Merge branch 'develop' of github.com:baidu/Paddle into feature/clean_mnist_v2

doc/api/trainer_config_helpers/layers.rst

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

python/paddle/trainer_config_helpers/layers.py

@@ -112,6 +112,8 @@ __all__ = [
     'priorbox_layer',
     'spp_layer',
     'pad_layer',
+    'eos_layer',
+    'layer_support',
 ]
 
 
@@ -708,6 +710,7 @@ class MixedLayerType(LayerOutput):
         # update the size which might be computed inside MixedLayer
         # according to the operator's output size
         self.size = ml.config.size
+        self.finalized = True
 
 
 @wrap_name_default("mixed")
@@ -1287,6 +1290,12 @@ def last_seq(input,
     """
     Get Last Timestamp Activation of a sequence.
 
+    The simple usage is:
+
+    .. code-block:: python
+
+       seq = last_seq(input=layer)
+
     :param agg_level: Aggregated level
     :param name: Layer name.
     :type name: basestring
@@ -1325,6 +1334,12 @@ def first_seq(input,
     """
     Get First Timestamp Activation of a sequence.
 
+    The simple usage is:
+
+    .. code-block:: python
+
+       seq = first_seq(input=layer)
+
     :param agg_level: aggregation level
     :param name: Layer name.
     :type name: basestring
@@ -1425,7 +1440,7 @@ def repeat_layer(input, num_repeats, name=None, layer_attr=None):
 
     .. code-block:: python
 
-       expand = repeat_layer(layer, 4)
+       expand = repeat_layer(input=layer, num_repeats=4)
 
     :param input: Input layer
     :type input: LayerOutput
@@ -1797,6 +1812,12 @@ def cos_sim(a, b, scale=1, size=1, name=None, layer_attr=None):
     Note that the above computation is for one sample. Multiple samples are
     processed in one batch.
 
+    The example usage is:
+
+    .. code-block:: python
+
+       cos = cos_sim(a=layer1, b=layer2, size=3)
+
     :param name: layer name
     :type name: basestring
     :param a: input layer a
@@ -1958,6 +1979,16 @@ def img_conv_layer(input,
     pieces. First 256/4 = 64 channels will process by first 32 filters. The
     rest channels will be processed by rest group of filters.
 
+    The example usage is:
+
+    ..  code-block:: python
+
+        conv = img_conv_layer(input=data, filter_size=1, filter_size_y=1,
+                              num_channels=8,
+                              num_filters=16, stride=1,
+                              bias_attr=False,
+                              act=ReluActivation())
+
     :param name: Layer name.
     :type name: basestring
     :param input: Layer Input.
@@ -2097,6 +2128,34 @@ def img_pool_layer(input,
 
     .. _pooling: http://ufldl.stanford.edu/tutorial/supervised/Pooling/
 
+    - ceil_mode=True:
+
+    ..  math::
+
+        w = 1 + int(ceil(input\_width + 2 * padding - pool\_size) / float(stride))
+        h = 1 + int(ceil(input\_height + 2 * padding\_y - pool\_size\_y) / float(stride\_y))
+
+    - ceil_mode=False:
+
+    ..  math::
+
+        w = 1 + int(floor(input\_width + 2 * padding - pool\_size) / float(stride))
+        h = 1 + int(floor(input\_height + 2 * padding\_y - pool\_size\_y) / float(stride\_y))
+
+    The example usage is:
+
+    ..  code-block:: python
+
+        maxpool = img_pool_layer(input=conv,
+                                 pool_size=3,
+                                 pool_size_y=5,
+                                 num_channels=8,
+                                 stride=1,
+                                 stride_y=2,
+                                 padding=1,
+                                 padding_y=2,
+                                 pool_type=MaxPooling())
+
     :param padding: pooling padding width.
     :type padding: int
     :param padding_y: pooling padding height. It's equal to padding by default.
@@ -2123,19 +2182,6 @@ def img_pool_layer(input,
     :param ceil_mode: Wether to use ceil mode to calculate output height and with.
                       Defalut is True. If set false, Otherwise use floor.
 
-                      - ceil_mode=True:
-
-                      ..  math::
-
-                          w = 1 + int(ceil(input_width + 2 * padding - pool_size) / float(stride))
-                          h = 1 + int(ceil(input_height + 2 * padding_y - pool_size_y) / float(stride_y))
-
-                      - ceil_mode=False:
-
-                      ..  math::
-
-                          w = 1 + int(floor(input_width + 2 * padding - pool_size) / float(stride))
-                          h = 1 + int(floor(input_height + 2 * padding_y - pool_size_y) / float(stride_y))
     :type ceil_mode: bool
     :return: LayerOutput object.
     :rtype: LayerOutput
@@ -2197,6 +2243,15 @@ def spp_layer(input,
     The details please refer to
     `Kaiming He's paper <https://arxiv.org/abs/1406.4729>`_.
 
+    The example usage is:
+
+    ..  code-block:: python
+
+        spp = spp_layer(input=data, 
+                        pyramid_height=2, 
+                        num_channels=16, 
+                        pool_type=MaxPooling())
+
     :param name: layer name.
     :type name: basestring
     :param input: layer's input.
@@ -2285,6 +2340,12 @@ def img_cmrnorm_layer(input,
     The details please refer to
     `Alex's paper <http://www.cs.toronto.edu/~fritz/absps/imagenet.pdf>`_.
 
+    The example usage is:
+
+    ..  code-block:: python
+    
+        norm = img_cmrnorm_layer(input=net, size=5)
+
     :param name: layer name.
     :type name: None|basestring
     :param input: layer's input.
@@ -2340,6 +2401,12 @@ def batch_norm_layer(input,
     The details of batch normalization please refer to this
     `paper <http://arxiv.org/abs/1502.03167>`_.
 
+    The example usage is:
+
+    ..  code-block:: python
+    
+        norm = batch_norm_layer(input=net, act=ReluActivation())
+
     :param name: layer name.
     :type name: basestring
     :param input: batch normalization input. Better be linear activation.
@@ -3903,13 +3970,13 @@ def conv_shift_layer(a, b, name=None, layer_attr=None):
 
     .. code-block:: python
 
-       conv_shift = conv_shift_layer(input=[layer1, layer2])
+       conv_shift = conv_shift_layer(a=layer1, b=layer2)
 
     :param name: layer name
     :type name: basestring
     :param a: Input layer a.
     :type a: LayerOutput
-    :param b: input layer b
+    :param b: input layer b.
     :type b: LayerOutput
     :param layer_attr: layer's extra attribute.
     :type layer_attr: ExtraLayerAttribute
@@ -4001,8 +4068,8 @@ def tensor_layer(a,
 @wrap_act_default()
 @layer_support()
 def selective_fc_layer(input,
-                       select,
                        size,
+                       select=None,
                        act=None,
                        name=None,
                        pass_generation=False,
@@ -4029,6 +4096,7 @@ def selective_fc_layer(input,
     :type input: LayerOutput|list|tuple
     :param select: The select layer. The output of select layer should be a
                    sparse binary matrix, and treat as the mask of selective fc.
+                   If is None, acts exactly like fc_layer.
     :type select: LayerOutput
     :param size: The layer dimension.
     :type size: int
@@ -4257,7 +4325,7 @@ def block_expand_layer(input,
 
     .. code-block:: python
 
-       block_expand = block_expand_layer(input,
+       block_expand = block_expand_layer(input=layer,
                                          num_channels=128,
                                          stride_x=1,
                                          stride_y=1,
@@ -4461,7 +4529,7 @@ def warp_ctc_layer(input,
         - You can set 'blank' to any value ranged in [0, num_classes], which
           should be consistent as that used in your labels.
         - As a native 'softmax' activation is interated to the warp-ctc library,
-         'linear' activation is expected instead in the 'input' layer.
+          'linear' activation is expected instead in the 'input' layer.
 
     The simple usage:
 
@@ -4594,6 +4662,13 @@ def crf_decoding_layer(input,
     this layer will also calculate error. output.value[i] is 1 for incorrect
     decoding or 0 for correct decoding.
 
+    The simple usage:
+
+    .. code-block:: python
+
+      crf_decoding = crf_decoding_layer(input=input,
+                                        size=label_dim)
+
     :param input: The first input layer.
     :type input: LayerOutput
     :param size: size of this layer.

python/paddle/v2/__init__.py

@@ -22,11 +22,12 @@ import data_feeder
 from . import dataset
 from . import reader
 import attr
+import pooling
 import py_paddle.swig_paddle as api
 
 __all__ = [
     'optimizer', 'layer', 'activation', 'parameters', 'init', 'trainer',
-    'event', 'data_type', 'attr', 'data_feeder', 'dataset', 'reader'
+    'event', 'data_type', 'attr', 'pooling', 'data_feeder', 'dataset', 'reader'
 ]
 
 

python/paddle/v2/pooling.py

@@ -0,0 +1,24 @@
+# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from paddle.trainer_config_helpers.poolings import *
+
+__all__ = ["Max", "CudnnMax", "Avg", "CudnnAvg", "Sum", "SquareRootN"]
+
+Max = MaxPooling
+CudnnMax = CudnnMaxPooling
+Avg = AvgPooling
+CudnnAvg = CudnnAvgPooling
+Sum = SumPooling
+SquareRootN = SquareRootNPooling

python/paddle/v2/tests/test_layer.py

@@ -19,18 +19,106 @@ import paddle.v2.activation as activation
 import paddle.v2.attr as attr
 import paddle.v2.data_type as data_type
 import paddle.v2.layer as layer
+import paddle.v2.pooling as pooling
 from paddle.trainer_config_helpers.config_parser_utils import \
     parse_network_config as parse_network
 
-pixel = layer.data(name='pixel', type=data_type.dense_vector(784))
+pixel = layer.data(name='pixel', type=data_type.dense_vector(128))
 label = layer.data(name='label', type=data_type.integer_value(10))
 weight = layer.data(name='weight', type=data_type.dense_vector(10))
 score = layer.data(name='score', type=data_type.dense_vector(1))
+
 hidden = layer.fc(input=pixel,
                   size=100,
                   act=activation.Sigmoid(),
                   param_attr=attr.Param(name='hidden'))
 inference = layer.fc(input=hidden, size=10, act=activation.Softmax())
+conv = layer.img_conv(
+    input=pixel,
+    filter_size=1,
+    filter_size_y=1,
+    num_channels=8,
+    num_filters=16,
+    act=activation.Linear())
+
+
+class ImageLayerTest(unittest.TestCase):
+    def test_conv_layer(self):
+        conv_shift = layer.conv_shift(a=pixel, b=score)
+        print layer.parse_network(conv, conv_shift)
+
+    def test_pooling_layer(self):
+        maxpool = layer.img_pool(
+            input=conv,
+            pool_size=2,
+            num_channels=16,
+            padding=1,
+            pool_type=pooling.Max())
+        spp = layer.spp(input=conv,
+                        pyramid_height=2,
+                        num_channels=16,
+                        pool_type=pooling.Max())
+        maxout = layer.maxout(input=conv, num_channels=16, groups=4)
+        print layer.parse_network(maxpool, spp, maxout)
+
+    def test_norm_layer(self):
+        norm1 = layer.img_cmrnorm(input=conv, size=5)
+        norm2 = layer.batch_norm(input=conv)
+        norm3 = layer.sum_to_one_norm(input=conv)
+        print layer.parse_network(norm1, norm2, norm3)
+
+
+class AggregateLayerTest(unittest.TestCase):
+    def test_aggregate_layer(self):
+        pool = layer.pool(
+            input=pixel,
+            pooling_type=pooling.Avg(),
+            agg_level=layer.AggregateLevel.EACH_SEQUENCE)
+        last_seq = layer.last_seq(input=pixel)
+        first_seq = layer.first_seq(input=pixel)
+        concat = layer.concat(input=[last_seq, first_seq])
+        seq_concat = layer.seq_concat(a=last_seq, b=first_seq)
+        print layer.parse_network(pool, last_seq, first_seq, concat, seq_concat)
+
+
+class MathLayerTest(unittest.TestCase):
+    def test_math_layer(self):
+        addto = layer.addto(input=[pixel, pixel])
+        linear_comb = layer.linear_comb(weights=weight, vectors=hidden, size=10)
+        interpolation = layer.interpolation(
+            input=[hidden, hidden], weight=score)
+        bilinear = layer.bilinear_interp(input=conv, out_size_x=4, out_size_y=4)
+        power = layer.power(input=pixel, weight=score)
+        scaling = layer.scaling(input=pixel, weight=score)
+        slope = layer.slope_intercept(input=pixel)
+        tensor = layer.tensor(a=pixel, b=pixel, size=1000)
+        cos_sim = layer.cos_sim(a=pixel, b=pixel)
+        trans = layer.trans(input=tensor)
+        print layer.parse_network(addto, linear_comb, interpolation, power,
+                                  scaling, slope, tensor, cos_sim, trans)
+
+
+class ReshapeLayerTest(unittest.TestCase):
+    def test_reshape_layer(self):
+        block_expand = layer.block_expand(
+            input=conv, num_channels=4, stride_x=1, block_x=1)
+        expand = layer.expand(
+            input=weight,
+            expand_as=pixel,
+            expand_level=layer.ExpandLevel.FROM_TIMESTEP)
+        repeat = layer.repeat(input=pixel, num_repeats=4)
+        reshape = layer.seq_reshape(input=pixel, reshape_size=4)
+        rotate = layer.rotate(input=pixel, height=16, width=49)
+        print layer.parse_network(block_expand, expand, repeat, reshape, rotate)
+
+
+class RecurrentLayerTest(unittest.TestCase):
+    def test_recurrent_layer(self):
+        word = layer.data(name='word', type=data_type.integer_value(12))
+        recurrent = layer.recurrent(input=word)
+        lstm = layer.lstmemory(input=word)
+        gru = layer.grumemory(input=word)
+        print layer.parse_network(recurrent, lstm, gru)
 
 
 class CostLayerTest(unittest.TestCase):
@@ -51,12 +139,120 @@ class CostLayerTest(unittest.TestCase):
         cost10 = layer.sum_cost(input=inference)
         cost11 = layer.huber_cost(input=score, label=label)
 
-        print dir(layer)
-        layer.parse_network(cost1, cost2)
-        print dir(layer)
-        #print layer.parse_network(cost3, cost4)
-        #print layer.parse_network(cost5, cost6)
-        #print layer.parse_network(cost7, cost8, cost9, cost10, cost11)
+        print layer.parse_network(cost1, cost2)
+        print layer.parse_network(cost3, cost4)
+        print layer.parse_network(cost5, cost6)
+        print layer.parse_network(cost7, cost8, cost9, cost10, cost11)
+
+        crf = layer.crf(input=inference, label=label)
+        crf_decoding = layer.crf_decoding(input=inference, size=3)
+        ctc = layer.ctc(input=inference, label=label)
+        warp_ctc = layer.warp_ctc(input=pixel, label=label)
+        nce = layer.nce(input=inference, label=label, num_classes=3)
+        hsigmoid = layer.hsigmoid(input=inference, label=label, num_classes=3)
+
+        print layer.parse_network(crf, crf_decoding, ctc, warp_ctc, nce,
+                                  hsigmoid)
+
+
+class OtherLayerTest(unittest.TestCase):
+    def test_sampling_layer(self):
+        maxid = layer.max_id(input=inference)
+        sampling_id = layer.sampling_id(input=inference)
+        eos = layer.eos(input=maxid, eos_id=5)
+        print layer.parse_network(maxid, sampling_id, eos)
+
+    def test_slicing_joining_layer(self):
+        pad = layer.pad(input=conv, pad_c=[2, 3], pad_h=[1, 2], pad_w=[3, 1])
+        print layer.parse_network(pad)
+
+
+class ProjOpTest(unittest.TestCase):
+    def test_projection(self):
+        input = layer.data(name='data', type=data_type.dense_vector(784))
+        word = layer.data(
+            name='word', type=data_type.integer_value_sequence(10000))
+        fc0 = layer.fc(input=input, size=100, act=activation.Sigmoid())
+        fc1 = layer.fc(input=input, size=200, act=activation.Sigmoid())
+        mixed0 = layer.mixed(
+            size=256,
+            input=[
+                layer.full_matrix_projection(input=fc0),
+                layer.full_matrix_projection(input=fc1)
+            ])
+        with layer.mixed(size=200) as mixed1:
+            mixed1 += layer.full_matrix_projection(input=fc0)
+            mixed1 += layer.identity_projection(input=fc1)
+
+        table = layer.table_projection(input=word)
+        emb0 = layer.mixed(size=512, input=table)
+        with layer.mixed(size=512) as emb1:
+            emb1 += table
+
+        scale = layer.scaling_projection(input=fc0)
+        scale0 = layer.mixed(size=100, input=scale)
+        with layer.mixed(size=100) as scale1:
+            scale1 += scale
+
+        dotmul = layer.dotmul_projection(input=fc0)
+        dotmul0 = layer.mixed(size=100, input=dotmul)
+        with layer.mixed(size=100) as dotmul1:
+            dotmul1 += dotmul
+
+        context = layer.context_projection(input=fc0, context_len=5)
+        context0 = layer.mixed(size=100, input=context)
+        with layer.mixed(size=100) as context1:
+            context1 += context
+
+        conv = layer.conv_projection(
+            input=input,
+            filter_size=1,
+            num_channels=1,
+            num_filters=128,
+            stride=1,
+            padding=0)
+        conv0 = layer.mixed(input=conv, bias_attr=True)
+        with layer.mixed(bias_attr=True) as conv1:
+            conv1 += conv
+
+        print layer.parse_network(mixed0)
+        print layer.parse_network(mixed1)
+        print layer.parse_network(emb0)
+        print layer.parse_network(emb1)
+        print layer.parse_network(scale0)
+        print layer.parse_network(scale1)
+        print layer.parse_network(dotmul0)
+        print layer.parse_network(dotmul1)
+        print layer.parse_network(conv0)
+        print layer.parse_network(conv1)
+
+    def test_operator(self):
+        ipt0 = layer.data(name='data', type=data_type.dense_vector(784))
+        ipt1 = layer.data(name='word', type=data_type.dense_vector(128))
+        fc0 = layer.fc(input=ipt0, size=100, act=activation.Sigmoid())
+        fc1 = layer.fc(input=ipt0, size=100, act=activation.Sigmoid())
+
+        dotmul_op = layer.dotmul_operator(a=fc0, b=fc1)
+        dotmul0 = layer.mixed(input=dotmul_op)
+        with layer.mixed() as dotmul1:
+            dotmul1 += dotmul_op
+
+        conv = layer.conv_operator(
+            img=ipt0,
+            filter=ipt1,
+            filter_size=1,
+            num_channels=1,
+            num_filters=128,
+            stride=1,
+            padding=0)
+        conv0 = layer.mixed(input=conv)
+        with layer.mixed() as conv1:
+            conv1 += conv
+
+        print layer.parse_network(dotmul0)
+        print layer.parse_network(dotmul1)
+        print layer.parse_network(conv0)
+        print layer.parse_network(conv1)
 
 
 if __name__ == '__main__':