Merge conflict with maxout layer

.gitignore

@@ -3,4 +3,6 @@ build/
 *.user
 
 .vscode
-.idea
+.idea
+.project
+.pydevproject

.travis.yml

@@ -2,9 +2,17 @@ language: cpp
 cache: ccache
 sudo: required
 dist: trusty
+os:
+  - linux
+  - osx
 env:
   - JOB=DOCS
   - JOB=BUILD_AND_TEST
+matrix:
+  exclude:
+    - os: osx
+      env: JOB=DOCS  # Only generate documentation in linux
+
 addons:
   apt:
     packages:
@@ -27,9 +35,11 @@ addons:
       - libgoogle-glog-dev
       - libgflags-dev
       - libgtest-dev
+      - graphviz
 before_install:
+  - if [[ "$TRAVIS_OS_NAME" == "linux" ]]; then sudo paddle/scripts/travis/before_install.linux.sh; fi
+  - if [[ "$TRAVIS_OS_NAME" == "osx" ]]; then paddle/scripts/travis/before_install.osx.sh; fi
   - pip install wheel protobuf sphinx breathe recommonmark
-  - sudo paddle/scripts/travis/before_install.sh
 script:
   - paddle/scripts/travis/main.sh
 notifications:

CMakeLists.txt

@@ -3,7 +3,7 @@ cmake_minimum_required(VERSION 2.8)
 project(paddle CXX C)
 set(PADDLE_MAJOR_VERSION 0)
 set(PADDLE_MINOR_VERSION 8)
-set(PADDLE_PATCH_VERSION 0b1)
+set(PADDLE_PATCH_VERSION 0b2)
 set(PADDLE_VERSION ${PADDLE_MAJOR_VERSION}.${PADDLE_MINOR_VERSION}.${PADDLE_PATCH_VERSION})
 
 set(CMAKE_MODULE_PATH ${CMAKE_MODULE_PATH} "${CMAKE_SOURCE_DIR}/cmake")
@@ -104,7 +104,7 @@ else()
 endif(NOT WITH_GPU)
 
 if(WITH_DOUBLE)
-    add_definitions(-DPADDLE_TYPE_DOUBLE -DHPPL_TYPE_DOUBLE)
+    add_definitions(-DPADDLE_TYPE_DOUBLE)
     set(ACCURACY double)
 else(WITH_DOUBLE)
     set(ACCURACY float)

cmake/cblas.cmake

@@ -17,10 +17,17 @@
 ## Find MKL First.
 set(MKL_ROOT $ENV{MKL_ROOT} CACHE PATH "Folder contains MKL")
 
-find_path(MKL_INCLUDE_DIR mkl.h PATHS ${MKL_ROOT}/include)
+find_path(MKL_INCLUDE_DIR mkl.h PATHS
-find_library(MKL_CORE_LIB NAMES mkl_core PATHS ${MKL_ROOT}/lib)
+  ${MKL_ROOT}/include)
-find_library(MKL_SEQUENTIAL_LIB NAMES mkl_sequential PATHS ${MKL_ROOT}/lib)
+find_library(MKL_CORE_LIB NAMES mkl_core PATHS
-find_library(MKL_INTEL_LP64 NAMES mkl_intel_lp64 PATHS ${MKL_ROOT}/lib)
+  ${MKL_ROOT}/lib
+  ${MKL_ROOT}/lib/intel64)
+find_library(MKL_SEQUENTIAL_LIB NAMES mkl_sequential PATHS
+  ${MKL_ROOT}/lib
+  ${MKL_ROOT}/lib/intel64)
+find_library(MKL_INTEL_LP64 NAMES mkl_intel_lp64 PATHS
+  ${MKL_ROOT}/lib
+  ${MKL_ROOT}/lib/intel64)
 
 
 if(MKL_INCLUDE_DIR AND MKL_CORE_LIB AND MKL_SEQUENTIAL_LIB AND MKL_INTEL_LP64)

cmake/flags.cmake

@@ -64,7 +64,9 @@ set(COMMON_FLAGS
     -Wdelete-non-virtual-dtor
     -Wno-unused-parameter
     -Wno-error=literal-suffix
-    -Wno-error=unused-local-typedefs)
+    -Wno-error=unused-local-typedefs
+    -Wno-error=unused-function  # Warnings in Numpy Header.
+)
 
 foreach(flag ${COMMON_FLAGS})
     safe_set_cflag(CMAKE_C_FLAGS ${flag})

cmake/util.cmake

@@ -184,3 +184,20 @@ macro(add_paddle_culib TARGET_NAME)
     cuda_add_library(${TARGET_NAME} STATIC ${ARGN})
     set(CUDA_NVCC_FLAGS ${NVCC_FLAG})
 endmacro()
+
+
+# Creates C resources file from files in given resource file
+function(create_resources res_file output)
+    # Create empty output file
+    file(WRITE ${output} "")
+    # Get short filename
+    string(REGEX MATCH "([^/]+)$" filename ${res_file})
+    # Replace filename spaces & extension separator for C compatibility
+    string(REGEX REPLACE "\\.| |-" "_" filename ${filename})
+    # Read hex data from file
+    file(READ ${res_file} filedata HEX)
+    # Convert hex data for C compatibility
+    string(REGEX REPLACE "([0-9a-f][0-9a-f])" "0x\\1," filedata ${filedata})
+    # Append data to output file
+    file(APPEND ${output} "const unsigned char ${filename}[] = {${filedata}};\nconst unsigned ${filename}_size = sizeof(${filename});\n")
+endfunction()

demo/mnist/.gitignore

@@ -0,0 +1,6 @@
+data/raw_data
+data/*.list
+mnist_vgg_model
+plot.png
+train.log
+*pyc

demo/mnist/data/generate_list.py

@@ -0,0 +1,21 @@
+# Copyright (c) 2016 Baidu, Inc. 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.
+
+o = open("./" + "train.list", "w")
+o.write("./data/raw_data/train" +"\n")
+o.close()
+
+o = open("./" + "test.list", "w")
+o.write("./data/raw_data/t10k" +"\n")
+o.close()

demo/mnist/data/get_mnist_data.sh

@@ -0,0 +1,22 @@
+#!/usr/bin/env sh
+# This scripts downloads the mnist data and unzips it.
+set -e
+DIR="$( cd "$(dirname "$0")" ; pwd -P )"
+rm -rf "$DIR/raw_data"
+mkdir "$DIR/raw_data"
+cd "$DIR/raw_data"
+
+echo "Downloading..."
+
+for fname in train-images-idx3-ubyte train-labels-idx1-ubyte t10k-images-idx3-ubyte t10k-labels-idx1-ubyte
+do
+    if [ ! -e $fname ]; then
+        wget --no-check-certificate http://yann.lecun.com/exdb/mnist/${fname}.gz
+        gunzip ${fname}.gz
+    fi
+done
+
+cd $DIR
+rm -f *.list
+python generate_list.py
+

demo/mnist/mnist_provider.py

@@ -0,0 +1,32 @@
+from paddle.trainer.PyDataProvider2 import *
+
+
+# Define a py data provider
+@provider(input_types={
+    'pixel': dense_vector(28 * 28),
+    'label': integer_value(10)
+})
+def process(settings, filename):  # settings is not used currently.
+    imgf = filename + "-images-idx3-ubyte"
+    labelf = filename + "-labels-idx1-ubyte"
+    f = open(imgf, "rb")
+    l = open(labelf, "rb")
+
+    f.read(16)
+    l.read(8)
+
+    # Define number of samples for train/test
+    if "train" in filename:
+        n = 60000
+    else:
+        n = 10000
+
+    for i in range(n):
+        label = ord(l.read(1))
+        pixels = []
+        for j in range(28 * 28):
+            pixels.append(float(ord(f.read(1))) / 255.0)
+        yield {"pixel": pixels, 'label': label}
+
+    f.close()
+    l.close()

demo/mnist/train.sh

@@ -0,0 +1,31 @@
+#!/bin/bash
+# Copyright (c) 2016 Baidu, Inc. 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.
+set -e
+config=vgg_16_mnist.py
+output=./mnist_vgg_model
+log=train.log
+
+paddle train \
+--config=$config \
+--dot_period=10 \
+--log_period=100 \
+--test_all_data_in_one_period=1 \
+--use_gpu=0 \
+--trainer_count=1 \
+--num_passes=100 \
+--save_dir=$output \
+2>&1 | tee $log
+
+python -m paddle.utils.plotcurve -i $log > plot.png

demo/mnist/vgg_16_mnist.py

@@ -0,0 +1,53 @@
+# Copyright (c) 2016 Baidu, Inc. 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 import *
+
+is_predict = get_config_arg("is_predict", bool, False)
+
+####################Data Configuration ##################
+
+
+if not is_predict:
+  data_dir='./data/'
+  define_py_data_sources2(train_list= data_dir + 'train.list',
+                        test_list= data_dir + 'test.list',
+                        module='mnist_provider',
+                        obj='process')
+
+######################Algorithm Configuration #############
+settings(
+    batch_size = 128,
+    learning_rate = 0.1 / 128.0,
+    learning_method = MomentumOptimizer(0.9),
+    regularization = L2Regularization(0.0005 * 128)
+)
+
+#######################Network Configuration #############
+
+data_size=1*28*28
+label_size=10
+img = data_layer(name='pixel', size=data_size)
+
+# small_vgg is predined in trainer_config_helpers.network
+predict = small_vgg(input_image=img,
+                    num_channels=1,
+                    num_classes=label_size)
+
+if not is_predict:
+    lbl = data_layer(name="label", size=label_size)
+    inputs(img, lbl)
+    outputs(classification_cost(input=predict, label=lbl))
+else:
+    outputs(predict)

demo/quick_start/preprocess.sh

@@ -20,6 +20,8 @@
 
 set -e
 
+export LC_ALL=C
+
 mkdir -p data/tmp
 python preprocess.py -i data/reviews_Electronics_5.json.gz
 # uniq and shuffle

demo/quick_start/train.sh

@@ -18,6 +18,8 @@ cfg=trainer_config.lr.py
 #cfg=trainer_config.emb.py
 #cfg=trainer_config.cnn.py
 #cfg=trainer_config.lstm.py
+#cfg=trainer_config.bidi-lstm.py
+#cfg=trainer_config.db-lstm.py
 paddle train \
   --config=$cfg \
   --save_dir=./output \

demo/quick_start/trainer_config.bidi-lstm.py

@@ -0,0 +1,62 @@
+# edit-mode: -*- python -*-
+
+# Copyright (c) 2016 Baidu, Inc. 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 import *
+
+dict_file = "./data/dict.txt"
+word_dict = dict()
+with open(dict_file, 'r') as f:
+    for i, line in enumerate(f):
+        w = line.strip().split()[0]
+        word_dict[w] = i
+
+is_predict = get_config_arg('is_predict', bool, False)
+trn = 'data/train.list' if not is_predict else None
+tst = 'data/test.list' if not is_predict else 'data/pred.list'
+process = 'process' if not is_predict else 'process_predict'
+define_py_data_sources2(train_list=trn,
+                        test_list=tst,
+                        module="dataprovider_emb",
+                        obj=process,
+                        args={"dictionary": word_dict})
+
+batch_size = 128 if not is_predict else 1
+settings(
+    batch_size=batch_size,
+    learning_rate=2e-3,
+    learning_method=AdamOptimizer(),
+    regularization=L2Regularization(8e-4),
+    gradient_clipping_threshold=25
+)
+
+bias_attr = ParamAttr(initial_std=0.,l2_rate=0.)
+data = data_layer(name="word", size=len(word_dict))
+emb = embedding_layer(input=data, size=128)
+
+bi_lstm = bidirectional_lstm(input=emb, size=128)
+dropout = dropout_layer(input=bi_lstm, dropout_rate=0.5)
+
+output = fc_layer(input=dropout, size=2,
+                  bias_attr=bias_attr,
+                  act=SoftmaxActivation())
+
+if is_predict:
+    maxid = maxid_layer(output)
+    outputs([maxid, output])
+else:
+    label = data_layer(name="label", size=2)
+    cls = classification_cost(input=output, label=label)
+    outputs(cls)

demo/quick_start/trainer_config.db-lstm.py

@@ -0,0 +1,73 @@
+# edit-mode: -*- python -*-
+
+# Copyright (c) 2016 Baidu, Inc. 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 import *
+
+dict_file = "./data/dict.txt"
+word_dict = dict()
+with open(dict_file, 'r') as f:
+    for i, line in enumerate(f):
+        w = line.strip().split()[0]
+        word_dict[w] = i
+
+is_predict = get_config_arg('is_predict', bool, False)
+trn = 'data/train.list' if not is_predict else None
+tst = 'data/test.list' if not is_predict else 'data/pred.list'
+process = 'process' if not is_predict else 'process_predict'
+define_py_data_sources2(train_list=trn,
+                        test_list=tst,
+                        module="dataprovider_emb",
+                        obj=process,
+                        args={"dictionary": word_dict})
+
+batch_size = 128 if not is_predict else 1
+settings(
+    batch_size=batch_size,
+    learning_rate=2e-3,
+    learning_method=AdamOptimizer(),
+    regularization=L2Regularization(8e-4),
+    gradient_clipping_threshold=25
+)
+
+bias_attr = ParamAttr(initial_std=0.,l2_rate=0.)
+
+data = data_layer(name="word", size=len(word_dict))
+emb = embedding_layer(input=data, size=128)
+
+hidden_0 = mixed_layer(size=128, input=[full_matrix_projection(input=emb)])
+lstm_0 = lstmemory(input=hidden_0, layer_attr=ExtraAttr(drop_rate=0.1))
+
+input_layers = [hidden_0, lstm_0]
+
+for i in range(1,8):
+    fc = fc_layer(input=input_layers, size=128)
+    lstm = lstmemory(input=fc, layer_attr=ExtraAttr(drop_rate=0.1),
+                    reverse=(i % 2) == 1,)
+    input_layers = [fc, lstm]
+
+lstm_last = pooling_layer(input=lstm, pooling_type=MaxPooling())
+
+output = fc_layer(input=lstm_last, size=2,
+                  bias_attr=bias_attr,
+                  act=SoftmaxActivation())
+
+if is_predict:
+    maxid = maxid_layer(output)
+    outputs([maxid, output])
+else:
+    label = data_layer(name="label", size=2)
+    cls = classification_cost(input=output, label=label)
+    outputs(cls)

demo/seqToseq/seqToseq_net.py

@@ -96,12 +96,12 @@ def gru_encoder_decoder(data_conf,
     encoded_vector = concat_layer(input=[src_forward, src_backward])
 
     with mixed_layer(size=decoder_size) as encoded_proj:
-        encoded_proj += full_matrix_projection(encoded_vector)
+        encoded_proj += full_matrix_projection(input=encoded_vector)
 
     backward_first = first_seq(input=src_backward)
     with mixed_layer(size=decoder_size,
                      act=TanhActivation(), ) as decoder_boot:
-        decoder_boot += full_matrix_projection(backward_first)
+        decoder_boot += full_matrix_projection(input=backward_first)
 
     def gru_decoder_with_attention(enc_vec, enc_proj, current_word):
         decoder_mem = memory(name='gru_decoder',
@@ -113,8 +113,8 @@ def gru_encoder_decoder(data_conf,
                                    decoder_state=decoder_mem, )
 
         with mixed_layer(size=decoder_size * 3) as decoder_inputs:
-            decoder_inputs += full_matrix_projection(context)
+            decoder_inputs += full_matrix_projection(input=context)
-            decoder_inputs += full_matrix_projection(current_word)
+            decoder_inputs += full_matrix_projection(input=current_word)
 
         gru_step = gru_step_layer(name='gru_decoder',
                                   input=decoder_inputs,

demo/sequence_tagging/data/get_data.sh

@@ -0,0 +1,21 @@
+#!/bin/bash
+# Copyright (c) 2016 Baidu, Inc. 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.
+set -e
+
+DIR="$( cd "$(dirname "$0")" ; pwd -P )"
+cd $DIR
+
+wget http://www.cnts.ua.ac.be/conll2000/chunking/train.txt.gz
+wget http://www.cnts.ua.ac.be/conll2000/chunking/test.txt.gz

demo/sequence_tagging/data/test.list

@@ -0,0 +1 @@
+data/test.txt.gz

demo/sequence_tagging/data/train.list

@@ -0,0 +1 @@
+data/train.txt.gz

demo/sequence_tagging/linear_crf.py

@@ -0,0 +1,84 @@
+# Copyright (c) 2016 Baidu, Inc. 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 import *
+
+import math
+
+define_py_data_sources2(train_list="data/train.list",
+                        test_list="data/test.list",
+                        module="dataprovider",
+                        obj="process")
+
+
+batch_size = 1
+settings(
+    learning_method=MomentumOptimizer(),
+    batch_size=batch_size,
+    regularization=L2Regularization(batch_size * 1e-4),
+    average_window=0.5,
+    learning_rate=1e-1,
+    learning_rate_decay_a=1e-5,
+    learning_rate_decay_b=0.25,
+)
+
+num_label_types=23
+
+def get_simd_size(size):
+    return int(math.ceil(float(size) / 8)) * 8
+
+# Currently, in order to use sparse_update=True,
+# the size has to be aligned.
+num_label_types = get_simd_size(num_label_types)
+
+features = data_layer(name="features", size=76328)
+word = data_layer(name="word", size=6778)
+pos = data_layer(name="pos", size=44)
+chunk = data_layer(name="chunk",
+                   size=num_label_types)
+
+crf_input = fc_layer(
+    input=features,
+    size=num_label_types,
+    act=LinearActivation(),
+    bias_attr=False,
+    param_attr=ParamAttr(initial_std=0, sparse_update=True))
+
+crf=crf_layer(
+    input=crf_input,
+    label=chunk,
+    param_attr=ParamAttr(name="crfw", initial_std=0),
+)
+
+crf_decoding=crf_decoding_layer(
+    size=num_label_types,
+    input=crf_input,
+    label=chunk,
+    param_attr=ParamAttr(name="crfw"),
+)
+
+sum_evaluator(
+    name="error",
+    input=crf_decoding,
+)
+
+chunk_evaluator(
+    name="chunk_f1",
+    input =[crf_decoding, chunk],
+    chunk_scheme="IOB",
+    num_chunk_types=11,
+)
+
+inputs(word, pos, chunk, features)
+outputs(crf)

demo/sequence_tagging/readme.md

@@ -0,0 +1,45 @@
+# Sequence Tagging
+
+This demo is a sequence model for assigning tags to each token in a sentence. The task is described at <a href = "http://www.cnts.ua.ac.be/conll2000/chunking">CONLL2000 Text Chunking</a> task.
+
+## Download data
+```bash
+cd demo/sequence_tagging
+./data/get_data.sh
+```
+
+## Train model
+```bash
+cd demo/sequence_tagging
+./train.sh
+```
+
+## Model description
+
+We provide two models. One is a linear CRF model (linear_crf.py) with is equivalent to the one at <a href="http://leon.bottou.org/projects/sgd#stochastic_gradient_crfs">leon.bottou.org/projects/sgd</a>. The second one is a stacked bidirectional RNN and CRF model (rnn_crf.py).
+<center>
+<table border="2" cellspacing="0" cellpadding="6" rules="all" frame="border">
+
+<thead>
+<th scope="col" class="left">Model name</th>
+<th scope="col" class="left">Number of parameters</th>
+<th scope="col" class="left">F1 score</th>
+</thead>
+
+<tbody>
+<tr>
+<td class="left">linear_crf</td>
+<td class="left"> 1.8M </td>
+<td class="left"> 0.937</td>
+</tr>
+
+<tr>
+<td class="left">rnn_crf</td>
+<td class="left"> 960K </td>
+<td class="left">0.941</td>
+</tr>
+
+</tbody>
+</table>
+</center>
+<br>

demo/sequence_tagging/rnn_crf.py

@@ -0,0 +1,130 @@
+# Copyright (c) 2016 Baidu, Inc. 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 import *
+
+import math
+
+define_py_data_sources2(train_list="data/train.list",
+                        test_list="data/test.list",
+                        module="dataprovider",
+                        obj="process")
+
+batch_size = 16
+settings(
+    learning_method=MomentumOptimizer(),
+    batch_size=batch_size,
+    regularization=L2Regularization(batch_size * 1e-5),
+    average_window=0.5,
+    learning_rate = 2e-3,
+    learning_rate_decay_a = 5e-7,
+    learning_rate_decay_b = 0.5,
+)
+
+word_dim=128
+hidden_dim = 128
+with_rnn = True
+
+initial_std=1/math.sqrt(hidden_dim)
+param_attr=ParamAttr(initial_std=initial_std)
+cpu_layer_attr=ExtraLayerAttribute(device=-1)
+
+default_device(0)
+
+num_label_types=23
+
+features = data_layer(name="features", size=76328)
+word = data_layer(name="word", size=6778)
+pos = data_layer(name="pos", size=44)
+chunk = data_layer(name="chunk",
+                   size=num_label_types,
+                   layer_attr=cpu_layer_attr)
+
+emb = embedding_layer(
+    input=word, size=word_dim, param_attr=ParamAttr(initial_std=0))
+
+hidden1 = mixed_layer(
+    size=hidden_dim,
+    act=STanhActivation(),
+    bias_attr=True,
+    input=[full_matrix_projection(emb),
+           table_projection(pos, param_attr=param_attr)]
+)
+
+if with_rnn:
+    rnn1 = recurrent_layer(
+        act=ReluActivation(),
+        bias_attr=True,
+        input=hidden1,
+        param_attr=ParamAttr(initial_std=0),
+    )
+
+hidden2 = mixed_layer(
+    size=hidden_dim,
+    act=STanhActivation(),
+    bias_attr=True,
+    input=[full_matrix_projection(hidden1)
+    ] + ([
+        full_matrix_projection(rnn1, param_attr=ParamAttr(initial_std=0))
+    ] if with_rnn else []),
+)
+
+if with_rnn:
+    rnn2=recurrent_layer(
+        reverse=True,
+        act=ReluActivation(),
+        bias_attr=True,
+        input=hidden2,
+        param_attr=ParamAttr(initial_std=0),
+    )
+
+crf_input = mixed_layer(
+    size=num_label_types,
+    bias_attr=False,
+    input=[
+        full_matrix_projection(hidden2),
+    ] + ([
+        full_matrix_projection(rnn2, param_attr=ParamAttr(initial_std=0))
+    ] if with_rnn else []),
+)
+
+crf = crf_layer(
+    input=crf_input,
+    label=chunk,
+    param_attr=ParamAttr(name="crfw", initial_std=0),
+    layer_attr=cpu_layer_attr,
+)
+
+crf_decoding = crf_decoding_layer(
+    size=num_label_types,
+    input=crf_input,
+    label=chunk,
+    param_attr=ParamAttr(name="crfw"),
+    layer_attr=cpu_layer_attr,
+)
+
+sum_evaluator(
+    name="error",
+    input=crf_decoding,
+)
+
+chunk_evaluator(
+    name="chunk_f1",
+    input =[crf_decoding, chunk],
+    chunk_scheme="IOB",
+    num_chunk_types=11,
+)
+
+inputs(word, pos, chunk, features)
+outputs(crf)

demo/sequence_tagging/train.sh

@@ -0,0 +1,10 @@
+#!/bin/bash
+
+paddle train \
+       --config rnn_crf.py \
+       --parallel_nn=1 \
+       --use_gpu=1 \
+       --dot_period=10 \
+       --log_period=1000 \
+       --test_period=0 \
+       --num_passes=10