gpu_inputs.addArg(*gpu_weight, *gpu_seq);
-  }
-  gpu_outputs.addArg(gpu_out, *gpu_seq, ADD_TO);
-
-  compare.getGpuFunction()->calc(gpu_inputs, gpu_outputs);
-
-  autotest::TensorCheckEqual(cpu_out, gpu_out);
+  // run Function
+  test.run();
 }
 
 void testMatrixProjectionBackward(int context_start,
@@ -88,63 +59,31 @@ void testMatrixProjectionBackward(int context_start,
                std::max(0, (int)(context_start + context_length - 1));
   if (pad == 0) is_padding = false;
 
-  FunctionCompare compare("ContextProjectionBackward",
-                          FuncConfig()
-                              .set("context_length", context_length)
-                              .set("context_start", context_start)
-                              .set("begin_pad", std::max(0, -context_start))
-                              .set("is_padding", is_padding)
-                              .set("total_pad", pad));
-
-  CpuMatrix cpu_in_grad(batch_size, input_dim);
-  cpu_in_grad.randomizeUniform();
-  GpuMatrix gpu_in_grad(batch_size, input_dim);
-  gpu_in_grad.copyFrom(cpu_in_grad);
-
-  CpuMatrix cpu_out_grad(batch_size, input_dim * context_length);
-  cpu_out_grad.randomizeUniform();
-  GpuMatrix gpu_out_grad(batch_size, input_dim * context_length);
-  gpu_out_grad.copyFrom(cpu_out_grad);
-
-  IVectorPtr cpu_seq;
-  generateSequenceStartPositions(batch_size, cpu_seq);
-  IVectorPtr gpu_seq = IVector::create(cpu_seq->getSize(), true);
-  gpu_seq->copyFrom(*cpu_seq);
-
-  auto cpu_w_grad =
-      is_padding ? std::make_shared<CpuMatrix>(pad, input_dim) : nullptr;
-  auto gpu_w_grad =
-      is_padding ? std::make_shared<GpuMatrix>(pad, input_dim) : nullptr;
-  if (is_padding) {
-    cpu_w_grad->randomizeUniform();
-    gpu_w_grad->copyFrom(*cpu_w_grad);
+  FunctionCompare test("ContextProjectionBackward",
+                       FuncConfig()
+                           .set("context_length", context_length)
+                           .set("context_start", context_start)
+                           .set("begin_pad", std::max(0, -context_start))
+                           .set("is_padding", is_padding)
+                           .set("total_pad", pad));
+
+  // prepare input arguments
+  test.addSequence(SequenceIdArg(TensorShape{batch_size}));
+  test.addInputs(SequenceArg(
+      VALUE_TYPE_FLOAT, TensorShape{batch_size, input_dim * context_length}));
+  test.addOutputs(
+      SequenceArg(VALUE_TYPE_FLOAT, TensorShape{batch_size, input_dim}),
+      ADD_TO);
+  if (is_padding) {  // weight
+    test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{pad, input_dim}),
+                    ADD_TO);
   }
 
-  BufferArgs cpu_inputs;
-  BufferArgs cpu_outputs;
-  cpu_inputs.addArg(cpu_out_grad, *cpu_seq);
-  cpu_outputs.addArg(cpu_in_grad, *cpu_seq, ADD_TO);
-  cpu_outputs.addArg(
-      cpu_w_grad ? *cpu_w_grad : CpuMatrix(nullptr, 0, input_dim), ADD_TO);
-
-  compare.getCpuFunction()->calc(cpu_inputs, cpu_outputs);
-
-  BufferArgs gpu_inputs;
-  BufferArgs gpu_outputs;
-  gpu_inputs.addArg(gpu_out_grad, *gpu_seq);
-  gpu_outputs.addArg(gpu_in_grad, *gpu_seq, ADD_TO);
-  gpu_outputs.addArg(
-      gpu_w_grad ? *gpu_w_grad : GpuMatrix(nullptr, 0, input_dim), ADD_TO);
-
-  compare.getGpuFunction()->calc(gpu_inputs, gpu_outputs);
-
-  autotest::TensorCheckErr(cpu_in_grad, gpu_in_grad);
-  if (is_padding) {
-    autotest::TensorCheckErr(*cpu_w_grad, *gpu_w_grad);
-  }
+  // run Function
+  test.run();
 }
 
-TEST(ContextProjection, projection) {
+TEST(ContextProjection, Projection) {
   for (auto context_start : {-5, -3, -1, 0, 3}) {
     for (auto context_length : {1, 2, 5, 7}) {
       for (auto trainable_padding : {false, true}) {

paddle/function/CosSimOp.cpp

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "CosSimOp.h"
+#include "paddle/math/Matrix.h"
+#include "paddle/math/Vector.h"
+
+namespace paddle {
+/**
+ * Cosine Similarity for CpuMatrix
+ *
+ * \param out_mat, output value, size: nSamples * 1.
+ * \param in1_mat, input value 1, size: nSamples * dim.
+ * \param in2_mat, input value 2, size: n2 * dim (n2 == 1 or n2 == nSamples).
+ * \param scale, default 1.0
+ *
+ */
+template <>
+void CosSimForward<DEVICE_TYPE_CPU>(CpuMatrix& out_mat,
+                                    const CpuMatrix& in1_mat,
+                                    const CpuMatrix& in2_mat,
+                                    real scale) {
+  CHECK(out_mat.getData() && in1_mat.getData() && in2_mat.getData());
+  size_t num_samples = out_mat.getHeight();
+  size_t dim = in1_mat.getWidth();
+  /// column vector [nSamples, 1]
+  real* out = out_mat.getData();
+  const real* x = in1_mat.getData();
+  const real* y = in2_mat.getData();
+
+  /// in2 might only have one row or full rows
+  CHECK(in2_mat.getHeight() == 1LU || in2_mat.getHeight() == num_samples);
+  size_t inc = (in2_mat.getHeight() == 1LU) ? 0 : dim;
+  for (size_t i = 0; i < num_samples; ++i, x += dim, y += inc) {
+    real square_sum_x = 0;
+    real square_sum_y = 0;
+    real xy = 0;
+    for (size_t j = 0; j < dim; ++j) {
+      square_sum_x += x[j] * x[j];
+      square_sum_y += y[j] * y[j];
+      xy += x[j] * y[j];
+    }
+    CHECK(square_sum_x > 0 && square_sum_y > 0);
+    out[i] = scale * xy / (std::sqrt(square_sum_x) * std::sqrt(square_sum_y));
+  }
+}
+
+/**
+ * Cosine Similarity
+ * for each row i,
+ *   out[i] = scale * cos(input1[i], input2[i])
+ *      = scale * <input1[i], input2[i]>/sqrt(|input1[i]|^2 * |input2[i]|^2)
+ * when input2 only has one row, then for each row i,
+ *   out[i] = cos(input1[i], input2[0])
+ *
+ * \param inputs[0] input matrix 1, size: nSamples * dim.
+ * \param inputs[1] input matrix 2, size: n2 * dim (n2 == 1 or n2 == nSamples).
+ * \param outputs[0] output matrix, size : nSamples * 1.
+ */
+
+template <DeviceType Device>
+class CosSimForwardFunc : public FunctionBase {
+  void init(const FuncConfig& config) override {
+    scale_ = config.get<real>("scale");
+  }
+
+  void calc(const BufferArgs& inputs, const BufferArgs& outputs) override {
+    CHECK_EQ(inputs.size(), 2UL);
+    CHECK_EQ(outputs.size(), 1UL);
+
+    CHECK_EQ(inputs[0].shape().ndims(), 2UL);
+    CHECK_EQ(inputs[1].shape().ndims(), 2UL);
+    CHECK_EQ(outputs[0].shape().ndims(), 2UL);
+
+    CHECK_EQ(inputs[0].shape()[0], outputs[0].shape()[0]);
+    CHECK_EQ(inputs[0].shape()[1], inputs[1].shape()[1]);
+    CHECK_EQ(outputs[0].shape()[1], 1UL);
+
+    CHECK(outputs[0].data() && inputs[0].data() && inputs[1].data());
+
+    CHECK_EQ(outputs[0].getArgType(), ASSIGN_TO);
+    auto out_mat = outputs[0].matrix<Device>();
+    const auto in1_mat = inputs[0].matrix<Device>();
+    const auto in2_mat = inputs[1].matrix<Device>();
+
+    CosSimForward<Device>(out_mat, in1_mat, in2_mat, scale_);
+  }
+
+private:
+  real scale_;
+};
+
+/**
+ * Cosine Similarity Derivative for CpuMatrix
+ *
+ * \param in1_grad  forward input grad 1, size: nSamples * dim.
+ * \param in2_grad  forward input grad 2,
+ *                  size: n2 * dim (n2 == 1 or n2 == nSamples).
+ *
+ * \param out_grad  backward loss output grad, size : nSamples * 1.
+ * \param out_val   forward output value, size: nSamples * 1.
+ * \param in1_val   forward input value 1, size: nSamples * dim.
+ * \param in2_val   forward input value 2,
+ *                  size: n2 * dim (n2 == 1 or n2 == nSamples).
+ * \param scale,    default 1.0
+ */
+template <>
+void CosSimBackward<DEVICE_TYPE_CPU>(const CpuMatrix& out_grad,
+                                     const CpuMatrix& out_val,
+                                     const CpuMatrix& in1_val,
+                                     const CpuMatrix& in2_val,
+                                     CpuMatrix& in1_grad,
+                                     CpuMatrix& in2_grad,
+                                     real scale) {
+  CHECK(out_grad.getData() && out_val.getData() && in1_val.getData() &&
+        in2_val.getData() && in1_grad.getData() && in2_grad.getData());
+  CHECK_EQ(out_val.useGpu_, false) << "Matrix type are GPU, CPU required";
+
+  const real* grad = out_grad.getData();
+  const real* out = out_val.getData();
+  const real* prev_out_x = in1_val.getData();
+  const real* prev_out_y = in2_val.getData();
+  real* prev_grad_x = in1_grad.getData();
+  real* prev_grad_y = in2_grad.getData();
+
+  size_t num_samples = out_grad.getHeight();
+  size_t dim = in1_val.getWidth();
+  CHECK_EQ(in2_val.getHeight(), in2_grad.getHeight());
+  CHECK(in2_val.getHeight() == 1LU || in2_val.getHeight() == num_samples);
+  size_t inc = (in2_val.getHeight() == 1LU) ? 0 : dim;
+  for (size_t i = 0; i < num_samples; ++i,
+              prev_out_x += dim,
+              prev_out_y += inc,
+              prev_grad_x += dim,
+              prev_grad_y += inc) {
+    real square_sum_x = 0;
+    real square_sum_y = 0;
+    real xy = 0;
+    for (size_t j = 0; j < dim; ++j) {
+      square_sum_x += prev_out_x[j] * prev_out_x[j];
+      square_sum_y += prev_out_y[j] * prev_out_y[j];
+      xy += prev_out_x[j] * prev_out_y[j];
+    }
+    CHECK(square_sum_x > 0 && square_sum_y > 0);
+    if (xy == 0) {
+      real reciprocal =
+          1.0f / (std::sqrt(square_sum_x) * std::sqrt(square_sum_y));
+      for (size_t j = 0; j < dim; ++j) {
+        prev_grad_x[j] += scale * grad[i] * prev_out_y[j] * reciprocal;
+        prev_grad_y[j] += scale * grad[i] * prev_out_x[j] * reciprocal;
+      }
+    } else {
+      real reciprocal_xy = 1.0f / xy;
+      real reciprocal_square_sum_x = 1.0f / square_sum_x;
+      real reciprocal_square_sum_y = 1.0f / square_sum_y;
+      for (size_t j = 0; j < dim; ++j) {
+        prev_grad_x[j] +=
+            out[i] * grad[i] * (prev_out_y[j] * reciprocal_xy -
+                                prev_out_x[j] * reciprocal_square_sum_x);
+        prev_grad_y[j] +=
+            out[i] * grad[i] * (prev_out_x[j] * reciprocal_xy -
+                                prev_out_y[j] * reciprocal_square_sum_y);
+      }
+    }
+  }
+}
+
+/**
+ * Cosine Similarity backward Derivative
+ *
+ * \param outputs[0] forward input grad 1, size: nSamples * dim.
+ * \param outputs[1] forward input grad 2,
+ *                  size: n2 * dim (n2 == 1 or n2 == nSamples).
+ *
+ * \param inputs[0] backward loss output grad, size : nSamples * 1.
+ * \param inputs[1] forward output value, size: nSamples * 1.
+ * \param inputs[2] forward input value 1, size: nSamples * dim.
+ * \param inputs[3] forward input value 2,
+ *                  size: n2 * dim (n2 == 1 or n2 == nSamples).
+ */
+template <DeviceType Device>
+class CosSimBackwardFunc : public FunctionBase {
+  void init(const FuncConfig& config) override {
+    scale_ = config.get<real>("scale");
+  }
+
+  void calc(const BufferArgs& inputs, const BufferArgs& outputs) override {
+    CHECK_EQ(inputs.size(), 4UL);
+    CHECK_EQ(outputs.size(), 2UL);
+    /// dim of out_grad and out_val == 1, column vector
+    CHECK_EQ(inputs[0].shape()[1], 1UL);
+    CHECK_EQ(inputs[1].shape()[1], 1UL);
+    /// nSamples of out_grad == out_val == in_val1 == in_grad1
+    CHECK_EQ(inputs[1].shape()[0], inputs[0].shape()[0]);
+    CHECK_EQ(inputs[0].shape()[0], inputs[0].shape()[0]);
+    CHECK_EQ(outputs[0].shape()[0], inputs[0].shape()[0]);
+    /// dim of in1_val1 == in_val2 == in_grad1 == in_grad2
+    CHECK_EQ(inputs[3].shape()[1], inputs[2].shape()[1]);
+    CHECK_EQ(outputs[0].shape()[1], inputs[2].shape()[1]);
+    CHECK_EQ(outputs[1].shape()[1], inputs[2].shape()[1]);
+
+    CHECK(inputs[0].data() && inputs[1].data() && inputs[2].data() &&
+          inputs[3].data() && outputs[0].data() && outputs[1].data());
+
+    CHECK_EQ(outputs[0].getArgType(), ADD_TO);
+    CHECK_EQ(outputs[1].getArgType(), ADD_TO);
+
+    const auto out_grad = inputs[0].matrix<Device>();
+    const auto out_val = inputs[1].matrix<Device>();
+    const auto in1_val = inputs[2].matrix<Device>();
+    const auto in2_val = inputs[3].matrix<Device>();
+    auto in1_grad = outputs[0].matrix<Device>();
+    auto in2_grad = outputs[1].matrix<Device>();
+
+    CosSimBackward<Device>(
+        out_grad, out_val, in1_val, in2_val, in1_grad, in2_grad, scale_);
+  }
+
+private:
+  real scale_;
+};
+
+REGISTER_TYPED_FUNC(CosSimForward, CPU, CosSimForwardFunc);
+REGISTER_TYPED_FUNC(CosSimBackward, CPU, CosSimBackwardFunc);
+#ifndef PADDLE_ONLY_CPU
+REGISTER_TYPED_FUNC(CosSimForward, GPU, CosSimForwardFunc);
+REGISTER_TYPED_FUNC(CosSimBackward, GPU, CosSimBackwardFunc);
+#endif
+}  // namespace paddle

paddle/function/CosSimOp.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#include "Function.h"
+
+namespace paddle {
+
+/**
+ * \brief   Cosine Similarity Forward.
+ * for each row i,
+ * out[i] = scale * cos(in1[i], in2[i])
+ *        = scale * \sum_j (in1[i][j] * in2[i][j]) /
+ *                  sqrt(sum_j (in1[i][j]^2) * sum_j (in2[i][j])^2)
+ *
+ * \param[out]  output            output value.
+ * \param[in]   intput1           input value.
+ * \param[in]   intput2           input value.
+ * \param[in]   scale             default 1.0.
+ *
+ */
+template <DeviceType Device>
+void CosSimForward(typename Tensor<real, Device>::Matrix& output,
+                   const typename Tensor<real, Device>::Matrix& input1,
+                   const typename Tensor<real, Device>::Matrix& input2,
+                   real scale);
+
+/**
+ * \brief   Cosine Similarity BackWard for Derivative.
+ *
+ * \param[in]       output grad           backward loss output grad.
+ * \param[in]       output val            forward-output value.
+ * \param[in]       input val1            forward input value 1.
+ * \param[in]       input val2            forward input value 2.
+ * \param[in/out]   input grad            forward input grad 1.
+ * \param[in/out]   input grad            forward input grad 2.
+ * \param[in]       scale                 default 1.0.
+ *
+ */
+template <DeviceType Device>
+void CosSimBackward(const typename Tensor<real, Device>::Matrix& out_grad,
+                    const typename Tensor<real, Device>::Matrix& out_value,
+                    const typename Tensor<real, Device>::Matrix& in1_value,
+                    const typename Tensor<real, Device>::Matrix& in2_value,
+                    typename Tensor<real, Device>::Matrix& in1_grad,
+                    typename Tensor<real, Device>::Matrix& in2_grad,
+                    real scale);
+
+}  // namespace paddle

paddle/function/CosSimOpGpu.cu

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "hl_base.h"
+#include "hl_device_functions.cuh"
+#include "CosSimOp.h"
+
+namespace paddle {
+
+template<int block_size>
+__global__ void KeCosSim(real* output,
+                         const real* input1,
+                         const real* input2,
+                         int width,
+                         int input1_height,
+                         int input2_height,
+                         real scale) {
+  const int ty = blockIdx.y;
+  int tid = threadIdx.x;
+
+  __shared__ real xx[block_size];
+  __shared__ real yy[block_size];
+  __shared__ real xy[block_size];
+
+  xx[tid] = 0.0;
+  yy[tid] = 0.0;
+  xy[tid] = 0.0;
+  __syncthreads();
+
+  input1 += ty * width;
+  if (input2_height > 1) {
+    input2 += ty * width;
+  }
+  for (int index = tid; index < width; index += block_size) {
+    real x = input1[index];
+    real y = input2[index];
+    xx[tid] += x * x;
+    yy[tid] += y * y;
+    xy[tid] += x * y;
+  }
+  __syncthreads();
+
+  for (int s = block_size / 2; s > 0; s >>= 1) {
+    if (tid < s) {
+      xx[tid] += xx[tid + s];
+      yy[tid] += yy[tid + s];
+      xy[tid] += xy[tid + s];
+    }
+    __syncthreads();
+  }
+  if (tid == 0) {
+    output[ty] = scale * xy[0] / (sqrt(xx[0]) * sqrt(yy[0]));
+  }
+}
+
+void hlCossim(real* output,
+              const real* input1,
+              const real* input2,
+              size_t width,
+              size_t input1_height,
+              size_t input2_height,
+              real scale) {
+  CHECK_NOTNULL(output);
+  CHECK_NOTNULL(input1);
+  CHECK_NOTNULL(input2);
+  const int block_size = 256;
+  dim3 threads(block_size, 1);
+  dim3 grid(1, input1_height);
+
+  KeCosSim<block_size><<<grid, threads, 0, STREAM_DEFAULT>>>
+    (output, input1, input2, width, input1_height, input2_height, scale);
+  CHECK_SYNC("hlCossim failed");
+}
+
+template <>
+void CosSimForward<DEVICE_TYPE_GPU>(GpuMatrix& out_mat,
+                                    const GpuMatrix& in1_mat,
+                                    const GpuMatrix& in2_mat,
+                                    real scale) {
+  CHECK(out_mat.getData() && in1_mat.getData() && in2_mat.getData());
+  CHECK(in1_mat.useGpu_ == true && in2_mat.useGpu_ == true)
+      << "Matrix type are not GPU";
+
+  size_t num_samples = out_mat.getHeight();
+  size_t dim = in1_mat.getWidth();
+  real* out = out_mat.getData();
+  const real* x = in1_mat.getData();
+  const real* y = in2_mat.getData();
+  hlCossim(out, x, y, dim, in1_mat.getHeight(), in2_mat.getHeight(), scale);
+}
+
+template<int block_size>
+__global__ void KeCosSimDerivative(const real* grad,
+                                   const real* output,
+                                   const real* prev_out_x,
+                                   const real* prev_out_y,
+                                   real* prev_grad_x,
+                                   real* prev_grad_y,
+                                   size_t width,
+                                   size_t input1_height,
+                                   size_t input2_height,
+                                   real scale) {
+  const int ty = blockIdx.y;
+  int tid = threadIdx.x;
+
+  __shared__ real xx[block_size];
+  __shared__ real yy[block_size];
+  __shared__ real xy[block_size];
+
+  xx[tid] = 0.0;
+  yy[tid] = 0.0;
+  xy[tid] = 0.0;
+  __syncthreads();
+
+  prev_out_x += ty * width;
+  prev_grad_x += ty * width;
+  if (input2_height > 1) {
+    prev_out_y += ty * width;
+    prev_grad_y += ty * width;
+  }
+  for (int index = tid; index < width; index += block_size) {
+    real x = prev_out_x[index];
+    real y = prev_out_y[index];
+    xx[tid] += x * x;
+    yy[tid] += y * y;
+    xy[tid] += x * y;
+  }
+  __syncthreads();
+
+  for (int s = block_size / 2; s > 0; s >>= 1) {
+    if (tid < s) {
+      xx[tid] += xx[tid + s];
+      yy[tid] += yy[tid + s];
+      xy[tid] += xy[tid + s];
+    }
+    __syncthreads();
+  }
+  if (xy[0] == 0) {
+    real reciprocal = 1.0 / (sqrt(xx[0]) * sqrt(yy[0]));
+    for (int index = tid; index < width; index += block_size) {
+      prev_grad_x[index] +=
+        scale * grad[ty] * prev_out_y[index] * reciprocal;
+      if (input2_height > 1) {
+        prev_grad_y[index] +=
+          scale * grad[ty] * prev_out_x[index] * reciprocal;
+      } else {
+        paddle::paddleAtomicAdd(prev_grad_y + index,
+          scale * grad[ty] * prev_out_x[index] * reciprocal);
+      }
+    }
+  } else {
+    real reciprocalXY = 1.0 / xy[0];
+    real reciprocalSquareSumX = 1.0 / xx[0];
+    real reciprocalSquareSumY = 1.0 / yy[0];
+    for (int index = tid; index < width; index += block_size) {
+      prev_grad_x[index] += output[ty] * grad[ty] *
+        (prev_out_y[index] * reciprocalXY -
+         prev_out_x[index] * reciprocalSquareSumX);
+      if (input2_height > 1) {
+        prev_grad_y[index] += output[ty] * grad[ty] *
+          (prev_out_x[index] * reciprocalXY -
+           prev_out_y[index] * reciprocalSquareSumY);
+      } else {
+        paddle::paddleAtomicAdd(prev_grad_y + index, output[ty] * grad[ty] *
+          (prev_out_x[index] * reciprocalXY -
+           prev_out_y[index] * reciprocalSquareSumY));
+      }
+    }
+  }
+}
+
+void hlCossimDerivative(const real* grad,
+                        const real* output,
+                        const real* prev_out_x,
+                        const real* prev_out_y,
+                        real* prev_grad_x,
+                        real* prev_grad_y,
+                        size_t width,
+                        size_t input1_height,
+                        size_t input2_height,
+                        real scale) {
+  CHECK_NOTNULL(grad);
+  CHECK_NOTNULL(output);
+  CHECK_NOTNULL(prev_out_x);
+  CHECK_NOTNULL(prev_out_y);
+  CHECK_NOTNULL(prev_grad_x);
+  CHECK_NOTNULL(prev_grad_y);
+  const int block_size = 256;
+  dim3 threads(block_size, 1);
+  dim3 grid(1, input1_height);
+  KeCosSimDerivative<block_size><<<grid, threads, 0, STREAM_DEFAULT>>>
+    (grad, output, prev_out_x, prev_out_y, prev_grad_x, prev_grad_y, width,
+        input1_height, input2_height, scale);
+  CHECK_SYNC("hlCossimDerivate failed");
+}
+
+template <>
+void CosSimBackward<DEVICE_TYPE_GPU>(const GpuMatrix& out_grad,
+                                     const GpuMatrix& out_val,
+                                     const GpuMatrix& in1_val,
+                                     const GpuMatrix& in2_val,
+                                     GpuMatrix& in1_grad,
+                                     GpuMatrix& in2_grad,
+                                     real scale) {
+  CHECK(out_grad.getData() && out_val.getData() && in1_val.getData() &&
+        in2_val.getData() && in1_grad.getData() && in2_grad.getData());
+  CHECK(out_grad.useGpu_ && out_val.useGpu_ && in1_val.useGpu_
+        && in2_val.useGpu_ && in1_grad.useGpu_ && in2_grad.useGpu_)
+        << "Matrix types are not equally GPU";
+
+  size_t dim = in1_val.getWidth();
+  const real* grad = out_grad.getData();
+  const real* out = out_val.getData();
+  const real* prev_out_x = in1_val.getData();
+  const real* prev_out_y = in2_val.getData();
+  real* prev_grad_x = in1_grad.getData();
+  real* prev_grad_y = in2_grad.getData();
+  hlCossimDerivative(grad,
+                     out,
+                     prev_out_x,
+                     prev_out_y,
+                     prev_grad_x,
+                     prev_grad_y,
+                     dim,
+                     in1_val.getHeight(),
+                     in2_val.getHeight(),
+                     scale);
+}
+
+}  // namespace paddle

paddle/function/CosSimOpTest.cpp

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include <gtest/gtest.h>
+#include "FunctionTest.h"
+#include "paddle/math/Matrix.h"
+
+using namespace paddle;  // NOLINT
+
+void testCosSimForward(size_t height_x,
+                       size_t height_y,
+                       size_t width,
+                       real scale) {
+  FunctionCompare test("CosSimForward", FuncConfig().set("scale", scale));
+  // prepare input arguments
+  test.addInputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{height_x, width}));
+  test.addInputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{height_y, width}));
+  test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{height_x, 1}),
+                  ASSIGN_TO);
+  // run Function
+  test.run();
+}
+
+void testCosSimBackward(size_t height_x,
+                        size_t height_y,
+                        size_t width,
+                        real scale) {
+  FunctionCompare test("CosSimBackward", FuncConfig().set("scale", scale));
+  // prepare input arguments
+  test.addInputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{height_x, 1}));
+  test.addInputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{height_x, 1}));
+  test.addInputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{height_x, width}));
+  test.addInputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{height_y, width}));
+  test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{height_x, width}),
+                  ADD_TO);
+  test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, TensorShape{height_y, width}),
+                  ADD_TO);
+  // run Function
+  test.run();
+}
+
+TEST(Matrix, cosSim) {
+  for (auto height_x : {10, 100, 1000}) {
+    for (auto height_y : {1, height_x}) {
+      for (auto width : {10, 100, 1000}) {
+        for (auto scale : {1.0, 2.0}) {
+          testCosSimForward(height_x, height_y, width, scale);
+          testCosSimBackward(height_x, height_y, width, scale);
+        }
+      }
+    }
+  }
+}

paddle/function/FunctionTest.h

@@ -69,6 +69,54 @@ public:
         gpuMemory_.back()->getBuf(), input.valueType(), input.shape()));
   }
 
+  // assume one copy of sequence is shared by different SequenceArgs
+  void addSequence(const SequenceIdArg& input) {
+    CHECK_EQ(input.shape().ndims(), 1UL);
+    size_t batchSize = input.shape()[0];
+    size_t numSeqs = batchSize / 10 + 1;
+    size_t sizeId = (numSeqs + 1) * sizeOfValuType(VALUE_TYPE_INT32);
+    cpuMemory_.emplace_back(std::make_shared<CpuMemoryHandle>(sizeId));
+    gpuMemory_.emplace_back(std::make_shared<GpuMemoryHandle>(sizeId));
+    cpuSeq_ = std::make_shared<SequenceIdArg>(cpuMemory_.back()->getBuf(),
+                                              TensorShape{numSeqs + 1});
+    gpuSeq_ = std::make_shared<SequenceIdArg>(gpuMemory_.back()->getBuf(),
+                                              TensorShape{numSeqs + 1});
+    /// init sequence Id
+    initArg(*cpuSeq_, batchSize);
+
+    // todo(tianbing), delete it
+    CHECK_EQ(cpuSeq_->shape().getElements(), cpuSeq_->numSeqs() + 1);
+
+    CpuIVector cpuSeq(cpuSeq_->shape().getElements(), (int*)cpuSeq_->data());
+    GpuIVector gpuSeq(gpuSeq_->shape().getElements(), (int*)gpuSeq_->data());
+    gpuSeq.copyFrom(cpuSeq);
+  }
+
+  void addInputs(const SequenceArg& input) {
+    CHECK_EQ(input.shape().ndims(), 2UL);
+    size_t batchSize = input.shape()[0];
+    if (!cpuSeq_ || !gpuSeq_) {  // sequence not exist
+      addSequence(SequenceIdArg(TensorShape{batchSize}));
+    }
+
+    size_t size =
+        input.shape().getElements() * sizeOfValuType(input.valueType());
+    cpuMemory_.emplace_back(std::make_shared<CpuMemoryHandle>(size));
+    gpuMemory_.emplace_back(std::make_shared<GpuMemoryHandle>(size));
+
+    /// SequenceArg
+    cpuInputs_.emplace_back(
+        std::make_shared<SequenceArg>(cpuMemory_.back()->getBuf(),
+                                      input.valueType(),
+                                      input.shape(),
+                                      *cpuSeq_));
+    gpuInputs_.emplace_back(
+        std::make_shared<SequenceArg>(gpuMemory_.back()->getBuf(),
+                                      input.valueType(),
+                                      input.shape(),
+                                      *gpuSeq_));
+  }
+
   // output need only contains shape, do not contains data.
   void addOutputs(const BufferArg& output, ArgType argType = ASSIGN_TO) {
     size_t size =
@@ -116,24 +164,31 @@ public:
         std::make_shared<SparseMatrixArg>(*gpuSparse_, argType));
   }
 
-  void addInputs(const SequenceArg& input) {
-    size_t batchSize = input.shape()[0];
-    size_t numSeqs = batchSize / 10 + 1;
-
-    size_t sizeId = (numSeqs + 1) * sizeOfValuType(VALUE_TYPE_INT32);
-    cpuMemory_.emplace_back(std::make_shared<CpuMemoryHandle>(sizeId));
-    gpuMemory_.emplace_back(std::make_shared<GpuMemoryHandle>(sizeId));
-
-    TensorShape seqsId({numSeqs + 1});
-    // void* cpuBuffer = cpuMemory_.back()->getBuf();
-    // void* gpuBuffer = gpuMemory_.back()->getBuf();
+  void addOutputs(const SequenceArg& output, ArgType argType = ASSIGN_TO) {
+    CHECK_EQ(output.shape().ndims(), 2UL);
+    size_t batchSize = output.shape()[0];
 
+    if (!cpuSeq_ || !gpuSeq_) {  // sequence not exist
+      addSequence(SequenceIdArg(TensorShape{batchSize}));
+    }
     size_t size =
-        input.shape().getElements() * sizeOfValuType(input.valueType());
+        output.shape().getElements() * sizeOfValuType(output.valueType());
     cpuMemory_.emplace_back(std::make_shared<CpuMemoryHandle>(size));
     gpuMemory_.emplace_back(std::make_shared<GpuMemoryHandle>(size));
 
-    // TODO: need be implemented.
+    /// SequenceArg
+    cpuOutputs_.emplace_back(
+        std::make_shared<SequenceArg>(cpuMemory_.back()->getBuf(),
+                                      output.valueType(),
+                                      output.shape(),
+                                      *cpuSeq_,
+                                      argType));
+    gpuOutputs_.emplace_back(
+        std::make_shared<SequenceArg>(gpuMemory_.back()->getBuf(),
+                                      output.valueType(),
+                                      output.shape(),
+                                      *gpuSeq_,
+                                      argType));
   }
 
   void addInputs(const SparseMatrixArg& input) {
@@ -193,14 +248,44 @@ public:
   std::shared_ptr<FunctionBase> getGpuFunction() const { return gpuFunc_; }
 
 protected:
+  // only init cpu argument, gpu argument copy from cpu argument.
+  void initArg(BufferArg& arg) {
+    CpuVector vector(arg.shape().getElements(), (real*)arg.data());
+    vector.uniform(0.001, 1);
+  }
+
+  void initArg(SequenceArg& arg) {
+    /// init only matrix
+    CpuVector vector(arg.shape().getElements(), (real*)arg.data());
+    vector.uniform(0.001, 1);
+  }
+
+  void initArg(SequenceIdArg& arg, size_t batchSize) {
+    size_t numSeqs = arg.numSeqs();
+    int* buf = reinterpret_cast<int*>(arg.data());
+    int pos = 0;
+    size_t maxLen = 2 * batchSize / numSeqs;
+    for (int i = 0; i < (int)numSeqs; ++i) {
+      int len = 1 + uniformRandom(std::min<int64_t>(
+                        maxLen, batchSize - pos - numSeqs + i));
+      buf[i] = pos;
+      pos += len;
+      VLOG(1) << " len=" << len;
+    }
+    buf[numSeqs] = batchSize;
+  }
+
   void initInputs() {
     for (size_t i = 0; i < cpuInputs_.size(); i++) {
       if (cpuInputs_[i]->isSparseArg()) {
         continue;  /// sparse matrix already init
       }
 
-      initArg(*cpuInputs_[i]);
-
+      if (cpuInputs_[i]->isSequenceArg()) {
+        initArg(dynamic_cast<SequenceArg&>(*cpuInputs_[i]));
+      } else {
+        initArg(*cpuInputs_[i]);
+      }
       // TODO: Need a BufferCopy used to copy from one BufferArg to another.
       CpuVector cpuVector(cpuInputs_[i]->shape().getElements(),
                           (real*)cpuInputs_[i]->data());
@@ -217,7 +302,11 @@ protected:
         continue;  /// sparse matrix already init
       }
 
-      initArg(*cpuOutputs_[i]);
+      if (cpuOutputs_[i]->isSequenceArg()) {
+        initArg(dynamic_cast<SequenceArg&>(*cpuOutputs_[i]));
+      } else {
+        initArg(*cpuOutputs_[i]);
+      }
 
       // TODO: Need a BufferCopy used to copy from one BufferArg to another.
       CpuVector cpuVector(cpuOutputs_[i]->shape().getElements(),
@@ -241,28 +330,6 @@ protected:
     }
   }
 
-  // only init cpu argument, gpu argument copy from cpu argument.
-  void initArg(BufferArg& arg) {
-    CpuVector vector(arg.shape().getElements(), (real*)arg.data());
-    vector.uniform(0.001, 1);
-  }
-
-  void initArg(SequenceIdArg& arg, size_t batchSize) {
-    size_t numSeqs = arg.numSeqs();
-    int* buf = reinterpret_cast<int*>(arg.data());
-    int pos = 0;
-    size_t maxLen = 2 * batchSize / numSeqs;
-    for (int i = 0; i < (int)numSeqs; ++i) {
-      int len = uniformRandom(
-                    std::min<int64_t>(maxLen, batchSize - pos - numSeqs + i)) +
-                1;
-      buf[i] = pos;
-      pos += len;
-      VLOG(1) << " len=" << len;
-    }
-    buf[numSeqs] = batchSize;
-  }
-
 protected:
   std::shared_ptr<FunctionBase> cpuFunc_;
   std::shared_ptr<FunctionBase> gpuFunc_;
@@ -274,6 +341,8 @@ protected:
   std::vector<BufferArgPtr> gpuOutputs_;
   std::shared_ptr<CpuSparseMatrix> cpuSparse_;
   std::shared_ptr<GpuSparseMatrix> gpuSparse_;
+  std::shared_ptr<SequenceIdArg> cpuSeq_;
+  std::shared_ptr<SequenceIdArg> gpuSeq_;
 };
 
 }  // namespace paddle

paddle/function/MulOpTest.cpp

@@ -60,7 +60,7 @@ TEST(MulOp, DDDMatrixMul) {
             if (transa && transb) {
               continue;
             }
-            VLOG(3) << setiosflags(std::ios::left) << std::setfill(' ')
+            VLOG(3) << std::setiosflags(std::ios::left) << std::setfill(' ')
                     << " transa=" << transa << " transb=" << transb
                     << " dimM=" << std::setw(5) << dimM
                     << " dimN=" << std::setw(5) << dimN
@@ -104,7 +104,7 @@ TEST(MuLOp, DSparseDMul) {
       for (const auto dimK : {3, 10}) {
         for (const auto nnz : {3, 10}) {
           for (const auto FORMAT : {SPARSE_CSR}) {
-            VLOG(3) << setiosflags(std::ios::left) << std::setfill(' ')
+            VLOG(3) << std::setiosflags(std::ios::left) << std::setfill(' ')
                     << " dimM=" << std::setw(5) << dimM
                     << " dimN=" << std::setw(5) << dimN
                     << " dimK=" << std::setw(5) << dimK
@@ -150,7 +150,7 @@ TEST(MulOp, DDSparseMul) {
       for (const auto dimK : {3, 10}) {
         for (const auto nnz : {3, 10}) {
           for (const auto FORMAT : {SPARSE_CSR, SPARSE_CSC}) {
-            VLOG(3) << setiosflags(std::ios::left) << std::setfill(' ')
+            VLOG(3) << std::setiosflags(std::ios::left) << std::setfill(' ')
                     << " dimM=" << std::setw(5) << dimM
                     << " dimN=" << std::setw(5) << dimN
                     << " dimK=" << std::setw(5) << dimK
@@ -197,7 +197,7 @@ TEST(MulOp, SparseDDMul) {
       for (const auto dimK : {3, 10}) {
         for (const auto nnz : {3, 10}) {
           for (const auto FORMAT : {SPARSE_CSC, SPARSE_CSR}) {
-            VLOG(3) << setiosflags(std::ios::left) << std::setfill(' ')
+            VLOG(3) << std::setiosflags(std::ios::left) << std::setfill(' ')
                     << " dimM=" << std::setw(5) << dimM
                     << " dimN=" << std::setw(5) << dimN
                     << " dimK=" << std::setw(5) << dimK

paddle/gserver/dataproviders/PyDataProvider2.cpp

@@ -647,7 +647,7 @@ public:
       DataBatch& gpuBatch = *batch;
       std::vector<Argument>& gpuArguments = gpuBatch.getStreams();
       gpuArguments.resize(cpuArguments.size());
-      gpuBatch.setSize(size);
+      gpuBatch.setSize(bsize);
       for (size_t i = 0; i < headers_.size(); ++i) {
         gpuArguments[i].resizeAndCopyFrom(
             cpuArguments[i], useGpu_, HPPL_STREAM_1);

paddle/gserver/evaluators/CTCErrorEvaluator.cpp

@@ -20,7 +20,7 @@ namespace paddle {
 /**
  * calculate sequence-to-sequence edit distance
  */
-class CTCErrorEvaluator : public Evaluator {
+class CTCErrorEvaluator : public NotGetableEvaluator {
 private:
   MatrixPtr outActivations_;
   int numTimes_, numClasses_, numSequences_, blank_;

paddle/gserver/evaluators/Evaluator.cpp

@@ -13,9 +13,9 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/gserver/evaluators/Evaluator.h"
-#include "paddle/utils/Stat.h"
-
 #include "paddle/gserver/gradientmachines/NeuralNetwork.h"
+#include "paddle/utils/Stat.h"
+#include "paddle/utils/StringUtil.h"
 
 DECLARE_int32(trainer_id);
 
@@ -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();
@@ -102,6 +122,10 @@ public:
   virtual void distributeEval(ParameterClient2* client) {
     mergeResultsOfAllClients(client);
   }
+
+  // Evaluator interface
+protected:
+  std::string getTypeImpl() const { return "classification_error"; }
 };
 
 /**
@@ -140,6 +164,10 @@ public:
   virtual void distributeEval(ParameterClient2* client) {
     mergeResultsOfAllClients(client);
   }
+
+  // Evaluator interface
+protected:
+  std::string getTypeImpl() const { return "seq_classification_error"; }
 };
 REGISTER_EVALUATOR(seq_classification_error,
                    SequenceClassificationErrorEvaluator);
@@ -230,6 +258,10 @@ public:
 private:
   IVectorPtr cpuLabel_;
   MatrixPtr cpuWeight_;
+
+  // Evaluator interface
+protected:
+  std::string getTypeImpl() const { return "sum"; }
 };
 /**
  * @brief column sum Evaluator
@@ -337,10 +369,18 @@ public:
   }
 
 private:
-  ColumnSumEvaluator() {}
   int32_t colIdx_;
   size_t colNum_;
   MatrixPtr sum_; /* cpu matrix */
+
+  // Evaluator interface
+protected:
+  std::string getTypeImpl() const {
+    if (colIdx_ == -1)
+      return "last-column-sum";
+    else
+      return "column-sum";
+  }
 };
 
 void AucEvaluator::start() {
@@ -449,6 +489,16 @@ double AucEvaluator::calcAuc() const {
   }
 }
 
+real AucEvaluator::getValueImpl() const { return calcAuc(); }
+
+std::string AucEvaluator::getTypeImpl() const {
+  if (colIdx_ == -1) {
+    return "last-column-auc";
+  } else {
+    return "auc";
+  }
+}
+
 // class RankAucEvaluator
 REGISTER_EVALUATOR(rankauc, RankAucEvaluator);
 
@@ -528,12 +578,15 @@ double RankAucEvaluator::calcRankAuc(real* outputData,
                                         : aucTmp / (clickSum * noClickSum);
 }
 
+std::string RankAucEvaluator::getTypeImpl() const { return "rankauc"; }
+
 // class PrecisionRecallEvaluator
 REGISTER_EVALUATOR(precision_recall, PrecisionRecallEvaluator);
 
 void PrecisionRecallEvaluator::start() {
   Evaluator::start();
   statsInfo_.clear();
+  values_.clear();
 }
 
 real PrecisionRecallEvaluator::evalImp(std::vector<Argument>& arguments) {
@@ -594,52 +647,23 @@ real PrecisionRecallEvaluator::evalImp(std::vector<Argument>& arguments) {
 }
 
 void PrecisionRecallEvaluator::printStats(std::ostream& os) const {
-  int label = config_.positive_label();
-  if (label != -1) {
-    CHECK(label >= 0 && label < (int)statsInfo_.size())
-        << "positive_label [" << label << "] should be in range [0, "
-        << statsInfo_.size() << ")";
-    double precision =
-        calcPrecision(statsInfo_[label].TP, statsInfo_[label].FP);
-    double recall = calcRecall(statsInfo_[label].TP, statsInfo_[label].FN);
-    os << "positive_label=" << label << " precision=" << precision
-       << " recall=" << recall
-       << " F1-score=" << calcF1Score(precision, recall);
-    return;
-  }
-
-  // micro average method: precision = (TP1+TP2)/(TP1+FP1+TP2+FP2)
-  // macro average method: precision = (precision1+precision2)/2
-  double microTotalTP = 0;
-  double microTotalFP = 0;
-  double microTotalFN = 0;
-  double macroAvgPrecision = 0;
-  double macroAvgRecall = 0;
-  size_t numLabels = statsInfo_.size();
-  for (size_t i = 0; i < numLabels; ++i) {
-    microTotalTP += statsInfo_[i].TP;
-    microTotalFP += statsInfo_[i].FP;
-    microTotalFN += statsInfo_[i].FN;
-    macroAvgPrecision += calcPrecision(statsInfo_[i].TP, statsInfo_[i].FP);
-    macroAvgRecall += calcRecall(statsInfo_[i].TP, statsInfo_[i].FN);
-  }
-  macroAvgPrecision /= numLabels;
-  macroAvgRecall /= numLabels;
-  double macroAvgF1Score = calcF1Score(macroAvgPrecision, macroAvgRecall);
-  os << "macro-average-precision=" << macroAvgPrecision
-     << " macro-average-recall=" << macroAvgRecall
-     << " macro-average-F1-score=" << macroAvgF1Score;
-
-  double microAvgPrecision = calcPrecision(microTotalTP, microTotalFP);
-  double microAvgRecall = calcPrecision(microTotalTP, microTotalFN);
-  double microAvgF1Score = calcF1Score(microAvgPrecision, microAvgRecall);
-  if (!isMultiBinaryLabel_) {
-    // precision and recall are equal in this case
-    os << " micro-average-precision=" << microAvgPrecision;
-  } else {
-    os << " micro-average-precision=" << microAvgPrecision
-       << " micro-average-recall=" << microAvgRecall
-       << " micro-average-F1-score=" << microAvgF1Score;
+  PrintStatsInfo info;
+  bool containMacroMicroInfo = getStatsInfo(&info);
+  os << "positive_label=" << config_.positive_label()
+     << " precision=" << info.precision << " recall=" << info.recall
+     << " F1-score=" << info.f1;
+  if (containMacroMicroInfo) {
+    os << "macro-average-precision=" << info.macroAvgPrecision
+       << " macro-average-recall=" << info.macroAvgRecall
+       << " macro-average-F1-score=" << info.macroAvgF1Score;
+    if (!isMultiBinaryLabel_) {
+      // precision and recall are equal in this case
+      os << " micro-average-precision=" << info.microAvgPrecision;
+    } else {
+      os << " micro-average-precision=" << info.microAvgPrecision
+         << " micro-average-recall=" << info.microAvgRecall
+         << " micro-average-F1-score=" << info.microAvgF1Score;
+    }
   }
 }
 
@@ -721,6 +745,60 @@ void PrecisionRecallEvaluator::calcStatsInfoMulti(const MatrixPtr& output,
   }
 }
 
+void PrecisionRecallEvaluator::storeLocalValues() const {
+  if (this->values_.size() == 0) {
+    PrintStatsInfo info;
+    bool containMacroMicroInfo = getStatsInfo(&info);
+    values_["precision"] = info.precision;
+    values_["recal"] = info.recall;
+    values_["F1-score"] = info.f1;
+    if (containMacroMicroInfo) {
+      values_["macro-average-precision"] = info.macroAvgPrecision;
+      values_["macro-average-recall"] = info.macroAvgRecall;
+      values_["macro-average-F1-score"] = info.macroAvgF1Score;
+      if (!isMultiBinaryLabel_) {
+        // precision and recall are equal in this case
+        values_["micro-average-precision"] = info.microAvgPrecision;
+      } else {
+        values_["micro-average-precision"] = info.microAvgPrecision;
+        values_["micro-average-recall"] = info.microAvgRecall;
+        values_["micro-average-F1-score"] = info.microAvgF1Score;
+      }
+    }
+  }
+}
+
+void PrecisionRecallEvaluator::getNames(std::vector<std::string>* names) {
+  this->storeLocalValues();
+  names->reserve(this->values_.size());
+  for (auto it = this->values_.begin(); it != this->values_.end(); ++it) {
+    names->push_back(this->config_.name() + "." + it->first);
+  }
+}
+
+real PrecisionRecallEvaluator::getValue(const std::string& name,
+                                        Error* err) const {
+  this->storeLocalValues();
+  std::vector<std::string> buffers;
+  paddle::str::split(name, '.', &buffers);
+  auto it = this->values_.find(buffers[buffers.size() - 1]);
+  if (it == this->values_.end()) {  // not found
+    *err = Error("No such key %s", name.c_str());
+    return .0f;
+  }
+
+  return it->second;
+}
+
+std::string PrecisionRecallEvaluator::getType(const std::string& name,
+                                              Error* err) const {
+  this->getValue(name, err);
+  if (!err->isOK()) {
+    return "";
+  }
+  return "precision_recall";
+}
+
 void PrecisionRecallEvaluator::distributeEval(ParameterClient2* client) {
   size_t size = 4 * statsInfo_.size();
   double* buf = new double[size];
@@ -740,6 +818,47 @@ void PrecisionRecallEvaluator::distributeEval(ParameterClient2* client) {
   delete[] buf;
 }
 
+bool PrecisionRecallEvaluator::getStatsInfo(
+    PrecisionRecallEvaluator::PrintStatsInfo* info) const {
+  int label = config_.positive_label();
+  if (label != -1) {
+    CHECK(label >= 0 && label < (int)statsInfo_.size())
+        << "positive_label [" << label << "] should be in range [0, "
+        << statsInfo_.size() << ")";
+    info->precision = calcPrecision(statsInfo_[label].TP, statsInfo_[label].FP);
+    info->recall = calcRecall(statsInfo_[label].TP, statsInfo_[label].FN);
+    info->f1 = calcF1Score(info->precision, info->recall);
+    return false;
+  }
+
+  // micro average method: precision = (TP1+TP2)/(TP1+FP1+TP2+FP2)
+  // macro average method: precision = (precision1+precision2)/2
+  double microTotalTP = 0;
+  double microTotalFP = 0;
+  double microTotalFN = 0;
+  info->macroAvgPrecision = 0;
+  info->macroAvgRecall = 0;
+  size_t numLabels = statsInfo_.size();
+  for (size_t i = 0; i < numLabels; ++i) {
+    microTotalTP += statsInfo_[i].TP;
+    microTotalFP += statsInfo_[i].FP;
+    microTotalFN += statsInfo_[i].FN;
+    info->macroAvgPrecision +=
+        calcPrecision(statsInfo_[i].TP, statsInfo_[i].FP);
+    info->macroAvgRecall += calcRecall(statsInfo_[i].TP, statsInfo_[i].FN);
+  }
+  info->macroAvgPrecision /= numLabels;
+  info->macroAvgRecall /= numLabels;
+  info->macroAvgF1Score =
+      calcF1Score(info->macroAvgPrecision, info->macroAvgRecall);
+
+  info->microAvgPrecision = calcPrecision(microTotalTP, microTotalFP);
+  info->microAvgRecall = calcPrecision(microTotalTP, microTotalFN);
+  info->microAvgF1Score =
+      calcF1Score(info->microAvgPrecision, info->microAvgRecall);
+  return true;
+}
+
 REGISTER_EVALUATOR(pnpair, PnpairEvaluator);
 void PnpairEvaluator::start() {
   Evaluator::start();
@@ -864,56 +983,35 @@ void PnpairEvaluator::calc(std::vector<PredictionResult>& predictArray) {
             << " calc total special pair: " << special;
 }
 
+std::string PnpairEvaluator::getTypeImpl() const { return "pnpair"; }
+
 ClassRegistrar<Evaluator> Evaluator::registrar_;
 Evaluator* Evaluator::create(const EvaluatorConfig& config) {
-  Evaluator* evaluator = nullptr;
-  if (config.type() == "classification_error") {
-    evaluator = new ClassificationErrorEvaluator();
-  } else if (config.type() == "sum") {
-    evaluator = new SumEvaluator();
-  } else if (config.type() == "last-column-sum") {
-    evaluator = new ColumnSumEvaluator(-1);
-  } else if (config.type() == "last-column-auc") {
-    evaluator = new AucEvaluator(-1);
-  } else {
-    evaluator = registrar_.createByType(config.type());
-  }
+  Evaluator* evaluator = registrar_.createByType(config.type());
   evaluator->init(config);
   return evaluator;
 }
+
+REGISTER_EVALUATOR(classification_error, ClassificationErrorEvaluator);
+REGISTER_EVALUATOR(sum, SumEvaluator);
+static InitFunction __reg_type_auc_sum__([]() {
+  Evaluator::registrar_.registerClass(
+      "last-column-sum", [] { return new ColumnSumEvaluator(-1); });
+  Evaluator::registrar_.registerClass("last-column-auc",
+                                      [] { return new AucEvaluator(-1); });
+});
+
 /**
  * @brief print value of each layer.
  *
  * The config file api is value_printer_evaluator.
  */
-class ValuePrinter : public Evaluator {
+class ValuePrinter : public NotGetableEvaluator {
 public:
-  ValuePrinter() {}
-
   virtual void eval(const NeuralNetwork& nn) {
     for (const std::string& name : config_.input_layers()) {
-      const Argument& argu = nn.getLayer(name)->getOutput();
-      if (argu.value) {
-        std::ostringstream os;
-        argu.value->print(os);
-        LOG(INFO) << "layer=" << name << " value matrix:\n" << os.str();
-      }
-      if (argu.ids) {
-        std::ostringstream os;
-        argu.ids->print(os, argu.ids->getSize());
-        LOG(INFO) << "layer=" << name << " ids vector:\n" << os.str();
-      }
-      if (auto startPos = argu.sequenceStartPositions) {
-        std::ostringstream os;
-        startPos->getVector(false)->print(os, startPos->getSize());
-        LOG(INFO) << "layer=" << name << " sequence pos vector:\n" << os.str();
-      }
-      if (auto subStartPos = argu.subSequenceStartPositions) {
-        std::ostringstream os;
-        subStartPos->getVector(false)->print(os, subStartPos->getSize());
-        LOG(INFO) << "layer=" << name << " sub-sequence pos vector:\n"
-                  << os.str();
-      }
+      nn.getLayer(name)->getOutput().printValueString(LOG(INFO),
+                                                      "layer=" + name + " ");
     }
   }
 
@@ -922,15 +1020,14 @@ public:
   virtual real evalImp(std::vector<Argument>& arguments) { return 0; }
 };
 REGISTER_EVALUATOR(value_printer, ValuePrinter);
+
 /**
  * @brief print gradient of each layer.
  *
  * The config file api is gradient_printer_evaluator.
  */
-class GradientPrinter : public Evaluator {
+class GradientPrinter : public NotGetableEvaluator {
 public:
-  GradientPrinter() {}
-
   virtual void eval(const NeuralNetwork& nn) {
     for (const std::string& name : config_.input_layers()) {
       const Argument& argu = nn.getLayer(name)->getOutput();
@@ -939,11 +1036,6 @@ public:
         argu.grad->print(os);
         LOG(INFO) << "layer=" << name << " grad matrix:\n" << os.str();
       }
-      if (auto startPos = argu.sequenceStartPositions) {
-        std::ostringstream os;
-        startPos->getVector(false)->print(os, startPos->getSize());
-        LOG(INFO) << "layer=" << name << " sequence pos vector:\n" << os.str();
-      }
     }
   }
 
@@ -957,7 +1049,7 @@ REGISTER_EVALUATOR(gradient_printer, GradientPrinter);
  *
  * The config file api is maxid_printer_evaluator.
  */
-class MaxIdPrinter : public Evaluator {
+class MaxIdPrinter : public NotGetableEvaluator {
 private:
   IVectorPtr maxIds_;
   MatrixPtr maxValues_;
@@ -999,7 +1091,7 @@ REGISTER_EVALUATOR(max_id_printer, MaxIdPrinter);
  *
  * The config file api is maxframe_printer_evaluator.
  */
-class MaxFramePrinter : public Evaluator {
+class MaxFramePrinter : public NotGetableEvaluator {
 private:
   IVectorPtr maxIds_;
   MatrixPtr maxValues_;
@@ -1086,7 +1178,7 @@ REGISTER_EVALUATOR(max_frame_printer, MaxFramePrinter);
  * The config file api is seqtext_printer_evaluator.
  *
  */
-class SequenceTextPrinter : public Evaluator {
+class SequenceTextPrinter : public NotGetableEvaluator {
 private:
   /// dict_file, which contains a list of tokens
   std::vector<std::string> dict_;
@@ -1253,4 +1345,6 @@ public:
 };
 REGISTER_EVALUATOR(classification_error_printer, ClassificationErrorPrinter);
 
+std::string DummyEvaluator::getTypeImpl() const { return "dummy"; }
+
 }  // namespace paddle

paddle/gserver/evaluators/Evaluator.h

@@ -19,6 +19,7 @@ limitations under the License. */
 #include "paddle/parameter/Argument.h"
 #include "paddle/pserver/ParameterClient2.h"
 #include "paddle/utils/ClassRegistrar.h"
+#include "paddle/utils/Error.h"
 
 namespace paddle {
 
@@ -117,12 +118,105 @@ public:
 
   static ClassRegistrar<Evaluator> registrar_;
 
+  /**
+   * @brief getNames will return all field names of current evaluator.
+   *
+   * The format of name is `evaluator_name.evaluator_fields`. If the evaluator
+   * has multiple field, the name could be `evaluator_name.field1`. For example
+   * the PrecisionRecallEvaluator contains `precision`, `recall` fields. The get
+   * names will return `precision_recall_evaluator.precision`,
+   * `precision_recall_evaluator.recal`, etc.
+   *
+   * Also, if current Evaluator is a combined evaluator. getNames will return
+   * all names of all evaluators inside the combined evaluator.
+   *
+   * @param names [out]: the field names of current evaluator.
+   * @note Never clear the names parameter inside getNames.
+   */
+  virtual void getNames(std::vector<std::string>* names) {
+    names->push_back(config_.name());
+  }
+
+  /**
+   * @brief getValue will return the current evaluate value of one field.
+   *
+   * @param name: The field name of current evaluator.
+   * @param err [out]: The error state.
+   *
+   * @return The evaluate value(metric).
+   */
+  virtual real getValue(const std::string& name, Error* err) const {
+    if (name != config_.name()) {
+      *err = Error("no such name of evaluator %s", name.c_str());
+      return .0f;
+    }
+    return this->getValueImpl();
+  }
+
+  /**
+   * @brief getType will return the evaluator type by field name.
+   *
+   * Evaluate Type is the current type of evaluator in string. Such as 'auc',
+   * 'precision_recall'. In combined evaluator, different name may get different
+   * evaluate type because it could be evaluated by different evaluator inside.
+   *
+   * @param name: The field name of current Evaluator.
+   * @param err: The error state. nullptr means don't care.
+   * @return the evaluator type string.
+   */
+  virtual std::string getType(const std::string& name, Error* err) const {
+    if (name != config_.name()) {
+      *err = Error("no such name of evaluator %s", name.c_str());
+      return std::string();
+    }
+    return this->getTypeImpl();
+  }
+
+protected:
+  /**
+   * @brief getValueImpl The simplest way to define getValue result. If this
+   * evaluator doesn't contain multiple fields, and do not throw any error, just
+   * implemented this method to get the evaluate result(metric).
+   * @return Evaluate result(metric).
+   */
+  virtual real getValueImpl() const {
+    return numSamples_ != .0 ? totalScore_ / numSamples_ : .0;
+  }
+
+  /**
+   * @brief getTypeImpl The simplest way to define getType result. If this
+   * evaluator doesn't combine many evaluators, the get type should only return
+   * itself type.
+   * @return Evaluator type.
+   */
+  virtual std::string getTypeImpl() const { return "base"; }
+
 protected:
   EvaluatorConfig config_;
   double numSamples_;
   double totalScore_;
 };
 
+/**
+ * @brief The NotGetableEvaluator class is the base class of evaluator that
+ * cannot get value in runtime. The most NotGetableEvaluator is Printer
+ * Evaluator, which is only used to debug network configuration.
+ */
+class NotGetableEvaluator : public Evaluator {
+  // Evaluator interface
+public:
+  void getNames(std::vector<std::string>* names) {}
+
+  real getValue(const std::string& name, Error* err) const {
+    *err = Error("Not implemented");
+    return .0f;
+  }
+  std::string getType(const std::string& name, Error* err) const {
+    *err = Error("Not implemented");
+    return "";
+  }
+};
+
 class DummyEvaluator : public Evaluator {
 public:
   DummyEvaluator() {}
@@ -135,6 +229,10 @@ public:
   }
   virtual void finish() {}
   virtual void printStats(std::ostream&) const {}
+
+  // Evaluator interface
+protected:
+  std::string getTypeImpl() const;
 };
 /**
  * @brief evaluate AUC using colIdx-th column as prediction.
@@ -191,6 +289,11 @@ private:
   }
 
   double calcAuc() const;
+
+  // Evaluator interface
+protected:
+  real getValueImpl() const;
+  std::string getTypeImpl() const;
 };
 
 /**
@@ -223,6 +326,10 @@ private:
                      real* clickData,
                      real* pvData,
                      size_t size);
+
+  // Evaluator interface
+protected:
+  std::string getTypeImpl() const;
 };
 /**
  * @brief precision, recall and f1 score Evaluator
@@ -272,6 +379,20 @@ private:
   IVectorPtr cpuLabel_;
   MatrixPtr cpuWeight_;
 
+  struct PrintStatsInfo {
+    double precision;
+    double recall;
+    double f1;
+    double macroAvgPrecision;
+    double macroAvgRecall;
+    double macroAvgF1Score;
+    double microAvgPrecision;
+    double microAvgRecall;
+    double microAvgF1Score;
+  };
+
+  bool getStatsInfo(PrintStatsInfo* info) const;
+
   void calcStatsInfo(const MatrixPtr& output,
                      const IVectorPtr& label,
                      const MatrixPtr& weight);
@@ -303,6 +424,15 @@ private:
       return 0;
     }
   }
+
+  mutable std::unordered_map<std::string, real> values_;
+
+  void storeLocalValues() const;
+  // Evaluator interface
+public:
+  void getNames(std::vector<std::string>* names);
+  real getValue(const std::string& name, Error* err) const;
+  std::string getType(const std::string& name, Error* err) const;
 };
 
 /*
@@ -349,8 +479,7 @@ public:
   virtual void finish() { calc(predictArray_); }
 
   virtual void printStats(std::ostream& os) const {
-    os << " pos/neg"
-       << "=" << pairArray_[0] / ((pairArray_[1] <= 0) ? 1.0 : pairArray_[1]);
+    os << " pos/neg=" << this->getValueImpl();
   }
 
   virtual void distributeEval(ParameterClient2* client) {
@@ -366,6 +495,13 @@ private:
   IVectorPtr cpuLabel_;
   IVectorPtr cpuInfo_;
   MatrixPtr cpuWeight_;
+
+  // Evaluator interface
+protected:
+  real getValueImpl() const {
+    return pairArray_[0] / ((pairArray_[1] <= 0) ? 1.0 : pairArray_[1]);
+  }
+  std::string getTypeImpl() const;
 };
 
 }  // namespace paddle

paddle/gserver/gradientmachines/NeuralNetwork.cpp

@@ -306,7 +306,6 @@ void NeuralNetwork::onPassEnd() {
 
 class CombinedEvaluator : public Evaluator {
 public:
-  CombinedEvaluator() {}
   void addEvaluator(std::unique_ptr<Evaluator>&& evaluator) {
     evaluators_.emplace_back(std::move(evaluator));
   }
@@ -346,6 +345,55 @@ public:
 
 protected:
   std::vector<std::unique_ptr<Evaluator>> evaluators_;
+
+  // Evaluator interface
+public:
+  /**
+   * @brief getNames will return all inside evaluators' names.
+   * @param names [out]: return names.
+   */
+  void getNames(std::vector<std::string>* names) {
+    for (auto& eval : evaluators_) {
+      eval->getNames(names);
+    }
+  }
+
+  /**
+   * @brief getValue could get all inside evaluators' value.
+   */
+  real getValue(const std::string& name, Error* err) const {
+    return this->getMethodHelper<real>(
+        name, err, [&name, err](const std::unique_ptr<Evaluator>& eval) {
+          return eval->getValue(name, err);
+        });
+  }
+
+  /**
+   * @brief getType could get all inside evaluators' type.
+   */
+  std::string getType(const std::string& name, Error* err) const {
+    return this->getMethodHelper<std::string>(
+        name, err, [&name, err](const std::unique_ptr<Evaluator>& eval) {
+          return eval->getType(name, err);
+        });
+  }
+
+private:
+  template <typename T>
+  T getMethodHelper(const std::string& name,
+                    Error* err,
+                    const std::function<T(const std::unique_ptr<Evaluator>&)>&
+                        callback) const {
+    for (auto& eval : evaluators_) {
+      std::vector<std::string> names;
+      eval->getNames(&names);
+      if (std::find(names.begin(), names.end(), name) != names.end()) {
+        return callback(eval);
+      }
+    }
+    *err = Error("No such key %s", name.c_str());
+    return T();
+  }
 };
 
 Evaluator* NeuralNetwork::makeEvaluator() const {

paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp

@@ -155,7 +155,8 @@ protected:
 public:
   explicit BootBiasLayer(const LayerConfig& config) : Layer(config) {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap) {
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override {
     if (!Layer::init(layerMap, parameterMap)) return false;
 
     if (biasParameter_) {
@@ -174,7 +175,7 @@ public:
     }
   }
 
-  virtual void forward(PassType passType) {
+  void forward(PassType passType) override {
     if (biases_) {
       MatrixPtr outV = getOutputValue();
       outV->addBias(*(biases_->getW()), 1);
@@ -182,7 +183,7 @@ public:
     }
   }
 
-  virtual void backward(const UpdateCallback& callback) {
+  void backward(const UpdateCallback& callback) override {
     if (biases_) {
       backwardActivation();
       biases_->getWGrad()->collectBias(*getOutputGrad(), 1);

paddle/gserver/layers/AddtoLayer.h

@@ -44,19 +44,20 @@ public:
   /**
    * Intialization of AddtoLayer.
    */
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
   /**
    * Forward propagation.
    * @note There is no weight matrix for each input,
    *       because it just a simple add operation.
    */
-  void forward(PassType passType);
+  void forward(PassType passType) override;
 
   /**
    * Backward propagation.
    */
-  void backward(const UpdateCallback& callback = nullptr);
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 }  // namespace paddle

paddle/gserver/layers/AgentLayer.h

@@ -35,7 +35,8 @@ public:
 
   ~AgentLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
   // if *numSamples* set,
   // real layer output will only use first *numSamples* rows
@@ -44,8 +45,8 @@ public:
     numSamples_ = numSamples;
   }
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr) {}
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override {}
 };
 
 /**
@@ -56,8 +57,8 @@ public:
   explicit SequenceAgentLayer(const LayerConfig& config) : AgentLayer(config) {}
   ~SequenceAgentLayer() {}
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr) {}
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override {}
 };
 
 /**
@@ -78,7 +79,8 @@ public:
 
   virtual ~GatherAgentLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
   // call before addRealLayer
   void copyIdAndSequenceInfo(const Argument& input,
@@ -88,8 +90,8 @@ public:
   // add one real layer, can call many times
   void addRealLayer(LayerPtr layer) { realLayers_.push_back(layer); }
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback) override;
 };
 
 /**
@@ -133,7 +135,8 @@ public:
 
   virtual ~ScatterAgentLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
   /**
    * @brief set real layer in generation
@@ -182,8 +185,8 @@ public:
     numSequences_ = numSequences;
   }
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback) override;
 };
 
 /**

paddle/gserver/layers/AverageLayer.h

@@ -38,12 +38,11 @@ public:
   explicit AverageLayer(const LayerConfig& config)
       : SequencePoolLayer(config) {}
 
-  ~AverageLayer() {}
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
-
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 
 protected:
   MatrixPtr outMtx_;

paddle/gserver/layers/BatchNormBaseLayer.h

@@ -52,7 +52,8 @@ public:
    */
   static Layer* create(const LayerConfig& config);
 
-  virtual bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
   /**
    * @brief Calculate feature map size. Some input uses frameHeight and

paddle/gserver/layers/BatchNormalizationLayer.h

@@ -33,9 +33,10 @@ public:
 
   ~BatchNormalizationLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 
 protected:
   /// Epsilon value used in the batch normalization formula.
@@ -58,7 +59,7 @@ protected:
   /// to batch, channels* imagePixels.
   void shrinkMat(const MatrixPtr& in, MatrixPtr& out);
 
-  void onPassEnd() { firstTest_ = true; }
+  void onPassEnd() override { firstTest_ = true; }
 
   MatrixPtr tmpMat_, tmpGrad_;
   MatrixPtr expandedIn_, expandedOut_;

paddle/gserver/layers/BilinearInterpLayer.h

@@ -38,9 +38,10 @@ public:
   virtual ~BilinearInterpLayer() {}
 
   size_t getSize();
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 }  // namespace paddle

paddle/gserver/layers/BlockExpandLayer.h

@@ -58,10 +58,11 @@ public:
 
   ~BlockExpandLayer() {}
 
-  virtual bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  virtual void forward(PassType passType);
-  virtual void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 }  // namespace paddle

paddle/gserver/layers/CRFDecodingLayer.h

@@ -32,9 +32,10 @@ namespace paddle {
 class CRFDecodingLayer : public CRFLayer {
 public:
   explicit CRFDecodingLayer(const LayerConfig& config) : CRFLayer(config) {}
-  virtual bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
-  virtual void forward(PassType passType);
-  virtual void backward(const UpdateCallback& callback);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback) override;
 
 protected:
   std::unique_ptr<LinearChainCRF> crf_;

paddle/gserver/layers/CRFLayer.h

@@ -29,9 +29,10 @@ namespace paddle {
 class CRFLayer : public Layer {
 public:
   explicit CRFLayer(const LayerConfig& config) : Layer(config) {}
-  virtual bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
-  virtual void forward(PassType passType);
-  virtual void backward(const UpdateCallback& callback);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback) override;
 
 protected:
   size_t numClasses_;

paddle/gserver/layers/CTCLayer.h

@@ -22,10 +22,11 @@ namespace paddle {
 class CTCLayer : public Layer {
 public:
   explicit CTCLayer(const LayerConfig& config) : Layer(config) {}
-  virtual bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
-  virtual void forward(PassType passType);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+  void forward(PassType passType) override;
   void forwardImp(const Argument& softmaxSeqs, const Argument& labelSeqs);
-  virtual void backward(const UpdateCallback& callback);
+  void backward(const UpdateCallback& callback) override;
   void backwardImp(const UpdateCallback& callback,
                    const Argument& softmaxSeqs,
                    const Argument& labelSeqs);

paddle/gserver/layers/ConcatenateLayer.cpp

@@ -28,10 +28,11 @@ public:
 
   ~ConcatenateLayer() {}
 
-  virtual bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  virtual void forward(PassType passType);
-  virtual void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 REGISTER_LAYER(concat, ConcatenateLayer);
@@ -101,10 +102,11 @@ public:
 
   ~ConcatenateLayer2() {}
 
-  virtual bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  virtual void forward(PassType passType);
-  virtual void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 
 protected:
   std::vector<std::unique_ptr<Projection>> projections_;

paddle/gserver/layers/ConvBaseLayer.h

@@ -80,7 +80,8 @@ protected:
 public:
   explicit ConvBaseLayer(const LayerConfig& config) : Layer(config) {}
 
-  virtual bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
   /**
    * imgSizeH_ and imgSizeW_ will be set according to the previous input layers

paddle/gserver/layers/ConvShiftLayer.cpp

@@ -47,10 +47,11 @@ public:
 
   ~ConvShiftLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 REGISTER_LAYER(conv_shift, ConvShiftLayer);

paddle/gserver/layers/ConvexCombinationLayer.cpp

@@ -49,10 +49,11 @@ public:
 
   ~ConvexCombinationLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 REGISTER_LAYER(convex_comb, ConvexCombinationLayer);

paddle/gserver/layers/CosSimLayer.cpp

@@ -26,15 +26,23 @@ bool CosSimLayer::init(const LayerMap& layerMap,
   Layer::init(layerMap, parameterMap);
 
   CHECK_EQ(inputLayers_.size(), 2LU);
+
+  createFunction(forward_,
+                 "CosSimForward",
+                 FuncConfig().set("scale", (real)config_.cos_scale()));
+  createFunction(backward_,
+                 "CosSimBackward",
+                 FuncConfig().set("scale", (real)config_.cos_scale()));
+
   return true;
 }
 
 void CosSimLayer::forward(PassType passType) {
   Layer::forward(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";
 
   {
     REGISTER_TIMER_INFO("CosFwResetTimer", getName().c_str());
@@ -42,26 +50,43 @@ void CosSimLayer::forward(PassType passType) {
   }
 
   MatrixPtr outV = getOutputValue();
-
   /* activation */ {
     REGISTER_TIMER_INFO("CosFwAtvTimer", getName().c_str());
     MatrixPtr prevOut1 = getInputValue(0);
     MatrixPtr prevOut2 = getInputValue(1);
-    outV->cosSim(*prevOut1, *prevOut2, config_.cos_scale());
+
+    CHECK(outV && prevOut1 && prevOut2);
+    BufferArgs inputs;
+    BufferArgs outputs;
+    inputs.addArg(*prevOut1);
+    inputs.addArg(*prevOut2);
+    outputs.addArg(*outV, ASSIGN_TO);
+    forward_[0]->calc(inputs, outputs);
   }
 }
 
 void CosSimLayer::backward(const UpdateCallback& callback) {
   /* activation */ {
     REGISTER_TIMER_INFO("CosBpAtvTimer", getName().c_str());
-    MatrixPtr outG = this->getOutputGrad();
-
-    outG->cosSimDerivative(*this->getOutputValue(),
-                           *getInputValue(0),
-                           *getInputValue(1),
-                           *getInputGrad(0),
-                           *getInputGrad(1),
-                           config_.cos_scale());
+    CHECK_EQ(backward_.size(), 1) << "Only one backward function needed";
+
+    const auto outG = this->getOutputGrad();
+    const auto outV = this->getOutputValue();
+    const auto inV1 = this->getInputValue(0);
+    const auto inV2 = this->getInputValue(1);
+    auto inG1 = this->getInputGrad(0);
+    auto inG2 = this->getInputGrad(1);
+    CHECK(outG && outV && inV1 && inV2 && inG1 && inG2);
+    BufferArgs inputs;
+    BufferArgs outputs;
+    inputs.addArg(*outG);
+    inputs.addArg(*outV);
+    inputs.addArg(*inV1);
+    inputs.addArg(*inV2);
+    outputs.addArg(*inG1, ADD_TO);
+    outputs.addArg(*inG2, ADD_TO);
+
+    backward_[0]->calc(inputs, outputs);
   }
 }
 

paddle/gserver/layers/CosSimLayer.h

@@ -28,7 +28,7 @@ namespace paddle {
  *
  * - Input1: A vector (batchSize * dataDim) *
  * - Input2: A vector (batchSize * dataDim) or (1 * dataDim) *
- * - Output: A vector (dataDim * 1)
+ * - Output: A vector (batchSize * 1)
  *
  * The config file api is cos_sim.
  */
@@ -38,10 +38,11 @@ public:
 
   ~CosSimLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 }  // namespace paddle

paddle/gserver/layers/CosSimVecMatLayer.cpp

@@ -18,7 +18,6 @@ limitations under the License. */
 #include "paddle/utils/Stat.h"
 
 namespace paddle {
-
 /**
  * @brief A layer for computing cosine similarity between a vector
  * and each row of a matrix
@@ -46,10 +45,11 @@ public:
 
   ~CosSimVecMatLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 REGISTER_LAYER(cos_vm, CosSimVecMatLayer);
@@ -97,11 +97,22 @@ bool CosSimVecMatLayer::init(const LayerMap& layerMap,
                            dataDim,
                            /* trans= */ false,
                            useGpu_);
+
+  CHECK(tmpRow0 && tmpRow1 && tmpRow2 && tmpRow3 && tmpMtx0 && tmpMtx1);
+
+  createFunction(forward_,
+                 "CosSimForward",
+                 FuncConfig().set("scale", (real)config_.cos_scale()));
+  createFunction(backward_,
+                 "CosSimBackward",
+                 FuncConfig().set("scale", (real)config_.cos_scale()));
+
   return true;
 }
 
 void CosSimVecMatLayer::forward(PassType passType) {
   Layer::forward(passType);
+  CHECK_EQ(forward_.size(), 1) << "Only one forward function needed";
 
   MatrixPtr inV0 = getInputValue(0);
   MatrixPtr inV1 = getInputValue(1);
@@ -117,17 +128,25 @@ void CosSimVecMatLayer::forward(PassType passType) {
   }
 
   MatrixPtr outV = getOutputValue();
-
+  CHECK(outV && inV0 && inV1);
   REGISTER_TIMER_INFO("FwCosVMTimer", getName().c_str());
   for (size_t i = 0; i < batchSize; i++) {
     tmpRow0->setData(inV0->rowBuf(i));
     tmpMtx0->setData(inV1->rowBuf(i));
     tmpRow2->setData(outV->rowBuf(i));
-    tmpRow2->cosSim(*(tmpMtx0), *(tmpRow0), config_.cos_scale());
+
+    BufferArgs inputs;
+    BufferArgs outputs;
+    inputs.addArg(*tmpMtx0);
+    inputs.addArg(*tmpRow0);
+    outputs.addArg(*tmpRow2, ASSIGN_TO);
+    forward_[0]->calc(inputs, outputs);
   }
 }
 
 void CosSimVecMatLayer::backward(const UpdateCallback& callback) {
+  CHECK_EQ(backward_.size(), 1) << "Only one forward function needed";
+
   MatrixPtr inV0 = getInputValue(0);
   MatrixPtr inV1 = getInputValue(1);
   MatrixPtr inG0 = getInputGrad(0);
@@ -136,27 +155,27 @@ void CosSimVecMatLayer::backward(const UpdateCallback& callback) {
   MatrixPtr outG = getOutputGrad();
 
   size_t batchSize = inV0->getHeight();
-
+  CHECK(inV0 && inV1 && inG0 && inG1 && outV && outG);
   REGISTER_TIMER_INFO("BwCosVMTimer", getName().c_str());
 
-  if (inG0 && inG1) {
-    for (size_t i = 0; i < batchSize; i++) {
-      tmpRow0->setData(inV0->rowBuf(i));
-      tmpRow1->setData(inG0->rowBuf(i));
-      tmpMtx0->setData(inV1->rowBuf(i));
-      tmpMtx1->setData(inG1->rowBuf(i));
-      tmpRow2->setData(outV->rowBuf(i));
-      tmpRow3->setData(outG->rowBuf(i));
-
-      tmpRow3->cosSimDerivative(*(tmpRow2),
-                                *(tmpMtx0),
-                                *(tmpRow0),
-                                *(tmpMtx1),
-                                *(tmpRow1),
-                                config_.cos_scale());
-    }
-  } else {
-    CHECK(!inG0 || !inG1) << "Not supported";
+  for (size_t i = 0; i < batchSize; i++) {
+    tmpRow0->setData(inV0->rowBuf(i));
+    tmpRow1->setData(inG0->rowBuf(i));
+    tmpMtx0->setData(inV1->rowBuf(i));
+    tmpMtx1->setData(inG1->rowBuf(i));
+    tmpRow2->setData(outV->rowBuf(i));
+    tmpRow3->setData(outG->rowBuf(i));
+
+    BufferArgs inputs;
+    BufferArgs outputs;
+    inputs.addArg(*tmpRow3);
+    inputs.addArg(*tmpRow2);
+    inputs.addArg(*tmpMtx0);
+    inputs.addArg(*tmpRow0);
+    outputs.addArg(*tmpMtx1, ADD_TO);
+    outputs.addArg(*tmpRow1, ADD_TO);
+
+    backward_[0]->calc(inputs, outputs);
   }
 }
 

paddle/gserver/layers/CostLayer.cpp

@@ -367,8 +367,6 @@ void LambdaCost::backward(const UpdateCallback& callback) {
   getInputGrad(0)->add(*marginGrad_);
 }
 
-void LambdaCost::onPassEnd() {}
-
 void LambdaCost::calcGrad(const real* outputScore,
                           const real* score,
                           real* gradData,
@@ -611,14 +609,15 @@ class SumCostLayer : public Layer {
 public:
   explicit SumCostLayer(const LayerConfig& config) : Layer(config) {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap) {
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override {
     bool ret = Layer::init(layerMap, parameterMap);
     if (!ret) return ret;
     CHECK_EQ(inputLayers_.size(), 1UL);
     return true;
   }
 
-  virtual void forward(PassType passType) {
+  void forward(PassType passType) override {
     Layer::forward(passType);
     const MatrixPtr& input = getInputValue(0);
 
@@ -629,7 +628,7 @@ public:
     output_.value->sumRows(*input, /* scaleSum= */ 1, /* scaleDest= */ 0);
   }
 
-  virtual void backward(const UpdateCallback& callback = nullptr) {
+  void backward(const UpdateCallback& callback = nullptr) override {
     getInputGrad(0)->add((real)1);
   }
 };

paddle/gserver/layers/CostLayer.h

@@ -32,15 +32,16 @@ class CostLayer : public Layer {
 public:
   explicit CostLayer(const LayerConfig& config) : Layer(config) {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
   LayerPtr getOutputLayer() { return inputLayers_[0]; }
 
   LayerPtr getLabelLayer() { return inputLayers_[1]; }
 
-  virtual void forward(PassType passType);
+  void forward(PassType passType) override;
 
-  virtual void backward(const UpdateCallback& callback = nullptr);
+  void backward(const UpdateCallback& callback = nullptr) override;
 
   virtual void forwardImp(Matrix& outputValue,
                           Argument& label,
@@ -68,11 +69,14 @@ public:
   explicit MultiClassCrossEntropy(const LayerConfig& config)
       : CostLayer(config) {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forwardImp(Matrix& output, Argument& label, Matrix& cost);
+  void forwardImp(Matrix& output, Argument& label, Matrix& cost) override;
 
-  void backwardImp(Matrix& outputValue, Argument& label, Matrix& outputGrad);
+  void backwardImp(Matrix& outputValue,
+                   Argument& label,
+                   Matrix& outputGrad) override;
 };
 
 /**
@@ -95,11 +99,14 @@ public:
   explicit MultiClassCrossEntropyWithSelfNorm(const LayerConfig& config)
       : CostLayer(config) {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forwardImp(Matrix& output, Argument& label, Matrix& cost);
+  void forwardImp(Matrix& output, Argument& label, Matrix& cost) override;
 
-  void backwardImp(Matrix& outputValue, Argument& label, Matrix& outputGrad);
+  void backwardImp(Matrix& outputValue,
+                   Argument& label,
+                   Matrix& outputGrad) override;
 
 protected:
   MatrixPtr sftMaxSum_;
@@ -117,11 +124,14 @@ public:
   explicit SoftBinaryClassCrossEntropy(const LayerConfig& config)
       : CostLayer(config) {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forwardImp(Matrix& output, Argument& label, Matrix& cost);
+  void forwardImp(Matrix& output, Argument& label, Matrix& cost) override;
 
-  void backwardImp(Matrix& outputValue, Argument& label, Matrix& outputGrad);
+  void backwardImp(Matrix& outputValue,
+                   Argument& label,
+                   Matrix& outputGrad) override;
 
 protected:
   MatrixPtr targetPerDim_;
@@ -139,11 +149,14 @@ public:
   explicit SumOfSquaresCostLayer(const LayerConfig& config)
       : CostLayer(config) {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forwardImp(Matrix& output, Argument& label, Matrix& cost);
+  void forwardImp(Matrix& output, Argument& label, Matrix& cost) override;
 
-  void backwardImp(Matrix& outputValue, Argument& label, Matrix& outputGrad);
+  void backwardImp(Matrix& outputValue,
+                   Argument& label,
+                   Matrix& outputGrad) override;
 };
 
 /**
@@ -162,17 +175,18 @@ class RankingCost : public Layer {
 public:
   explicit RankingCost(const LayerConfig& config) : Layer(config) {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
   LayerPtr getOutputLayer(size_t i) { return inputLayers_[i]; }
 
   LayerPtr getLabelLayer() { return inputLayers_[2]; }
 
-  void forward(PassType passType);
+  void forward(PassType passType) override;
 
-  void backward(const UpdateCallback& callback = nullptr);
+  void backward(const UpdateCallback& callback = nullptr) override;
 
-  void onPassEnd();
+  void onPassEnd() override;
 
   void forwardImp(Matrix& output, Argument& label, Matrix& cost) {
     (void)output;
@@ -214,17 +228,16 @@ class LambdaCost : public Layer {
 public:
   explicit LambdaCost(const LayerConfig& config) : Layer(config) {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
   LayerPtr getOutputLayer() { return inputLayers_[0]; }
 
   LayerPtr getScoreLayer() { return inputLayers_[1]; }
 
-  void forward(PassType passType);
+  void forward(PassType passType) override;
 
-  void backward(const UpdateCallback& callback = nullptr);
-
-  void onPassEnd();
+  void backward(const UpdateCallback& callback = nullptr) override;
 
   real calcNDCG(const real* outputScore, const real* score, int size);
   void calcGrad(const real* outputScore,
@@ -256,11 +269,14 @@ public:
   explicit MultiBinaryLabelCrossEntropy(const LayerConfig& config)
       : CostLayer(config) {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forwardImp(Matrix& output, Argument& label, Matrix& cost);
+  void forwardImp(Matrix& output, Argument& label, Matrix& cost) override;
 
-  void backwardImp(Matrix& outputValue, Argument& label, Matrix& outputGrad);
+  void backwardImp(Matrix& outputValue,
+                   Argument& label,
+                   Matrix& outputGrad) override;
 };
 
 /**
@@ -282,13 +298,16 @@ class HuberTwoClass : public CostLayer {
 public:
   explicit HuberTwoClass(const LayerConfig& config) : CostLayer(config) {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forwardImp(Matrix& output, Argument& label, Matrix& cost);
+  void forwardImp(Matrix& output, Argument& label, Matrix& cost) override;
 
   void forwardImpIn(Matrix& output, Argument& label, Matrix& cost);
 
-  void backwardImp(Matrix& outputValue, Argument& label, Matrix& outputGrad);
+  void backwardImp(Matrix& outputValue,
+                   Argument& label,
+                   Matrix& outputGrad) override;
 
   void backwardImpIn(Matrix& outputValue, Argument& label, Matrix& outputGrad);
 };

paddle/gserver/layers/CudnnBatchNormLayer.h

@@ -35,14 +35,15 @@ public:
 
   ~CudnnBatchNormLayer();
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
   /**
    * reshape tensor of ioDesc_.
    */
   void reshape(int batchSize);
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 
 protected:
   /**

paddle/gserver/layers/CudnnConvLayer.h

@@ -45,9 +45,10 @@ public:
 
   ~CudnnConvLayer();
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback) override;
   void addBiases();
   void bpropBiases();
 };

paddle/gserver/layers/CudnnPoolLayer.h

@@ -45,7 +45,8 @@ public:
                         hl_pooling_mode_t* mode = nullptr);
   explicit CudnnPoolLayer(const LayerConfig& config);
   ~CudnnPoolLayer();
-  virtual bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
   /**
    * Reshape input and output tensor descriptor.
@@ -53,8 +54,8 @@ public:
    * So reshaping is needed.
    */
   void reshape(int batchSize);
-  virtual void forward(PassType passType);
-  virtual void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 }  // namespace paddle

paddle/gserver/layers/DataLayer.h

@@ -33,13 +33,13 @@ public:
   /**
    * Prefetch sparse matrix/ids only.
    */
-  void prefetch() { output_ = data_; }
+  void prefetch() override { output_ = data_; }
 
   /**
    * Forward propagation. Copy data_ (value, in, grad, ids, cpuSequenceDims,
    * sequenceStartPositions, subSequenceStartPositions, strs) to output_.
    */
-  virtual void forward(PassType passType) {
+  void forward(PassType passType) override {
     Layer::forward(passType);
     copyDataToOutput(output_);
     if (FLAGS_show_layer_stat) {
@@ -50,9 +50,9 @@ public:
   /**
    * Data layer's backward propagation do nothing.
    */
-  virtual void backward(const UpdateCallback& callback) { (void)callback; }
+  void backward(const UpdateCallback& callback) override { (void)callback; }
 
-  virtual void copyOutputToOtherDevice() {
+  void copyOutputToOtherDevice() override {
     for (size_t i = 0; i != outputOtherDevice_.size(); i++) {
       copyDataToOutput(outputOtherDevice_[i]);
     }

paddle/gserver/layers/DataNormLayer.h

@@ -44,10 +44,11 @@ public:
 
   ~DataNormLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 
 protected:
   int mode_;

paddle/gserver/layers/EosIdCheckLayer.cpp

@@ -27,14 +27,14 @@ class EosIdCheckLayer : public Layer {
 public:
   explicit EosIdCheckLayer(const LayerConfig& config) : Layer(config) {}
 
-  virtual bool init(const LayerMap& layerMap,
-                    const ParameterMap& parameterMap) {
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override {
     bool ret = Layer::init(layerMap, parameterMap);
     CHECK_EQ(1UL, inputLayers_.size());
     return ret;
   }
 
-  virtual void forward(PassType passType) {
+  void forward(PassType passType) override {
     Layer::forward(passType);
 
     const Argument& input = getInput(0);
@@ -42,7 +42,7 @@ public:
     output_.ids->isEqualTo(*input.ids, config_.eos_id());
   }
 
-  virtual void backward(const UpdateCallback& callback) {}
+  void backward(const UpdateCallback& callback) override {}
 };
 
 REGISTER_LAYER(eos_id, EosIdCheckLayer);

paddle/gserver/layers/ExpandConvBaseLayer.h

@@ -48,7 +48,8 @@ public:
 
   ~ExpandConvBaseLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
   size_t getOutputSize();
   /**

paddle/gserver/layers/ExpandConvLayer.h

@@ -35,10 +35,11 @@ public:
 
   ~ExpandConvLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback) override;
 };
 
 }  // namespace paddle

paddle/gserver/layers/ExpandConvTransLayer.h

@@ -34,10 +34,11 @@ public:
 
   ~ExpandConvTransLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback) override;
 };
 
 }  // namespace paddle

paddle/gserver/layers/ExpandLayer.h

@@ -53,10 +53,11 @@ public:
 
   ~ExpandLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 }  // namespace paddle

paddle/gserver/layers/FeatureMapExpandLayer.cpp

@@ -46,10 +46,11 @@ public:
 
   ~FeatureMapExpandLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 REGISTER_LAYER(featmap_expand, FeatureMapExpandLayer);

paddle/gserver/layers/FullyConnectedLayer.h

@@ -36,13 +36,14 @@ public:
   explicit FullyConnectedLayer(const LayerConfig& config) : Layer(config) {}
   ~FullyConnectedLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
   Weight& getWeight(int idx) { return *weights_[idx]; }
 
-  void prefetch();
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void prefetch() override;
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 }  // namespace paddle

paddle/gserver/layers/GatedRecurrentLayer.h

@@ -50,17 +50,18 @@ class GatedRecurrentLayer : public Layer, public GruCompute {
 public:
   explicit GatedRecurrentLayer(const LayerConfig& config) : Layer(config) {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
+  void forward(PassType passType) override;
 
-  void backward(const UpdateCallback& callback);
+  void backward(const UpdateCallback& callback) override;
 
-  void resetState();
+  void resetState() override;
 
-  void setState(LayerStatePtr state);
+  void setState(LayerStatePtr state) override;
 
-  LayerStatePtr getState();
+  LayerStatePtr getState() override;
 
 protected:
   void forwardSequence(int batchSize,

paddle/gserver/layers/GetOutputLayer.cpp

@@ -22,17 +22,18 @@ public:
 
   ~GetOutputLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap) {
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override {
     if (!Layer::init(layerMap, parameterMap)) return false;
     CHECK_EQ(1U, inputLayers_.size());
     CHECK_NE(inputArgument_[0], "");
     return true;
   }
 
-  void forward(PassType passType) {
+  void forward(PassType passType) override {
     output_ = getPrev(0)->getOutput(inputArgument_[0]);
   }
-  void backward(const UpdateCallback& callback = nullptr) {}
+  void backward(const UpdateCallback& callback = nullptr) override {}
 };
 
 REGISTER_LAYER(get_output, GetOutputLayer);

paddle/gserver/layers/GruStepLayer.cpp

@@ -55,10 +55,11 @@ public:
 
   ~GruStepLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 REGISTER_LAYER(gru_step, GruStepLayer);

paddle/gserver/layers/HierarchicalSigmoidLayer.h

@@ -61,9 +61,10 @@ class HierarchicalSigmoidLayer : public Layer {
 public:
   explicit HierarchicalSigmoidLayer(const LayerConfig& config)
       : Layer(config) {}
-  virtual bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
-  virtual void forward(PassType passType);
-  virtual void backward(const UpdateCallback& callback);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback) override;
 
 protected:
   /**

paddle/gserver/layers/InterpolationLayer.cpp

@@ -43,10 +43,11 @@ public:
 
   ~InterpolationLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 REGISTER_LAYER(interpolation, InterpolationLayer);

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/layers/LstmLayer.h

@@ -74,17 +74,18 @@ class LstmLayer : public Layer, public LstmCompute {
 public:
   explicit LstmLayer(const LayerConfig &config) : Layer(config) {}
 
-  bool init(const LayerMap &layerMap, const ParameterMap &parameterMap);
+  bool init(const LayerMap &layerMap,
+            const ParameterMap &parameterMap) override;
 
-  void forward(PassType passType);
+  void forward(PassType passType) override;
 
-  void backward(const UpdateCallback &callback);
+  void backward(const UpdateCallback &callback) override;
 
-  void resetState();
+  void resetState() override;
 
-  void setState(LayerStatePtr state);
+  void setState(LayerStatePtr state) override;
 
-  LayerStatePtr getState();
+  LayerStatePtr getState() override;
 
 protected:
   /**

paddle/gserver/layers/LstmStepLayer.cpp

@@ -35,10 +35,11 @@ public:
 
   ~LstmStepLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 REGISTER_LAYER(lstm_step, LstmStepLayer);

paddle/gserver/layers/MDLstmLayer.cpp

@@ -181,11 +181,12 @@ class MDLstmLayer : public LstmLayer {
 public:
   explicit MDLstmLayer(const LayerConfig& config) : LstmLayer(config) {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
+  void forward(PassType passType) override;
 
-  void backward(const UpdateCallback& callback);
+  void backward(const UpdateCallback& callback) override;
 
 protected:
   void forwardOneSequence(int start, CoordIterator& coordIter);

paddle/gserver/layers/MaxIdLayer.cpp

@@ -30,8 +30,8 @@ private:
 public:
   explicit MaxIdLayer(const LayerConfig& config) : Layer(config) {}
 
-  virtual bool init(const LayerMap& layerMap,
-                    const ParameterMap& parameterMap) {
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override {
     bool ret = Layer::init(layerMap, parameterMap);
     CHECK_EQ(1UL, inputLayers_.size());
 
@@ -40,7 +40,7 @@ public:
     return ret;
   }
 
-  virtual void forward(PassType passType) {
+  void forward(PassType passType) override {
     Layer::forward(passType);
     const Argument& input = getInput(0);
     size_t batchSize = input.getBatchSize();
@@ -54,7 +54,7 @@ public:
     input.value->rowMax(*output_.ids, *output_.in);
   }
 
-  virtual void backward(const UpdateCallback& callback) {}
+  void backward(const UpdateCallback& callback) override {}
 };
 
 REGISTER_LAYER(maxid, MaxIdLayer);

paddle/gserver/layers/MaxLayer.h

@@ -42,14 +42,13 @@ protected:
 public:
   explicit MaxLayer(const LayerConfig& config) : SequencePoolLayer(config) {}
 
-  ~MaxLayer() {}
-
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap) {
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override {
     return SequencePoolLayer::init(layerMap, parameterMap);
   }
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 }  // namespace paddle

paddle/gserver/layers/MaxOutLayer.h

@@ -45,10 +45,11 @@ public:
   explicit MaxOutLayer(const LayerConfig& config) : Layer(config) {}
   virtual ~MaxOutLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 }  // namespace paddle

paddle/gserver/layers/MixedLayer.h

@@ -35,21 +35,22 @@ public:
 
   ~MixedLayer() {}
 
-  virtual bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  virtual void prefetch();
-  virtual void forward(PassType passType);
-  virtual void backward(const UpdateCallback& callback = nullptr);
-  virtual void resetState();
+  void prefetch() override;
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
+  void resetState() override;
   /**
    * setState() should be called after getState().
    * Argument state consists of all projections states.
    */
-  virtual void setState(LayerStatePtr state);
+  void setState(LayerStatePtr state) override;
   /**
    * Return state which consists of all projections states.
    */
-  virtual LayerStatePtr getState();
+  LayerStatePtr getState() override;
 
 protected:
   std::vector<std::unique_ptr<Projection>> projections_;

paddle/gserver/layers/MultiplexLayer.cpp

@@ -69,10 +69,11 @@ public:
 
   ~MultiplexLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 
 private:
   /**

paddle/gserver/layers/NCELayer.cpp

@@ -61,7 +61,8 @@ public:
         rand_(0, config.num_classes() - 1),
         prepared_(false) {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap) {
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override {
     /* Initialize the basic parent class */
     Layer::init(layerMap, parameterMap);
 
@@ -146,7 +147,7 @@ public:
     prepared_ = true;
   }
 
-  void prefetch() {
+  void prefetch() override {
     prepareSamples();
     IVector::resizeOrCreate(labelIds_, samples_.size(), useGpu_);
     int* ids = labelIds_->getData();
@@ -163,7 +164,7 @@ public:
     }
   }
 
-  void forward(PassType passType) {
+  void forward(PassType passType) override {
     Layer::forward(passType);
 
     CHECK(!useGpu_) << "GPU is not supported";
@@ -199,7 +200,7 @@ public:
     forwardCost();
   }
 
-  void backward(const UpdateCallback& callback) {
+  void backward(const UpdateCallback& callback) override {
     Matrix::resizeOrCreate(sampleOut_.grad,
                            1,
                            samples_.size(),

paddle/gserver/layers/NormLayer.h

@@ -30,7 +30,8 @@ class NormLayer : public Layer {
 public:
   explicit NormLayer(const LayerConfig& config) : Layer(config) {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap) {
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override {
     Layer::init(layerMap, parameterMap);
     return true;
   }
@@ -56,9 +57,10 @@ protected:
 public:
   explicit ResponseNormLayer(const LayerConfig& config) : NormLayer(config) {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
-  void forward(PassType passType) { LOG(FATAL) << "Not implemented"; }
-  void backward(const UpdateCallback& callback = nullptr) {
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+  void forward(PassType passType) override { LOG(FATAL) << "Not implemented"; }
+  void backward(const UpdateCallback& callback = nullptr) override {
     LOG(FATAL) << "Not implemented";
   }
 };

paddle/gserver/layers/NormProjectionLayer.h

@@ -36,9 +36,10 @@ public:
 
   size_t getSize();
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 
 protected:
   TensorShape shape_;

paddle/gserver/layers/OuterProdLayer.cpp

@@ -38,10 +38,11 @@ public:
 
   ~OuterProdLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 REGISTER_LAYER(out_prod, OuterProdLayer);

paddle/gserver/layers/PadLayer.h

@@ -29,9 +29,10 @@ public:
 
   ~PadLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 
 protected:
   void setOutDims(const size_t batchSize);

paddle/gserver/layers/ParameterReluLayer.h

@@ -56,9 +56,10 @@ public:
 
   ~ParameterReluLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 }  // namespace paddle

paddle/gserver/layers/PoolLayer.h

@@ -46,7 +46,8 @@ public:
    */
   static Layer* create(const LayerConfig& config);
 
-  virtual bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 };
 
 }  // namespace paddle

paddle/gserver/layers/PoolProjectionLayer.h

@@ -40,7 +40,7 @@ public:
 
   size_t getSize();
 
-  virtual void forward(PassType passType);
-  virtual void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 }  // namespace paddle

paddle/gserver/layers/PowerLayer.cpp

@@ -40,10 +40,11 @@ public:
 
   ~PowerLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 REGISTER_LAYER(power, PowerLayer);

paddle/gserver/layers/PrintLayer.cpp

@@ -19,38 +19,17 @@ namespace paddle {
 class PrintLayer : public Layer {
 public:
   explicit PrintLayer(const LayerConfig& config) : Layer(config) {}
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback) {}
-};
 
-void PrintLayer::forward(PassType passType) {
-  Layer::forward(passType);
-  for (size_t i = 0; i != inputLayers_.size(); ++i) {
-    const auto& argu = getInput(i);
-    const std::string& name = inputLayers_[i]->getName();
-    if (argu.value) {
-      std::ostringstream os;
-      argu.value->print(os);
-      LOG(INFO) << "layer=" << name << " value matrix:\n" << os.str();
-    }
-    if (argu.ids) {
-      std::ostringstream os;
-      argu.ids->print(os, argu.ids->getSize());
-      LOG(INFO) << "layer=" << name << " ids vector:\n" << os.str();
-    }
-    if (auto startPos = argu.sequenceStartPositions) {
-      std::ostringstream os;
-      startPos->getVector(false)->print(os, startPos->getSize());
-      LOG(INFO) << "layer=" << name << " sequence pos vector:\n" << os.str();
-    }
-    if (auto subStartPos = argu.subSequenceStartPositions) {
-      std::ostringstream os;
-      subStartPos->getVector(false)->print(os, subStartPos->getSize());
-      LOG(INFO) << "layer=" << name << " sub-sequence pos vector:\n"
-                << os.str();
+  void forward(PassType passType) override {
+    Layer::forward(passType);
+    for (size_t i = 0; i != inputLayers_.size(); ++i) {
+      getInput(i).printValueString(LOG(INFO),
+                                   "layer=" + inputLayers_[i]->getName() + " ");
     }
   }
-}
+
+  void backward(const UpdateCallback& callback) override {}
+};
 
 REGISTER_LAYER(print, PrintLayer);
 

paddle/gserver/layers/PriorBox.cpp

@@ -30,10 +30,11 @@ namespace paddle {
 class PriorBoxLayer : public Layer {
 public:
   explicit PriorBoxLayer(const LayerConfig& config) : Layer(config) {}
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback) {}
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback) override {}
 
 protected:
   int numPriors_;

paddle/gserver/layers/RecurrentLayer.cpp

@@ -45,17 +45,18 @@ class RecurrentLayer : public Layer {
 public:
   explicit RecurrentLayer(const LayerConfig& config) : Layer(config) {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
+  void forward(PassType passType) override;
 
-  void backward(const UpdateCallback& callback);
+  void backward(const UpdateCallback& callback) override;
 
-  void resetState();
+  void resetState() override;
 
-  void setState(LayerStatePtr state);
+  void setState(LayerStatePtr state) override;
 
-  LayerStatePtr getState();
+  LayerStatePtr getState() override;
 
 protected:
   /**

paddle/gserver/layers/RecurrentLayerGroup.cpp

@@ -33,15 +33,15 @@ public:
   void initSubNetwork(NeuralNetwork* rootNetwork,
                       const ModelConfig& config,
                       const std::vector<ParameterType>& parameterTypes,
-                      bool useGpu);
+                      bool useGpu) override;
 
-  void forward(PassType passType) {
+  void forward(PassType passType) override {
     REGISTER_TIMER_INFO("RecurrentGroupFwTime", getName().c_str());
     const std::vector<Argument> inArgs;
     std::vector<Argument> outArgs;
     network_->forward(inArgs, &outArgs, passType);
   }
-  void backward(const UpdateCallback& callback) {
+  void backward(const UpdateCallback& callback) override {
     REGISTER_TIMER_INFO("RecurrentGroupBwTime", getName().c_str());
     network_->backward(nullptr);
 
@@ -53,7 +53,8 @@ public:
   /**
    * @see Layer.accessSubNetwork
    */
-  void accessSubNetwork(const std::function<void(NeuralNetwork&)>& callback) {
+  void accessSubNetwork(
+      const std::function<void(NeuralNetwork&)>& callback) override {
     callback(*network_);
   }
 

paddle/gserver/layers/ResizeLayer.cpp

@@ -20,18 +20,19 @@ namespace paddle {
 /**
  * @brief A layer for resizing a minibatch matrix h*w to h'*w'
  * @note
- * origin matrix height * witdth)
+ * origin matrix height * width)
  * resize matrix: (height * width / size) * size
  */
 class ResizeLayer : public Layer {
 public:
   explicit ResizeLayer(const LayerConfig& config) : Layer(config) {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
+  void forward(PassType passType) override;
 
-  void backward(const UpdateCallback& callback);
+  void backward(const UpdateCallback& callback) override;
 };
 
 REGISTER_LAYER(resize, ResizeLayer);

paddle/gserver/layers/SamplingIdLayer.cpp

@@ -35,8 +35,8 @@ public:
   explicit SamplingIdLayer(const LayerConfig& config)
       : Layer(config), rand1_(0, 1) {}
 
-  virtual bool init(const LayerMap& layerMap,
-                    const ParameterMap& parameterMap) {
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override {
     bool ret = Layer::init(layerMap, parameterMap);
     CHECK_EQ(1UL, inputLayers_.size());
     if (useGpu_) {
@@ -48,7 +48,7 @@ public:
     return ret;
   }
 
-  void forward(PassType passType) {
+  void forward(PassType passType) override {
     Layer::forward(passType);
     if (useGpu_) {
       for (size_t i = 0; i < inputLayers_.size(); i++) {
@@ -83,7 +83,7 @@ public:
     output_.ids->copyFrom(ids.data(), batchSize);
   }
 
-  virtual void backward(const UpdateCallback& callback) {}
+  void backward(const UpdateCallback& callback) override {}
 };
 
 REGISTER_LAYER(sampling_id, SamplingIdLayer);

paddle/gserver/layers/ScalingLayer.cpp

@@ -37,10 +37,11 @@ public:
 
   ~ScalingLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 REGISTER_LAYER(scaling, ScalingLayer);

paddle/gserver/layers/SelectiveFullyConnectedLayer.h

@@ -65,9 +65,10 @@ public:
       : Layer(config), selCols_(nullptr) {}
 
   ~SelectiveFullyConnectedLayer() {}
-  void prefetch();
+  void prefetch() override;
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
   Weight& getWeight(int idx) { return *weights_[idx]; }
 
@@ -90,8 +91,8 @@ public:
   void fillSelectiveData(
       const std::shared_ptr<std::vector<std::pair<int*, size_t>>>& candidates);
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 
 private:
   /**

paddle/gserver/layers/SequenceConcatLayer.cpp

@@ -21,9 +21,11 @@ namespace paddle {
 
 /**
  * A layer for concatenating the first sequence with the second sequence
- * following the first
- * Input: two sequences each containing some instances
+ * Input: two sequences each containing the same number of instances
+ *        seq1 = [a1, a2, ..., an]
+ *        seq2 = [b1, b2, ..., bn]
  * Output: a concatenated sequence of the two input sequences
+ *        out = [a1, b1, a2, b2, ..., an, bn]
  */
 
 class SequenceConcatLayer : public Layer {
@@ -35,10 +37,11 @@ public:
 
   ~SequenceConcatLayer() {}
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 REGISTER_LAYER(seqconcat, SequenceConcatLayer);
@@ -167,13 +170,17 @@ void SequenceConcatLayer::backward(const UpdateCallback& callback) {
     size_t rightNumIns = 0;
     for (size_t seqId = 0; seqId < numSequences1; ++seqId) {
       leftNumIns = starts1[seqId + 1] - starts1[seqId];
-      inputGrad1->subMatrix(starts1[seqId], leftNumIns)
-          ->add(*(outputGrad->subMatrix(offset, leftNumIns)));
+      if (inputGrad1) {
+        inputGrad1->subMatrix(starts1[seqId], leftNumIns)
+            ->add(*(outputGrad->subMatrix(offset, leftNumIns)));
+      }
       offset += leftNumIns;
 
       rightNumIns = starts2[seqId + 1] - starts2[seqId];
-      inputGrad2->subMatrix(starts2[seqId], rightNumIns)
-          ->add(*(outputGrad->subMatrix(offset, rightNumIns)));
+      if (inputGrad2) {
+        inputGrad2->subMatrix(starts2[seqId], rightNumIns)
+            ->add(*(outputGrad->subMatrix(offset, rightNumIns)));
+      }
       offset += rightNumIns;
     }
   }

paddle/gserver/layers/SequenceLastInstanceLayer.cpp

@@ -42,12 +42,11 @@ public:
   explicit SequenceLastInstanceLayer(const LayerConfig& config)
       : SequencePoolLayer(config) {}
 
-  ~SequenceLastInstanceLayer() {}
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
-
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 REGISTER_LAYER(seqlastins, SequenceLastInstanceLayer);

paddle/gserver/layers/SequencePoolLayer.h

@@ -46,12 +46,11 @@ protected:
 public:
   explicit SequencePoolLayer(const LayerConfig& config) : Layer(config) {}
 
-  virtual ~SequencePoolLayer() {}
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
-
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 }  // namespace paddle

paddle/gserver/layers/SequenceReshapeLayer.cpp

@@ -20,9 +20,12 @@ limitations under the License. */
 namespace paddle {
 
 /**
- * A layer for reshaping the sequence
- * Input: a sequence
- * Output: a sequence
+ *  A layer for reshaping the sequence. Assume the input sequence has
+ *  T instances, the dimension of each instance is M, and the input
+ *  reshape_dim is N, then the output sequence has T*M/N instances,
+ *  the dimension of each instance is N.
+ *
+ *  Note that T*M/N must be an integer.
  */
 
 class SequenceReshapeLayer : public Layer {
@@ -34,12 +37,11 @@ protected:
 public:
   explicit SequenceReshapeLayer(const LayerConfig& config) : Layer(config) {}
 
-  ~SequenceReshapeLayer() {}
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
 
-  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
-
-  void forward(PassType passType);
-  void backward(const UpdateCallback& callback = nullptr);
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
 };
 
 REGISTER_LAYER(seqreshape, SequenceReshapeLayer);