diff --git a/demo/image_classification/api_v2_resnet.py b/demo/image_classification/api_v2_resnet.py
new file mode 100644
index 0000000000000000000000000000000000000000..19d20540780becf504973a23b50445d4b65dc2ef
--- /dev/null
+++ b/demo/image_classification/api_v2_resnet.py
@@ -0,0 +1,74 @@
+# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import paddle.v2 as paddle
+
+__all__ = ['resnet_cifar10']
+
+
+def conv_bn_layer(input,
+                  ch_out,
+                  filter_size,
+                  stride,
+                  padding,
+                  active_type=paddle.activation.Relu(),
+                  ch_in=None):
+    tmp = paddle.layer.img_conv(
+        input=input,
+        filter_size=filter_size,
+        num_channels=ch_in,
+        num_filters=ch_out,
+        stride=stride,
+        padding=padding,
+        act=paddle.activation.Linear(),
+        bias_attr=False)
+    return paddle.layer.batch_norm(input=tmp, act=active_type)
+
+
+def shortcut(ipt, n_in, n_out, stride):
+    if n_in != n_out:
+        return conv_bn_layer(ipt, n_out, 1, stride, 0,
+                             paddle.activation.Linear())
+    else:
+        return ipt
+
+
+def basicblock(ipt, ch_out, stride):
+    ch_in = ch_out * 2
+    tmp = conv_bn_layer(ipt, ch_out, 3, stride, 1)
+    tmp = conv_bn_layer(tmp, ch_out, 3, 1, 1, paddle.activation.Linear())
+    short = shortcut(ipt, ch_in, ch_out, stride)
+    return paddle.layer.addto(input=[tmp, short], act=paddle.activation.Relu())
+
+
+def layer_warp(block_func, ipt, features, count, stride):
+    tmp = block_func(ipt, features, stride)
+    for i in range(1, count):
+        tmp = block_func(tmp, features, 1)
+    return tmp
+
+
+def resnet_cifar10(ipt, depth=32):
+    # depth should be one of 20, 32, 44, 56, 110, 1202
+    assert (depth - 2) % 6 == 0
+    n = (depth - 2) / 6
+    nStages = {16, 64, 128}
+    conv1 = conv_bn_layer(
+        ipt, ch_in=3, ch_out=16, filter_size=3, stride=1, padding=1)
+    res1 = layer_warp(basicblock, conv1, 16, n, 1)
+    res2 = layer_warp(basicblock, res1, 32, n, 2)
+    res3 = layer_warp(basicblock, res2, 64, n, 2)
+    pool = paddle.layer.img_pool(
+        input=res3, pool_size=8, stride=1, pool_type=paddle.pooling.Avg())
+    return pool
diff --git a/demo/image_classification/api_v2_train.py b/demo/image_classification/api_v2_train.py
new file mode 100644
index 0000000000000000000000000000000000000000..585f61c6fa4c89c8621815a168742429ac236898
--- /dev/null
+++ b/demo/image_classification/api_v2_train.py
@@ -0,0 +1,91 @@
+# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License
+
+import sys
+import paddle.v2 as paddle
+from api_v2_vgg import vgg_bn_drop
+from api_v2_resnet import resnet_cifar10
+
+
+def main():
+    datadim = 3 * 32 * 32
+    classdim = 10
+
+    # PaddlePaddle init
+    paddle.init(use_gpu=True, trainer_count=1)
+
+    image = paddle.layer.data(
+        name="image", type=paddle.data_type.dense_vector(datadim))
+
+    # Add neural network config
+    # option 1. resnet
+    net = resnet_cifar10(image, depth=32)
+    # option 2. vgg
+    # net = vgg_bn_drop(image)
+
+    out = paddle.layer.fc(input=net,
+                          size=classdim,
+                          act=paddle.activation.Softmax())
+
+    lbl = paddle.layer.data(
+        name="label", type=paddle.data_type.integer_value(classdim))
+    cost = paddle.layer.classification_cost(input=out, label=lbl)
+
+    # Create parameters
+    parameters = paddle.parameters.create(cost)
+
+    # Create optimizer
+    momentum_optimizer = paddle.optimizer.Momentum(
+        momentum=0.9,
+        regularization=paddle.optimizer.L2Regularization(rate=0.0002 * 128),
+        learning_rate=0.1 / 128.0,
+        learning_rate_decay_a=0.1,
+        learning_rate_decay_b=50000 * 100,
+        learning_rate_schedule='discexp',
+        batch_size=128)
+
+    # End batch and end pass event handler
+    def event_handler(event):
+        if isinstance(event, paddle.event.EndIteration):
+            if event.batch_id % 100 == 0:
+                print "\nPass %d, Batch %d, Cost %f, %s" % (
+                    event.pass_id, event.batch_id, event.cost, event.metrics)
+            else:
+                sys.stdout.write('.')
+                sys.stdout.flush()
+        if isinstance(event, paddle.event.EndPass):
+            result = trainer.test(
+                reader=paddle.reader.batched(
+                    paddle.dataset.cifar.test10(), batch_size=128),
+                reader_dict={'image': 0,
+                             'label': 1})
+            print "\nTest with Pass %d, %s" % (event.pass_id, result.metrics)
+
+    # Create trainer
+    trainer = paddle.trainer.SGD(cost=cost,
+                                 parameters=parameters,
+                                 update_equation=momentum_optimizer)
+    trainer.train(
+        reader=paddle.reader.batched(
+            paddle.reader.shuffle(
+                paddle.dataset.cifar.train10(), buf_size=50000),
+            batch_size=128),
+        num_passes=5,
+        event_handler=event_handler,
+        reader_dict={'image': 0,
+                     'label': 1})
+
+
+if __name__ == '__main__':
+    main()
diff --git a/demo/image_classification/api_v2_vgg.py b/demo/image_classification/api_v2_vgg.py
new file mode 100644
index 0000000000000000000000000000000000000000..1e0e6b93adde30425f17aa9cd07542275f4fec37
--- /dev/null
+++ b/demo/image_classification/api_v2_vgg.py
@@ -0,0 +1,47 @@
+# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import paddle.v2 as paddle
+
+__all__ = ['vgg_bn_drop']
+
+
+def vgg_bn_drop(input):
+    def conv_block(ipt, num_filter, groups, dropouts, num_channels=None):
+        return paddle.networks.img_conv_group(
+            input=ipt,
+            num_channels=num_channels,
+            pool_size=2,
+            pool_stride=2,
+            conv_num_filter=[num_filter] * groups,
+            conv_filter_size=3,
+            conv_act=paddle.activation.Relu(),
+            conv_with_batchnorm=True,
+            conv_batchnorm_drop_rate=dropouts,
+            pool_type=paddle.pooling.Max())
+
+    conv1 = conv_block(input, 64, 2, [0.3, 0], 3)
+    conv2 = conv_block(conv1, 128, 2, [0.4, 0])
+    conv3 = conv_block(conv2, 256, 3, [0.4, 0.4, 0])
+    conv4 = conv_block(conv3, 512, 3, [0.4, 0.4, 0])
+    conv5 = conv_block(conv4, 512, 3, [0.4, 0.4, 0])
+
+    drop = paddle.layer.dropout(input=conv5, dropout_rate=0.5)
+    fc1 = paddle.layer.fc(input=drop, size=512, act=paddle.activation.Linear())
+    bn = paddle.layer.batch_norm(
+        input=fc1,
+        act=paddle.activation.Relu(),
+        layer_attr=paddle.attr.Extra(drop_rate=0.5))
+    fc2 = paddle.layer.fc(input=bn, size=512, act=paddle.activation.Linear())
+    return fc2
diff --git a/demo/semantic_role_labeling/api_train_v2.py b/demo/semantic_role_labeling/api_train_v2.py
index c582724185926853f3372f1265c1464fb92b2b80..e946a792b5f51ab54355aac0b6e9aef51ae815fb 100644
--- a/demo/semantic_role_labeling/api_train_v2.py
+++ b/demo/semantic_role_labeling/api_train_v2.py
@@ -165,7 +165,7 @@ def main():
 
     trn_reader = paddle.reader.batched(
         paddle.reader.shuffle(
-            conll05.test, buf_size=8192), batch_size=10)
+            conll05.test(), buf_size=8192), batch_size=10)
 
     trainer.train(
         reader=trn_reader, event_handler=event_handler, num_passes=10000)
diff --git a/paddle/api/GradientMachine.cpp b/paddle/api/GradientMachine.cpp
index 538ca2999f8f05afc45ac2d2f526133c8024f066..dcb5fe086fdccf8ec62ee52cbaaac4b7dbbe2f9d 100644
--- a/paddle/api/GradientMachine.cpp
+++ b/paddle/api/GradientMachine.cpp
@@ -142,6 +142,20 @@ Parameter* GradientMachine::getParameter(size_t i) throw(RangeError) {
   }
 }
 
+size_t GradientMachine::getNonStaticParameterSize() const {
+  return m->machine->getNonStaticParameters().size();
+}
+
+Parameter* GradientMachine::getNonStaticParameter(size_t i) throw(RangeError) {
+  auto params = m->machine->getNonStaticParameters();
+  if (i < params.size()) {
+    return Parameter::createFromSharedPtr(
+        &m->machine->getNonStaticParameters()[i]);
+  } else {
+    throw RangeError();
+  }
+}
+
 void GradientMachine::randParameters() { m->machine->randParameters(); }
 
 Arguments* GradientMachine::getLayerOutput(const std::string& layerName) const
diff --git a/paddle/api/PaddleAPI.h b/paddle/api/PaddleAPI.h
index d99e9a4ad48ea4764c7a1ea56c507d754d56853b..762f86ac79461558b6a2eb7105ffd05961f5d3e2 100644
--- a/paddle/api/PaddleAPI.h
+++ b/paddle/api/PaddleAPI.h
@@ -771,6 +771,9 @@ public:
   size_t getParameterSize() const;
   Parameter* getParameter(size_t i) throw(RangeError);
 
+  size_t getNonStaticParameterSize() const;
+  Parameter* getNonStaticParameter(size_t i) throw(RangeError);
+
   void randParameters();
 
   Arguments* getLayerOutput(const std::string& layerName) const
diff --git a/paddle/py_paddle/util.py b/paddle/py_paddle/util.py
index a708def1d2d7f6da2998a5905f9473accc6db969..1c9455fab5f9c1179bddffb100cd53fe8adfb6b1 100644
--- a/paddle/py_paddle/util.py
+++ b/paddle/py_paddle/util.py
@@ -195,6 +195,12 @@ def __monkeypatch_gradient_machine__():
 
     swig_paddle.GradientMachine.getParameters = getParameters
 
+    def getNonStaticParameters(self):
+        return (self.getNonStaticParameter(i)
+                for i in xrange(self.getNonStaticParameterSize()))
+
+    swig_paddle.GradientMachine.getNonStaticParameters = getNonStaticParameters
+
     def getLayerOutputs(self, layerNames):
         """
         getLayerOutputs. get outputs of layers and return a numpy matrix dict.
diff --git a/python/paddle/v2/__init__.py b/python/paddle/v2/__init__.py
index d548d1adaafacdb097dbe476fdc76651c9f46b6b..8ab8cd2f85d5d7bcf86c2f57b350dfcd99177b69 100644
--- a/python/paddle/v2/__init__.py
+++ b/python/paddle/v2/__init__.py
@@ -20,6 +20,7 @@ import event
 import data_type
 import topology
 import data_feeder
+import networks
 from . import dataset
 from . import reader
 import attr
diff --git a/python/paddle/v2/config_base.py b/python/paddle/v2/config_base.py
index 035f96b0f2e978a413a1ebe0ec115f75ff07befc..fa2ccec6c3270541dd6b13fdfd2323d10ceac642 100644
--- a/python/paddle/v2/config_base.py
+++ b/python/paddle/v2/config_base.py
@@ -22,6 +22,7 @@ class Layer(object):
     def __init__(self, name=None, parent_layers=None):
         assert isinstance(parent_layers, dict)
         self.name = name
+        self.__contex__ = {}
         self.__parent_layers__ = parent_layers
 
     def to_proto(self, context):
@@ -39,16 +40,38 @@ class Layer(object):
                                self.__parent_layers__[layer_name])
             kwargs[layer_name] = v1_layer
 
-        if self.name is None:
+        if self.context_name() is None:
             return self.to_proto_impl(**kwargs)
-        elif self.name not in context:
-            context[self.name] = self.to_proto_impl(**kwargs)
-
-        return context[self.name]
+        elif self.context_name() not in context:
+            context[self.context_name()] = self.to_proto_impl(**kwargs)
+        self.__contex__ = context
+        if self.use_context_name():
+            return context[self.context_name()]
+        else:
+            return context[self.name]
 
     def to_proto_impl(self, **kwargs):
         raise NotImplementedError()
 
+    def context_name(self):
+        """
+        Context name means the context which stores `to_proto_impl` result.
+        If multiple layer share same context_name, the `to_proto_impl` of them
+        will be invoked only once.
+        """
+        return self.name
+
+    def use_context_name(self):
+        return False
+
+    def calculate_size(self):
+        """
+        lazy calculate size of the layer, should be called when to_proto_impl of
+        this layer is called.
+        :return:
+        """
+        return self.__contex__[self.context_name()].size
+
 
 def __convert_to_v2__(method_name, parent_names, is_default_name=True):
     if is_default_name:
diff --git a/python/paddle/v2/dataset/__init__.py b/python/paddle/v2/dataset/__init__.py
index 90803628e3e0c52fbe5846c76ca065b8ab18c147..a1b21bab3bac8b304abb4ae292b1c1e9f3e719de 100644
--- a/python/paddle/v2/dataset/__init__.py
+++ b/python/paddle/v2/dataset/__init__.py
@@ -13,5 +13,10 @@
 # limitations under the License.
 
 import mnist
+import imikolov
+import imdb
+import cifar
+import movielens
+import conll05
 
-__all__ = ['mnist', 'cifar', 'imdb', 'conll05', 'imikolov', 'movielens']
+__all__ = ['mnist', 'imikolov', 'imdb', 'cifar', 'movielens', 'conll05']
diff --git a/python/paddle/v2/dataset/conll05.py b/python/paddle/v2/dataset/conll05.py
index 52f19d2115f545213962199914f7e40d13dbe92b..e96a701c1a944e2d6d84f897157cb357c5aa0824 100644
--- a/python/paddle/v2/dataset/conll05.py
+++ b/python/paddle/v2/dataset/conll05.py
@@ -12,10 +12,10 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import paddle.v2.dataset.common
 import tarfile
 import gzip
 import itertools
+from common import download
 
 __all__ = ['test, get_dict', 'get_embedding']
 """
@@ -173,33 +173,24 @@ def reader_creator(corpus_reader,
             yield word_idx, ctx_n2_idx, ctx_n1_idx, \
               ctx_0_idx, ctx_p1_idx, ctx_p2_idx, pred_idx, mark, label_idx
 
-    return reader()
+    return reader
 
 
 def get_dict():
-    word_dict = load_dict(
-        common.download(WORDDICT_URL, 'conll05st', WORDDICT_MD5))
-    verb_dict = load_dict(
-        common.download(VERBDICT_URL, 'conll05st', VERBDICT_MD5))
-    label_dict = load_dict(
-        common.download(TRGDICT_URL, 'conll05st', TRGDICT_MD5))
+    word_dict = load_dict(download(WORDDICT_URL, 'conll05st', WORDDICT_MD5))
+    verb_dict = load_dict(download(VERBDICT_URL, 'conll05st', VERBDICT_MD5))
+    label_dict = load_dict(download(TRGDICT_URL, 'conll05st', TRGDICT_MD5))
     return word_dict, verb_dict, label_dict
 
 
 def get_embedding():
-    return common.download(EMB_URL, 'conll05st', EMB_MD5)
+    return download(EMB_URL, 'conll05st', EMB_MD5)
 
 
 def test():
     word_dict, verb_dict, label_dict = get_dict()
     reader = corpus_reader(
-        common.download(DATA_URL, 'conll05st', DATA_MD5),
+        download(DATA_URL, 'conll05st', DATA_MD5),
         words_name='conll05st-release/test.wsj/words/test.wsj.words.gz',
         props_name='conll05st-release/test.wsj/props/test.wsj.props.gz')
     return reader_creator(reader, word_dict, verb_dict, label_dict)
-
-
-if __name__ == '__main__':
-    print get_embedding()
-    for f in test():
-        print f
diff --git a/python/paddle/v2/layer.py b/python/paddle/v2/layer.py
index 67111f1315fbb0f55c1db0f6fe89fc988c8d83f6..2f55611aaa1d3ae22f5d7f184b38e622271881ea 100644
--- a/python/paddle/v2/layer.py
+++ b/python/paddle/v2/layer.py
@@ -65,19 +65,24 @@ to be in a Python function but could be anywhere.
 Also, the creation of a protobuf message is hidden in the invocation of
 paddle.v2.parameters.create, no longer exposed to users.
 """
+
+import collections
+import inspect
 from config_base import Layer, __convert_to_v2__
 import paddle.trainer_config_helpers as conf_helps
 from paddle.trainer_config_helpers.config_parser_utils import \
     parse_network_config as __parse__
-
-from paddle.trainer_config_helpers.default_decorators import wrap_name_default
 from paddle.trainer_config_helpers.default_decorators import wrap_act_default
 from paddle.trainer_config_helpers.default_decorators import \
     wrap_bias_attr_default
+from paddle.trainer_config_helpers.default_decorators import wrap_name_default
 from paddle.trainer_config_helpers.layers import layer_support
+from paddle.trainer.config_parser import \
+    RecurrentLayerGroupWithoutOutLinksBegin, RecurrentLayerGroupSetOutLink, \
+    RecurrentLayerGroupEnd, model_type
 
-import data_type
 import activation
+import data_type
 
 __all__ = ['parse_network', 'data']
 
@@ -130,6 +135,137 @@ class DataLayerV2(Layer):
         return getattr(conf_helps, self.__method_name__)(name=self.name, **args)
 
 
+class WithExtraParent(Layer):
+    def extra_parent(self):
+        return self.__extra_parent__
+
+    def __init__(self, name=None, parent_layers=None):
+        self.__extra_parent__ = []
+        super(WithExtraParent, self).__init__(
+            name=name, parent_layers=parent_layers)
+
+    def append_extra_parent(self, parent):
+        self.__extra_parent__.append(parent)
+
+    def to_proto(self, context):
+        """
+        function to set proto attribute
+        """
+        kwargs = dict()
+        for p in self.__extra_parent__:
+            p.to_proto(context=context)
+
+        for layer_name in self.__parent_layers__:
+            if not isinstance(self.__parent_layers__[layer_name],
+                              collections.Sequence):
+                v1_layer = self.__parent_layers__[layer_name].to_proto(
+                    context=context)
+            else:
+                v1_layer = map(lambda x: x.to_proto(context=context),
+                               self.__parent_layers__[layer_name])
+            kwargs[layer_name] = v1_layer
+
+        if self.context_name() is None:
+            return self.to_proto_impl(context=context, **kwargs)
+        elif self.context_name() not in context:
+            context[self.context_name()] = self.to_proto_impl(
+                context=context, **kwargs)
+
+        if self.use_context_name():
+            return context[self.context_name()]
+        else:
+            return context[self.name]
+
+
+class MemoryV2(WithExtraParent):
+    def __init__(self, name, **kwargs):
+        self.name = name
+        super(MemoryV2, self).__init__(name=name, parent_layers=dict())
+        self.__kwargs__ = kwargs
+        self.__boot_layer_name__ = None
+        if 'boot_layer' in kwargs:
+            begin_of_current_rnn = []
+            # TODO(yuyang18): Fix inspect, it could be wrong when user invoke a
+            # function inside step.
+            st = inspect.stack()
+            for i in xrange(len(st)):
+                locs = inspect.stack()[i][0].f_locals
+                keys = locs.keys()
+                for key in keys:
+                    val = locs[key]
+                    if isinstance(val, RecurrentLayerInput):
+                        begin_of_current_rnn.append(val)
+                    elif isinstance(val, collections.Sequence):
+                        for v in val:
+                            if isinstance(v, RecurrentLayerInput):
+                                begin_of_current_rnn.append(v)
+
+                if begin_of_current_rnn:
+                    break
+            assert begin_of_current_rnn is not None
+            for extra in begin_of_current_rnn:
+                self.append_extra_parent(extra)
+                assert isinstance(extra, WithExtraParent)
+                extra.append_extra_parent(kwargs['boot_layer'])
+                self.__boot_layer_name__ = kwargs['boot_layer'].name
+
+    def to_proto_impl(self, context, **kwargs):
+        args = dict()
+        for each in kwargs:
+            args[each] = kwargs[each]
+        for each in self.__kwargs__:
+            args[each] = self.__kwargs__[each]
+
+        if self.__boot_layer_name__ is not None:
+            args['boot_layer'] = context[self.__boot_layer_name__]
+
+        size = args.get('size', None)
+        if size is not None:
+            if callable(size):
+                real_size = size()
+            else:
+                real_size = size
+            args['size'] = real_size
+        return conf_helps.memory(name=self.name, **args)
+
+    def context_name(self):
+        return self.name + "#memory"
+
+    def use_context_name(self):
+        """
+        memory layer will have the same name with some layer
+        :return:
+        """
+        return True
+
+
+class LayerOutputV2(Layer):
+    """
+    LayerOutputV2 is used to store the result of LayerOutput in v1 api.
+    It will not store it's parents because layer_output has been parsed already.
+    """
+
+    def __init__(self, layer_output):
+        assert isinstance(layer_output, conf_helps.LayerOutput)
+        self.layer_output = layer_output
+        super(LayerOutputV2, self).__init__(
+            name=layer_output.name, parent_layers=dict())
+
+    def to_proto_impl(self):
+        return self.layer_output
+
+
+class StaticInputV2(object):
+    def __init__(self, input, is_seq=False, size=None):
+        assert isinstance(input, LayerV2)
+        self.name = input.name
+        self.input = input
+        self.is_seq = is_seq
+        self.size = size
+        # TODO(qiaolongfei): add size
+        # assert input.size is not None or size is not None
+
+
 class MixedLayerV2(Layer):
     """
     This class is use to support `with` grammar. If not, the following code
@@ -161,7 +297,6 @@ class MixedLayerV2(Layer):
         other_kwargs['act'] = act
         other_kwargs['bias_attr'] = bias_attr
         other_kwargs['layer_attr'] = layer_attr
-
         parent_layers = {"input": self.__inputs__}
         super(MixedLayerV2, self).__init__(name, parent_layers)
         self.__other_kwargs__ = other_kwargs
@@ -171,7 +306,7 @@ class MixedLayerV2(Layer):
             self.__inputs__.append(other)
             return self
         else:
-            raise MixedLayerTypeV2.AddToSealedMixedLayerExceptionV2()
+            raise MixedLayerV2.AddToSealedMixedLayerExceptionV2()
 
     def __enter__(self):
         assert len(self.__inputs__) == 0
@@ -186,6 +321,13 @@ class MixedLayerV2(Layer):
             args[each] = kwargs[each]
         for each in self.__other_kwargs__:
             args[each] = self.__other_kwargs__[each]
+        size = args.get('size', None)
+        if size is not None:
+            if callable(size):
+                real_size = size()
+            else:
+                real_size = size
+            args['size'] = real_size
         return getattr(conf_helps, self.__method_name__)(**args)
 
 
@@ -202,14 +344,51 @@ def mixed(size=0,
     return MixedLayerV2(size, input, name, act, bias_attr, layer_attr)
 
 
+class RecurrentLayerInput(WithExtraParent):
+    def __init__(self, recurrent_name, index, parent_layers):
+        assert len(parent_layers) == 1
+        self.__parents__ = parent_layers.values()[0]
+        super(RecurrentLayerInput, self).__init__(
+            name=self.__parents__[index].name, parent_layers=parent_layers)
+        self.__recurrent_name__ = recurrent_name
+
+    def context_name(self):
+        return self.__recurrent_name__ + ".begin"
+
+    def to_proto_impl(self, context, **kwargs):
+        model_type('recurrent_nn')
+        RecurrentLayerGroupWithoutOutLinksBegin(
+            name=self.__recurrent_name__,
+            in_links=map(lambda x: x.name, self.__parents__))
+        return self
+
+
+class RecurrentLayerOutput(Layer):
+    def __init__(self, recurrent_name, index, parent_layers):
+        assert len(parent_layers) == 1
+        self.__parents__ = parent_layers.values()[0]
+        super(RecurrentLayerOutput, self).__init__(
+            name=self.__parents__[index].name, parent_layers=parent_layers)
+        self.__recurrent_name__ = recurrent_name
+
+    def context_name(self):
+        return self.__recurrent_name__ + ".end"
+
+    def to_proto_impl(self, **kwargs):
+        for l in self.__parents__:
+            RecurrentLayerGroupSetOutLink(l.name)
+        RecurrentLayerGroupEnd(name=self.__recurrent_name__)
+
+
 LayerV2 = Layer
 data = DataLayerV2
 AggregateLevel = conf_helps.layers.AggregateLevel
 ExpandLevel = conf_helps.layers.ExpandLevel
+memory = MemoryV2
 
 
 def __layer_name_mapping__(inname):
-    if inname in ['data_layer', 'memory', 'mixed_layer']:
+    if inname in ['data_layer', 'memory', 'mixed_layer', 'recurrent_group']:
         # Do Not handle these layers
         return
     elif inname == 'maxid_layer':
@@ -231,8 +410,10 @@ def __layer_name_mapping__(inname):
 def __layer_name_mapping_parent_names__(inname):
     all_args = getattr(conf_helps, inname).argspec.args
     return filter(
-        lambda x: x in ['input1', 'input2','label', 'input', 'a', 'b', 'expand_as',
-                        'weights', 'vectors', 'weight', 'score', 'left', 'right'],
+        lambda x: x in ['input1', 'input2', 'label', 'input', 'a', 'b',
+                        'expand_as',
+                        'weights', 'vectors', 'weight', 'score', 'left',
+                        'right', 'output_mem'],
         all_args)
 
 
@@ -267,3 +448,54 @@ operator_list = [
 for op in operator_list:
     globals()[op[0]] = __convert_to_v2__(
         op[0], parent_names=op[1], is_default_name=False)
+
+
+@wrap_name_default()
+def recurrent_group(step, input, name=None):
+    if not isinstance(input, collections.Sequence):
+        input = [input]
+
+    non_static_inputs = filter(lambda x: not isinstance(x, StaticInputV2),
+                               input)
+    actual_input = [
+        RecurrentLayerInput(
+            recurrent_name=name,
+            index=i,
+            parent_layers={'recurrent_inputs': non_static_inputs})
+        for i in xrange(len(non_static_inputs))
+    ]
+
+    def __real_step__(*args):
+        rnn_input = list(args)
+        static_inputs = filter(lambda x: isinstance(x, StaticInputV2), input)
+        for static_input in static_inputs:
+            mem_name = "__%s_memory__" % static_input.input.name
+            mem = memory(
+                name=mem_name,
+                is_seq=static_input.is_seq,
+                size=static_input.input.calculate_size,
+                boot_layer=static_input.input)
+            with mixed(
+                    name=mem_name,
+                    size=static_input.input.calculate_size,
+                    act=activation.Identity()) as mix:
+                mix += identity_projection(input=mem)
+            rnn_input.insert(input.index(static_input), mix)
+        return step(*rnn_input)
+
+    actual_output = __real_step__(*actual_input)
+
+    if not isinstance(actual_output, collections.Sequence):
+        actual_output = [actual_output]
+
+    retv = [
+        RecurrentLayerOutput(
+            recurrent_name=name,
+            index=i,
+            parent_layers={'recurrent_outputs': actual_output})
+        for i in xrange(len(actual_output))
+    ]
+    if len(retv) == 1:
+        return retv[0]
+    else:
+        return retv
diff --git a/python/paddle/v2/tests/CMakeLists.txt b/python/paddle/v2/tests/CMakeLists.txt
index 46b5d08b8761ea58530f5aefe5d1947408727f85..572deaff356712cac23cd7911cdf289db100564c 100644
--- a/python/paddle/v2/tests/CMakeLists.txt
+++ b/python/paddle/v2/tests/CMakeLists.txt
@@ -1,12 +1,16 @@
+add_test(NAME test_v2_api
+        COMMAND bash ${PROJ_ROOT}/python/paddle/v2/tests/run_tests.sh ${PYTHON_EXECUTABLE})
+
 add_test(NAME test_v2_layer
         COMMAND ${PROJ_ROOT}/paddle/.set_python_path.sh -d ${PROJ_ROOT}/python/
         ${PYTHON_EXECUTABLE} ${PROJ_ROOT}/python/paddle/v2/tests/test_layer.py
         WORKING_DIRECTORY ${PROJ_ROOT}/python/paddle)
 
-add_test(NAME test_v2_api
-        COMMAND bash ${PROJ_ROOT}/python/paddle/v2/tests/run_tests.sh ${PYTHON_EXECUTABLE})
+add_test(NAME test_v2_rnn_layer
+        COMMAND ${PROJ_ROOT}/paddle/.set_python_path.sh -d ${PROJ_ROOT}/python/
+        ${PYTHON_EXECUTABLE} ${PROJ_ROOT}/python/paddle/v2/tests/test_rnn_layer.py)
 
-add_test(NAME topology_test
+add_test(NAME test_topology
         COMMAND ${PROJ_ROOT}/paddle/.set_python_path.sh -d ${PROJ_ROOT}/python/
         ${PYTHON_EXECUTABLE} ${PROJ_ROOT}/python/paddle/v2/tests/test_topology.py
         WORKING_DIRECTORY ${PROJ_ROOT}/python/paddle)
diff --git a/python/paddle/v2/tests/test_rnn_layer.py b/python/paddle/v2/tests/test_rnn_layer.py
new file mode 100644
index 0000000000000000000000000000000000000000..5fbbd20eb76bb9daab2bcf98c4adad989106a377
--- /dev/null
+++ b/python/paddle/v2/tests/test_rnn_layer.py
@@ -0,0 +1,155 @@
+# Copyright PaddlePaddle contributors. All Rights Reserved
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import difflib
+import unittest
+
+import paddle.trainer_config_helpers as conf_helps
+import paddle.v2.activation as activation
+import paddle.v2.data_type as data_type
+import paddle.v2.layer as layer
+from paddle.trainer_config_helpers.config_parser_utils import \
+    parse_network_config as parse_network
+
+
+class RNNTest(unittest.TestCase):
+    def test_simple_rnn(self):
+        dict_dim = 10
+        word_dim = 8
+        hidden_dim = 8
+
+        def parse_old_rnn():
+            def step(y):
+                mem = conf_helps.memory(name="rnn_state", size=hidden_dim)
+                out = conf_helps.fc_layer(
+                    input=[y, mem],
+                    size=hidden_dim,
+                    act=activation.Tanh(),
+                    bias_attr=True,
+                    name="rnn_state")
+                return out
+
+            def test():
+                data = conf_helps.data_layer(name="word", size=dict_dim)
+                embd = conf_helps.embedding_layer(input=data, size=word_dim)
+                conf_helps.recurrent_group(name="rnn", step=step, input=embd)
+
+            return str(parse_network(test))
+
+        def parse_new_rnn():
+            def new_step(y):
+                mem = layer.memory(name="rnn_state", size=hidden_dim)
+                out = layer.fc(input=[y, mem],
+                               size=hidden_dim,
+                               act=activation.Tanh(),
+                               bias_attr=True,
+                               name="rnn_state")
+                return out
+
+            data = layer.data(
+                name="word", type=data_type.integer_value(dict_dim))
+            embd = layer.embedding(input=data, size=word_dim)
+            rnn_layer = layer.recurrent_group(
+                name="rnn", step=new_step, input=embd)
+            return str(layer.parse_network(rnn_layer))
+
+        diff = difflib.unified_diff(parse_old_rnn().splitlines(1),
+                                    parse_new_rnn().splitlines(1))
+        print ''.join(diff)
+
+    def test_sequence_rnn_multi_input(self):
+        dict_dim = 10
+        word_dim = 8
+        hidden_dim = 8
+        label_dim = 3
+
+        def parse_old_rnn():
+            def test():
+                data = conf_helps.data_layer(name="word", size=dict_dim)
+                label = conf_helps.data_layer(name="label", size=label_dim)
+                emb = conf_helps.embedding_layer(input=data, size=word_dim)
+                boot_layer = conf_helps.data_layer(name="boot", size=10)
+                boot_layer = conf_helps.fc_layer(
+                    name='boot_fc', input=boot_layer, size=10)
+
+                def step(y, wid):
+                    z = conf_helps.embedding_layer(input=wid, size=word_dim)
+                    mem = conf_helps.memory(
+                        name="rnn_state",
+                        size=hidden_dim,
+                        boot_layer=boot_layer)
+                    out = conf_helps.fc_layer(
+                        input=[y, z, mem],
+                        size=hidden_dim,
+                        act=conf_helps.TanhActivation(),
+                        bias_attr=True,
+                        name="rnn_state")
+                    return out
+
+                out = conf_helps.recurrent_group(
+                    name="rnn", step=step, input=[emb, data])
+
+                rep = conf_helps.last_seq(input=out)
+                prob = conf_helps.fc_layer(
+                    size=label_dim,
+                    input=rep,
+                    act=conf_helps.SoftmaxActivation(),
+                    bias_attr=True)
+
+                conf_helps.outputs(
+                    conf_helps.classification_cost(
+                        input=prob, label=label))
+
+            return str(parse_network(test))
+
+        def parse_new_rnn():
+            data = layer.data(
+                name="word", type=data_type.dense_vector(dict_dim))
+            label = layer.data(
+                name="label", type=data_type.dense_vector(label_dim))
+            emb = layer.embedding(input=data, size=word_dim)
+            boot_layer = layer.data(
+                name="boot", type=data_type.dense_vector(10))
+            boot_layer = layer.fc(name='boot_fc', input=boot_layer, size=10)
+
+            def step(y, wid):
+                z = layer.embedding(input=wid, size=word_dim)
+                mem = layer.memory(
+                    name="rnn_state", size=hidden_dim, boot_layer=boot_layer)
+                out = layer.fc(input=[y, z, mem],
+                               size=hidden_dim,
+                               act=activation.Tanh(),
+                               bias_attr=True,
+                               name="rnn_state")
+                return out
+
+            out = layer.recurrent_group(
+                name="rnn", step=step, input=[emb, data])
+
+            rep = layer.last_seq(input=out)
+            prob = layer.fc(size=label_dim,
+                            input=rep,
+                            act=activation.Softmax(),
+                            bias_attr=True)
+
+            cost = layer.classification_cost(input=prob, label=label)
+
+            return str(layer.parse_network(cost))
+
+        diff = difflib.unified_diff(parse_old_rnn().splitlines(1),
+                                    parse_new_rnn().splitlines(1))
+        print ''.join(diff)
+
+
+if __name__ == '__main__':
+    unittest.main()
diff --git a/python/paddle/v2/trainer.py b/python/paddle/v2/trainer.py
index 5003f55f3e0d15149d28d1478e0487d6873d6e0a..e743a49523ff21627ea2abfb76cee8b9ffd685e2 100644
--- a/python/paddle/v2/trainer.py
+++ b/python/paddle/v2/trainer.py
@@ -120,7 +120,8 @@ class SGD(ITrainer):
                     feeder(data_batch), out_args, pass_type)
                 self.__gradient_machine__.eval(pass_evaluator)
                 self.__gradient_machine__.eval(batch_evaluator)
-                for each_param in self.__gradient_machine__.getParameters():
+                for each_param in self.__gradient_machine__.getNonStaticParameters(
+                ):
                     updater.update(each_param)
                 # Get cost. We use numpy to calculate total cost for this batch.
                 cost_vec = out_args.getSlotValue(0)