diff --git a/RELEASE.md b/RELEASE.md index f1588cb15bb8d1f79732d6b5195775cb13acef54..146f7afa7dfbc152500b82fde28445ae3155c16c 100644 --- a/RELEASE.md +++ b/RELEASE.md @@ -1,11 +1,103 @@ # Release v0.10.0 +We are glad to release version 0.10.0. In this version, we are happy to release the new +[Python API](http://research.baidu.com/paddlepaddles-new-api-simplifies-deep-learning-programs/). + +- Our old Python API is kind of out of date. It's hard to learn and hard to + use. To write a PaddlePaddle program using the old API, we'd have to write + at least two Python files: one `data provider` and another one that defines + the network topology. Users start a PaddlePaddle job by running the + `paddle_trainer` C++ program, which calls Python interpreter to run the + network topology configuration script and then start the training loop, + which iteratively calls the data provider function to load minibatches. + This prevents us from writing a Python program in a modern way, e.g., in the + Jupyter Notebook. + +- The new API, which we often refer to as the *v2 API*, allows us to write + much shorter Python programs to define the network and the data in a single + .py file. Also, this program can run in Jupyter Notebook, since the entry + point is in Python program and PaddlePaddle runs as a shared library loaded + and invoked by this Python program. + +Basing on the new API, we delivered an online interative +book, [Deep Learning 101](http://book.paddlepaddle.org/index.en.html) +and [its Chinese version](http://book.paddlepaddle.org/). + +We also worked on updating our online documentation to describe the new API. +But this is an ongoing work. We will release more documentation improvements +in the next version. + +We also worked on bring the new API to distributed model training (via MPI and +Kubernetes). This work is ongoing. We will release more about it in the next +version. + ## New Features +* We release [new Python API](http://research.baidu.com/paddlepaddles-new-api-simplifies-deep-learning-programs/). +* Deep Learning 101 book in [English](http://book.paddlepaddle.org/index.en.html) and [Chinese](http://book.paddlepaddle.org/). +* Support rectangle input for CNN. +* Support stride pooling for seqlastin and seqfirstin. +* Expose `seq_concat_layer/seq_reshape_layer` in `trainer_config_helpers`. +* Add dataset package: CIFAR, MNIST, IMDB, WMT14, CONLL05, movielens, imikolov. +* Add Priorbox layer for Single Shot Multibox Detection. +* Add smooth L1 cost. +* Add data reader creator and data reader decorator for v2 API. +* Add the CPU implementation of cmrnorm projection. + ## Improvements +* Support Python virtualenv for `paddle_trainer`. +* Add pre-commit hooks, used for automatically format our code. +* Upgrade protobuf to version 3.x. +* Add an option to check data type in Python data provider. +* Speedup the backward of average layer on GPU. +* Documentation refinement. +* Check dead links in documents using Travis-CI. +* Add a example for explaining `sparse_vector`. +* Add ReLU in layer_math.py +* Simplify data processing flow for Quick Start. +* Support CUDNN Deconv. +* Add data feeder in v2 API. +* Support predicting the samples from sys.stdin for sentiment demo. +* Provide multi-proccess interface for image preprocessing. +* Add benchmark document for v1 API. +* Add ReLU in `layer_math.py`. +* Add packages for automatically downloading public datasets. +* Rename `Argument::sumCost` to `Argument::sum` since class `Argument` is nothing with cost. +* Expose Argument::sum to Python +* Add a new `TensorExpression` implementation for matrix-related expression evaluations. +* Add lazy assignment for optimizing the calculation of a batch of multiple expressions. +* Add abstract calss `Function` and its implementation: + * `PadFunc` and `PadGradFunc`. + * `ContextProjectionForwardFunc` and `ContextProjectionBackwardFunc`. + * `CosSimBackward` and `CosSimBackwardFunc`. + * `CrossMapNormalFunc` and `CrossMapNormalGradFunc`. + * `MulFunc`. +* Add class `AutoCompare` and `FunctionCompare`, which make it easier to write unit tests for comparing gpu and cpu version of a function. +* Generate `libpaddle_test_main.a` and remove the main function inside the test file. +* Support dense numpy vector in PyDataProvider2. +* Clean code base, remove some copy-n-pasted code snippets: + * Extract `RowBuffer` class for `SparseRowMatrix`. + * Clean the interface of `GradientMachine`. + * Use `override` keyword in layer. + * Simplify `Evaluator::create`, use `ClassRegister` to create `Evaluator`s. +* Check MD5 checksum when downloading demo's dataset. +* Add `paddle::Error` which intentially replace `LOG(FATAL)` in Paddle. + ## Bug Fixes +* Check layer input types for `recurrent_group`. +* Don't run `clang-format` with .cu source files. +* Fix bugs with `LogActivation`. +* Fix the bug that runs `test_layerHelpers` multiple times. +* Fix the bug that the seq2seq demo exceeds protobuf message size limit. +* Fix the bug in dataprovider converter in GPU mode. +* Fix a bug in `GatedRecurrentLayer`. +* Fix bug for `BatchNorm` when testing more than one models. +* Fix broken unit test of paramRelu. +* Fix some compile-time warnings about `CpuSparseMatrix`. +* Fix `MultiGradientMachine` error when `trainer_count > batch_size`. +* Fix bugs that prevents from asynchronous data loading in `PyDataProvider2`. # Release v0.9.0 diff --git a/cmake/package.cmake b/cmake/package.cmake index 211593f358eb34cf1a5692697247511893dfeb93..ff49a2d08e8f6004320acfce266339aa301eb9c4 100644 --- a/cmake/package.cmake +++ b/cmake/package.cmake @@ -1,5 +1,4 @@ set(CPACK_PACKAGE_NAME paddle) -set(CPACK_PACKAGE_DESCRIPTION_SUMMARY "") set(CPACK_PACKAGE_VERSION_MAJOR ${PADDLE_MAJOR_VERSION}) set(CPACK_PACKAGE_VERSION_MINOR ${PADDLE_MINOR_VERSION}) set(CPACK_PACKAGE_VERSION_PATCH ${PADDLE_PATCH_VERSION}) @@ -10,8 +9,9 @@ set(CPACK_DEBIAN_PACKAGE_ARCHITECTURE amd64) set(CPACK_DEBIAN_PACKAGE_MAINTAINER PaddlePaddle Dev ) set(CPACK_PACKAGE_DESCRIPTION_SUMMARY "Paddle") set(CPACK_PACKAGE_DESCRIPTION "") -set(CPACK_DEBIAN_PACKAGE_DEPENDS "libatlas3-base, libgflags2, libgoogle-glog0, libprotobuf8, libpython2.7, libstdc++6, python-numpy, python-pip, python-pip-whl, python-protobuf") +set(CPACK_DEBIAN_PACKAGE_DEPENDS "libpython2.7-dev, libstdc++6, python-pip, curl, libgfortran3, python-pip-whl") set(CPACK_DEBIAN_PACKAGE_SECTION Devel) +set(CPACK_DEBIAN_PACKAGE_VERSION ${PADDLE_VERSION}) set(CPACK_DEBIAN_PACKAGE_CONTROL_EXTRA "${PROJ_ROOT}/paddle/scripts/deb/postinst") #set(CPACK_GENERATOR "DEB") # Start cpack diff --git a/demo/sentiment/trainer_config.py b/demo/sentiment/trainer_config.py index 2defecd178262900c03c1eda60b351dc44629d1f..f1cadaa728ac58107e15f77b5994d31da088caf7 100644 --- a/demo/sentiment/trainer_config.py +++ b/demo/sentiment/trainer_config.py @@ -29,7 +29,7 @@ settings( batch_size=128, learning_rate=2e-3, learning_method=AdamOptimizer(), - average_window=0.5, + model_average=ModelAverage(0.5), regularization=L2Regularization(8e-4), gradient_clipping_threshold=25) diff --git a/demo/seqToseq/seqToseq_net.py b/demo/seqToseq/seqToseq_net.py index e523a34d5a95120d1f0a583be8bbdbff5678d1ab..3d1f86ec3b7eda4fceaf3a1e406e3d0a1a4a2f60 100644 --- a/demo/seqToseq/seqToseq_net.py +++ b/demo/seqToseq/seqToseq_net.py @@ -69,7 +69,8 @@ def gru_encoder_decoder(data_conf, encoder_size=512, decoder_size=512, beam_size=3, - max_length=250): + max_length=250, + error_clipping=50): """ A wrapper for an attention version of GRU Encoder-Decoder network is_generating: whether this config is used for generating @@ -90,9 +91,19 @@ def gru_encoder_decoder(data_conf, input=src_word_id, size=word_vector_dim, param_attr=ParamAttr(name='_source_language_embedding')) - src_forward = simple_gru(input=src_embedding, size=encoder_size) + src_forward = simple_gru( + input=src_embedding, + size=encoder_size, + naive=True, + gru_layer_attr=ExtraLayerAttribute( + error_clipping_threshold=error_clipping)) src_backward = simple_gru( - input=src_embedding, size=encoder_size, reverse=True) + input=src_embedding, + size=encoder_size, + reverse=True, + naive=True, + gru_layer_attr=ExtraLayerAttribute( + error_clipping_threshold=error_clipping)) encoded_vector = concat_layer(input=[src_forward, src_backward]) with mixed_layer(size=decoder_size) as encoded_proj: @@ -117,11 +128,13 @@ def gru_encoder_decoder(data_conf, decoder_inputs += full_matrix_projection(input=context) decoder_inputs += full_matrix_projection(input=current_word) - gru_step = gru_step_layer( + gru_step = gru_step_naive_layer( name='gru_decoder', input=decoder_inputs, output_mem=decoder_mem, - size=decoder_size) + size=decoder_size, + layer_attr=ExtraLayerAttribute( + error_clipping_threshold=error_clipping)) with mixed_layer( size=target_dict_dim, bias_attr=True, diff --git a/doc/getstarted/index_cn.rst b/doc/getstarted/index_cn.rst index c6a4d3121c5857cd434acecb389d68f4d4c7a532..cadf092f8f42ca16bbeb23bd21e0d018af8e43cc 100644 --- a/doc/getstarted/index_cn.rst +++ b/doc/getstarted/index_cn.rst @@ -2,7 +2,8 @@ ============ .. toctree:: - :maxdepth: 2 + :maxdepth: 1 build_and_install/index_cn.rst - basic_usage/index_cn.rst + +- `深度学习入门课程 `_ diff --git a/doc/getstarted/index_en.rst b/doc/getstarted/index_en.rst index 55d95d8015e56ddae3363d19315db0fad841caad..9f771e93e8b63eb98e31ec12667bd1aa007af20e 100644 --- a/doc/getstarted/index_en.rst +++ b/doc/getstarted/index_en.rst @@ -2,7 +2,8 @@ GET STARTED ============ .. toctree:: - :maxdepth: 2 + :maxdepth: 1 build_and_install/index_en.rst - basic_usage/index_en.rst + +- `Deep Learning 101 `_ diff --git a/doc/howto/deep_model/rnn/hierarchical_layer_cn.rst b/doc/howto/deep_model/rnn/hierarchical_layer_cn.rst index 4b328fc9d38bc5dfec35d5e0f0d46136aeeb41bc..79048e92482851af6c2dd7d055868ebcaa7a298b 100644 --- a/doc/howto/deep_model/rnn/hierarchical_layer_cn.rst +++ b/doc/howto/deep_model/rnn/hierarchical_layer_cn.rst @@ -19,18 +19,18 @@ 在 PaddlePaddle中，下面这些Layer能够接受双层序列作为输入，完成相应的计算。 -pooling_layer -============== +pooling +======== -pooling_layer 的使用示例如下，详细见 :ref:`api_trainer_config_helpers_layers_pooling_layer` 配置API。 +pooling 的使用示例如下，详细见 :ref:`api_v2.layer_pooling` 配置API。 .. code-block:: bash - seq_pool = pooling_layer(input=layer, - pooling_type=AvgPooling(), - agg_level=AggregateLevel.EACH_SEQUENCE) + seq_pool = pooling(input=layer, + pooling_type=pooling.Max(), + agg_level=AggregateLevel.EACH_SEQUENCE) -- `pooling_type` 目前支持两种，分别是：MaxPooling()和AvgPooling()。 +- `pooling_type` 目前支持两种，分别是：pooling.Max()和pooling.Avg()。 - `agg_level=AggregateLevel.EACH_TIMESTEP` 时（默认值）： @@ -47,7 +47,7 @@ pooling_layer 的使用示例如下，详细见 :ref:`api_trainer_config_helpers last_seq 和 first_seq ===================== -last_seq 的使用示例如下（ :ref:`api_trainer_config_helpers_layers_first_seq` 类似），详细见 :ref:`api_trainer_config_helpers_layers_last_seq` 配置API。 +last_seq 的使用示例如下（ :ref:`api_v2.layer_first_seq` 类似），详细见 :ref:`api_v2.layer_last_seq` 配置API。 .. code-block:: bash @@ -65,16 +65,16 @@ last_seq 的使用示例如下（ :ref:`api_trainer_config_helpers_layers_first_ - 输入：必须是一个双层序列 - 输出：一个单层序列，其中每个元素是双层序列中每个subseq最后一个（或第一个）元素。 -expand_layer -============ +expand +====== -expand_layer 的使用示例如下，详细见 :ref:`api_trainer_config_helpers_layers_expand_layer` 配置API。 +expand 的使用示例如下，详细见 :ref:`api_v2.layer_expand` 配置API。 .. code-block:: bash - expand = expand_layer(input=layer1, - expand_as=layer2, - expand_level=ExpandLevel.FROM_TIMESTEP) + ex = expand(input=layer1, + expand_as=layer2, + expand_level=ExpandLevel.FROM_TIMESTEP) - `expand_level=ExpandLevel.FROM_TIMESTEP` 时（默认值）： diff --git a/doc/howto/deep_model/rnn/index_cn.rst b/doc/howto/deep_model/rnn/index_cn.rst index 9ecab5594cff47cde4700b7ce0f58013a960a16e..9e805ca85191b793c8798a239927a318c70b96f5 100644 --- a/doc/howto/deep_model/rnn/index_cn.rst +++ b/doc/howto/deep_model/rnn/index_cn.rst @@ -4,7 +4,6 @@ RNN相关模型 .. toctree:: :maxdepth: 1 - rnn_config_cn.rst recurrent_group_cn.md hierarchical_layer_cn.rst hrnn_rnn_api_compare_cn.rst diff --git a/doc/howto/deep_model/rnn/index_en.rst b/doc/howto/deep_model/rnn/index_en.rst index 7adc79873d699fdfd5a85034bcef964dd1f19132..13a153b05c578e0af82ee29db5ea27fd4b6d6f59 100644 --- a/doc/howto/deep_model/rnn/index_en.rst +++ b/doc/howto/deep_model/rnn/index_en.rst @@ -1,7 +1,2 @@ RNN Models ========== - -.. toctree:: - :maxdepth: 1 - - rnn_config_en.rst diff --git a/doc/index_cn.rst b/doc/index_cn.rst index 460fedb5658a8ea9bbe8b602ee2b5df66502fa62..9279bac7f4b2898c18979630a8d6dfcb2dba70e0 100644 --- a/doc/index_cn.rst +++ b/doc/index_cn.rst @@ -5,7 +5,6 @@ PaddlePaddle 文档 :maxdepth: 1 getstarted/index_cn.rst - tutorials/index_cn.md howto/index_cn.rst api/index_cn.rst faq/index_cn.rst diff --git a/doc/index_en.rst b/doc/index_en.rst index 1d9cca7de720ebc23fe816f32d158930d91c07e7..168c7667c61da677905585d6c4b5037ce80b3765 100644 --- a/doc/index_en.rst +++ b/doc/index_en.rst @@ -5,8 +5,6 @@ PaddlePaddle Documentation :maxdepth: 1 getstarted/index_en.rst - tutorials/index_en.md howto/index_en.rst api/index_en.rst about/index_en.rst - \ No newline at end of file diff --git a/doc_theme/templates/layout.html b/doc_theme/templates/layout.html index 034740369ed10a748856e2205d3315f51a7de62f..65e61c5f298e19adc6330c378779a6edf418752e 100644 --- a/doc_theme/templates/layout.html +++ b/doc_theme/templates/layout.html @@ -114,10 +114,7 @@
@@ -137,7 +134,7 @@ {{ toctree }} {% endblock %} - {% if toc %} + {% if False %} {% endif %}
@@ -168,7 +165,8 @@ VERSION:'{{ release|e }}', COLLAPSE_INDEX:false, FILE_SUFFIX:'{{ '' if no_search_suffix else file_suffix }}', - HAS_SOURCE: {{ has_source|lower }} + HAS_SOURCE: {{ has_source|lower }}, + SOURCELINK_SUFFIX: ".txt", }; {%- for scriptfile in script_files %} diff --git a/paddle/function/CMakeLists.txt b/paddle/function/CMakeLists.txt index 1522510e8bb9816cb468fcf406e22560163950cc..233a53709a80f06dd2a06995b159c1aef10e2788 100644 --- a/paddle/function/CMakeLists.txt +++ b/paddle/function/CMakeLists.txt @@ -12,7 +12,7 @@ endif() add_library(paddle_function STATIC ${cpp_files} ${cu_objs}) add_dependencies(paddle_function ${external_project_dependencies}) - +add_dependencies(paddle_function gen_proto_cpp) if(WITH_GPU) if(WITH_TESTING) diff --git a/paddle/gserver/tests/sequence_layer_group.conf b/paddle/gserver/tests/sequence_layer_group.conf index 68d150d553588c864de56ce1e6f283cc42fbbf2f..50f2d89d0271b2eaa460e57636eb09b6d6aeda18 100644 --- a/paddle/gserver/tests/sequence_layer_group.conf +++ b/paddle/gserver/tests/sequence_layer_group.conf @@ -48,8 +48,7 @@ lstm = lstmemory_group( size=hidden_dim, act=TanhActivation(), gate_act=SigmoidActivation(), - state_act=TanhActivation(), - lstm_layer_attr=ExtraLayerAttribute(error_clipping_threshold=50)) + state_act=TanhActivation()) lstm_last = last_seq(input=lstm) diff --git a/paddle/gserver/tests/sequence_nest_layer_group.conf b/paddle/gserver/tests/sequence_nest_layer_group.conf index 88cb42798baff79fa6a86ef11dabf1781575c0b4..c01b95f7a29ae73c2b3ccd5b56ad1d316cbc72ec 100644 --- a/paddle/gserver/tests/sequence_nest_layer_group.conf +++ b/paddle/gserver/tests/sequence_nest_layer_group.conf @@ -51,8 +51,7 @@ def lstm_group(lstm_group_input): size=hidden_dim, act=TanhActivation(), gate_act=SigmoidActivation(), - state_act=TanhActivation(), - lstm_layer_attr=ExtraLayerAttribute(error_clipping_threshold=50)) + state_act=TanhActivation()) return lstm_output diff --git a/paddle/scripts/deb/postinst b/paddle/scripts/deb/postinst new file mode 100644 index 0000000000000000000000000000000000000000..91620b1ee7569cd17927f44112dfa9279ddbdd32 --- /dev/null +++ b/paddle/scripts/deb/postinst @@ -0,0 +1,6 @@ +#!/bin/bash +set -e +echo "Post install paddle debian package." +echo "Install some python package used for paddle. You can run " +echo " pip install /usr/opt/paddle/share/wheels/*.whl to install them." +find /usr/ -name '*paddle*.whl' | xargs pip install diff --git a/paddle/scripts/docker/build.sh b/paddle/scripts/docker/build.sh index a750c436dc50f906a35313490f667d9a24cc0c00..4172063d923f939dac7229573bc087ec8c62b844 100644 --- a/paddle/scripts/docker/build.sh +++ b/paddle/scripts/docker/build.sh @@ -5,13 +5,8 @@ set -e # Set BASE_IMAGE according to env variables if [ ${WITH_GPU} == "ON" ]; then BASE_IMAGE="nvidia/cuda:8.0-cudnn5-runtime-ubuntu14.04" - # additional packages to install when building gpu images - GPU_DOCKER_PKG="python-pip python-dev" else - BASE_IMAGE="python:2.7.13-slim" - # FIXME: python base image uses different python version than WITH_GPU - # need to change PYTHONHOME to /usr/local when using python base image - CPU_DOCKER_PYTHON_HOME_ENV="ENV PYTHONHOME /usr/local" + BASE_IMAGE="ubuntu:14.04" fi DOCKERFILE_GPU_ENV="" @@ -66,10 +61,7 @@ if [ ${WITH_DOC} == "ON" ]; then rm -rf /paddle/build_doc fi # generate deb package for current build -# FIXME(typhoonzero): should we remove paddle/scripts/deb ? -# FIXME: CPACK_DEBIAN_PACKAGE_DEPENDS removes all dev dependencies, must -# install them in docker -cpack -D CPACK_GENERATOR='DEB' -D CPACK_DEBIAN_PACKAGE_DEPENDS="" .. +cpack -D CPACK_GENERATOR='DEB' .. if [[ ${WOBOQ:-OFF} == 'ON' ]]; then apt-get install -y clang-3.8 llvm-3.8 libclang-3.8-dev @@ -97,32 +89,30 @@ fi paddle version -if [[ -n ${APT_MIRROR} ]]; then - MIRROR_UPDATE="sed -i '${APT_MIRROR}' /etc/apt/sources.list && \\" -else - MIRROR_UPDATE="\\" -fi - cat > /paddle/build/Dockerfile < ENV HOME /root ENV LANG en_US.UTF-8 # Use Fix locales to en_US.UTF-8 -RUN ${MIRROR_UPDATE} - apt-get update && \ - apt-get install -y libgfortran3 libpython2.7 ${GPU_DOCKER_PKG} && \ - apt-get clean -y && \ - pip install --upgrade pip && \ - pip install -U 'protobuf==3.1.0' requests numpy +EOF + +if [[ -n ${APT_MIRROR} ]]; then +cat >> /paddle/build/Dockerfile <> /paddle/build/Dockerfile <=3.2.2', 'numpy>=1.8.0', # The numpy is required. - 'protobuf>=${PROTOBUF_VERSION}' # The paddle protobuf version + 'protobuf==${PROTOBUF_VERSION}' # The paddle protobuf version ], url='http://www.paddlepaddle.org/', license='Apache 2.0', diff --git a/python/paddle/trainer_config_helpers/attrs.py b/python/paddle/trainer_config_helpers/attrs.py index 41ad05c3eb467b9a2e09315f980410d5e9b3853f..7ae9e5cb3050fa6f70fa84785a1ddbdc68c70235 100644 --- a/python/paddle/trainer_config_helpers/attrs.py +++ b/python/paddle/trainer_config_helpers/attrs.py @@ -208,12 +208,15 @@ class ExtraLayerAttribute(object): drop_rate=None, device=None): self.attr = dict() - if isinstance(error_clipping_threshold, float): - assert error_clipping_threshold > 0 - self.attr["error_clipping_threshold"] = error_clipping_threshold - - if isinstance(drop_rate, float): - assert drop_rate > 0 + if error_clipping_threshold is not None: + error_clipping_threshold = float(error_clipping_threshold) + if error_clipping_threshold < 0: + raise ValueError("Error clipping must > 0") + self.attr['error_clipping_threshold'] = error_clipping_threshold + if drop_rate is not None: + drop_rate = float(drop_rate) + if drop_rate < 0: + raise ValueError("Dropout rate must > 0") self.attr["drop_rate"] = drop_rate if isinstance(device, int): diff --git a/python/paddle/trainer_config_helpers/layers.py b/python/paddle/trainer_config_helpers/layers.py index 46fd752d527fa63578a1e01865356780955bc87a..31652613fb3a55636b32babbc4bde60d65776c61 100755 --- a/python/paddle/trainer_config_helpers/layers.py +++ b/python/paddle/trainer_config_helpers/layers.py @@ -84,6 +84,7 @@ __all__ = [ 'GeneratedInput', 'SubsequenceInput', 'gru_step_layer', + 'gru_step_naive_layer', 'recurrent_layer', 'BaseGeneratedInput', 'conv_operator', @@ -2286,7 +2287,7 @@ def img_pool_layer(input, type_name = pool_type.name + '-projection' \ if ( - isinstance(pool_type, AvgPooling) or isinstance(pool_type, MaxPooling)) \ + isinstance(pool_type, AvgPooling) or isinstance(pool_type, MaxPooling)) \ else pool_type.name pool_size_y = pool_size if pool_size_y is None else pool_size_y @@ -3086,6 +3087,78 @@ def gru_step_layer(input, activation=act) +@wrap_bias_attr_default() +@wrap_param_attr_default() +@wrap_act_default(param_names=['gate_act'], act=SigmoidActivation()) +@wrap_act_default(act=TanhActivation()) +@wrap_name_default('gru_step_naive') +@layer_support(ERROR_CLIPPING, DROPOUT) +def gru_step_naive_layer(input, + output_mem, + size=None, + name=None, + act=None, + gate_act=None, + bias_attr=None, + param_attr=None, + layer_attr=None): + """ + GRU Step Layer, but using MixedLayer to generate. It support ERROR_CLIPPING + and DROPOUT. + + :param input: + :param output_mem: + :param size: + :param name: + :param act: + :param gate_act: + :param bias_attr: + :param param_attr: + :param layer_attr: + :return: + """ + if input.size % 3 != 0: + raise ValueError("GruStep input size must be divided by 3") + if size is None: + size = input.size / 3 + + def __gate__(gate_name, offset): + with mixed_layer( + name=name + "_" + gate_name, + size=size, + layer_attr=layer_attr, + bias_attr=bias_attr, + act=gate_act) as gate: + gate += identity_projection(input=input, offset=offset) + gate += full_matrix_projection( + input=output_mem, param_attr=param_attr) + return gate + + update_gate = __gate__("update", 0) + reset_gate = __gate__("reset", size) + + with mixed_layer( + name=name + "_reset_output", bias_attr=False) as reset_output: + reset_output += dotmul_operator(a=output_mem, b=reset_gate) + + with mixed_layer( + name=name + "_output_candidate", + size=size, + layer_attr=layer_attr, + bias_attr=bias_attr, + act=act) as output_candidate: + output_candidate += identity_projection(input=input, offset=2 * size) + output_candidate += full_matrix_projection( + input=reset_output, param_attr=param_attr) + + with mixed_layer(name=name) as output: + output += identity_projection(output_mem) + output += dotmul_operator(a=output_mem, b=update_gate, scale=-1.0) + output += dotmul_operator(a=output_candidate, b=update_gate) + + return output + + @wrap_name_default() @layer_support() def get_output_layer(input, arg_name, name=None, layer_attr=None): diff --git a/python/paddle/trainer_config_helpers/networks.py b/python/paddle/trainer_config_helpers/networks.py index cadde11ff81658cb309cd1bf7a44bac6374c1e44..fb533a47e0b0585be6f0e019086993f8b3aa7f38 100755 --- a/python/paddle/trainer_config_helpers/networks.py +++ b/python/paddle/trainer_config_helpers/networks.py @@ -825,7 +825,8 @@ def gru_unit(input, gru_param_attr=None, act=None, gate_act=None, - gru_layer_attr=None): + gru_layer_attr=None, + naive=False): """ Define calculations that a gated recurrent unit performs in a single time step. This function itself is not a recurrent layer, so that it can not be @@ -857,7 +858,12 @@ def gru_unit(input, out_mem = memory(name=name, size=size) - gru_out = gru_step_layer( + if naive: + __step__ = gru_step_naive_layer + else: + __step__ = gru_step_layer + + gru_out = __step__( name=name, input=input, output_mem=out_mem, @@ -879,7 +885,8 @@ def gru_group(input, gru_param_attr=None, act=None, gate_act=None, - gru_layer_attr=None): + gru_layer_attr=None, + naive=False): """ gru_group is a recurrent layer group version of Gated Recurrent Unit. It does exactly the same calculation as the grumemory layer does. A promising @@ -928,7 +935,8 @@ def gru_group(input, gru_param_attr=gru_param_attr, act=act, gate_act=gate_act, - gru_layer_attr=gru_layer_attr) + gru_layer_attr=gru_layer_attr, + naive=naive) return recurrent_group( name='%s_recurrent_group' % name, @@ -949,7 +957,8 @@ def simple_gru(input, gru_param_attr=None, act=None, gate_act=None, - gru_layer_attr=None): + gru_layer_attr=None, + naive=False): """ You maybe see gru_step_layer, grumemory in layers.py, gru_unit, gru_group, simple_gru in network.py. The reason why there are so many interfaces is @@ -1018,7 +1027,8 @@ def simple_gru(input, gru_param_attr=gru_param_attr, act=act, gate_act=gate_act, - gru_layer_attr=gru_layer_attr) + gru_layer_attr=gru_layer_attr, + naive=naive) @wrap_name_default('simple_gru2') diff --git a/python/paddle/trainer_config_helpers/tests/configs/protostr/projections.protostr b/python/paddle/trainer_config_helpers/tests/configs/protostr/projections.protostr index 2afc3afef6d39ce9b8eef05948861284775d5011..d8bd7b9dfb71a392d0dc53872a0d72f47530530f 100644 --- a/python/paddle/trainer_config_helpers/tests/configs/protostr/projections.protostr +++ b/python/paddle/trainer_config_helpers/tests/configs/protostr/projections.protostr @@ -320,6 +320,7 @@ layers { } } drop_rate: 0.5 + error_clipping_threshold: 40.0 } parameters { name: "___embedding_0__.w0" diff --git a/python/paddle/v2/layer.py b/python/paddle/v2/layer.py index 384de9b9d57f88e84ab6067846174bb037502dc0..89cca7acd34b8dea0572169338649b5e9ff6536a 100644 --- a/python/paddle/v2/layer.py +++ b/python/paddle/v2/layer.py @@ -356,6 +356,9 @@ def mixed(size=0, return MixedLayerV2(size, input, name, act, bias_attr, layer_attr) +mixed.__doc__ = conf_helps.mixed_layer.__doc__ + + class RecurrentLayerInput(Layer): def __init__(self, recurrent_name, index, parent_layers): parents_len = len(parent_layers) @@ -404,6 +407,8 @@ data.__name__ = 'data' AggregateLevel = conf_helps.layers.AggregateLevel ExpandLevel = conf_helps.layers.ExpandLevel memory = MemoryV2 +memory.__name__ = 'memory' +memory.__doc__ = conf_helps.memory.__doc__ def __layer_name_mapping__(inname): @@ -512,6 +517,9 @@ def recurrent_group(step, input, name=None): return retv +recurrent_group.__doc__ = conf_helps.recurrent_group.__doc__ + + @wrap_name_default() def beam_search(step, input, @@ -579,6 +587,8 @@ def beam_search(step, return tmp +beam_search.__doc__ = conf_helps.beam_search.__doc__ + __projection_names__ = filter(lambda x: x.endswith('_projection'), dir(conf_helps)) diff --git a/python/setup.py.in b/python/setup.py.in index 228e762d56f672d7b31ede2b2b92c77f9a126f3c..5dfb46192ae54fdc36b0867312cf156aefb84f84 100644 --- a/python/setup.py.in +++ b/python/setup.py.in @@ -15,6 +15,9 @@ setup(name='paddle', description='Parallel Distributed Deep Learning', install_requires=[ "requests", + "numpy", + "protobuf==${PROTOBUF_VERSION}", + "matplotlib", ], packages=packages, package_dir={