Merge pull request #516 from wangyang59/gan

Demo: Generative Adverserial Nets

.gitignore

@@ -8,3 +8,4 @@ build/
 .cproject
 .pydevproject
 Makefile
+.test_env/

demo/gan/.gitignore

@@ -0,0 +1,11 @@
+output/
+uniform_params/
+cifar_params/
+mnist_params/
+*.png
+.pydevproject
+.project
+*.log
+*.pyc
+data/mnist_data/
+data/cifar-10-batches-py/

demo/gan/README.md

@@ -0,0 +1,13 @@
+# Generative Adversarial Networks (GAN) 
+
+This demo implements GAN training described in the original GAN paper (https://arxiv.org/abs/1406.2661) and DCGAN (https://arxiv.org/abs/1511.06434).
+
+The general training procedures are implemented in gan_trainer.py. The neural network configurations are specified in gan_conf.py (for synthetic data) and gan_conf_image.py (for image data).
+
+In order to run the model, first download the corresponding data by running the shell script in ./data.
+Then you can run the command below. The flag -d specifies the training data (cifar, mnist or uniform) and flag --useGpu specifies whether to use gpu for training (0 is cpu, 1 is gpu).  
+
+$python gan_trainer.py -d cifar --use_gpu 1
+
+The generated images will be stored in ./cifar_samples/
+The corresponding models will be stored in ./cifar_params/

demo/gan/data/download_cifar.sh

@@ -0,0 +1,18 @@
+# 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
+wget https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz
+tar zxf cifar-10-python.tar.gz
+rm cifar-10-python.tar.gz
+

demo/gan/data/get_mnist_data.sh

@@ -0,0 +1,19 @@
+#!/usr/bin/env sh
+# This script downloads the mnist data and unzips it.
+set -e
+DIR="$( cd "$(dirname "$0")" ; pwd -P )"
+rm -rf "$DIR/mnist_data"
+mkdir "$DIR/mnist_data"
+cd "$DIR/mnist_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
+
+

demo/gan/gan_conf.py

@@ -0,0 +1,134 @@
+# 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 *
+
+mode = get_config_arg("mode", str, "generator")
+assert mode in set(["generator",
+                    "discriminator",
+                    "generator_training",
+                    "discriminator_training"])
+
+is_generator_training = mode == "generator_training"
+is_discriminator_training = mode == "discriminator_training"
+is_generator = mode == "generator"
+is_discriminator = mode == "discriminator"
+
+# The network structure below follows the ref https://arxiv.org/abs/1406.2661
+# Here we used two hidden layers and batch_norm
+
+print('mode=%s' % mode)
+# the dim of the noise (z) as the input of the generator network
+noise_dim = 10
+# the dim of the hidden layer
+hidden_dim = 10
+# the dim of the generated sample
+sample_dim = 2
+
+settings(
+    batch_size=128,
+    learning_rate=1e-4,
+    learning_method=AdamOptimizer(beta1=0.5)
+)
+
+def discriminator(sample):
+    """
+    discriminator ouputs the probablity of a sample is from generator
+    or real data.
+    The output has two dimenstional: dimension 0 is the probablity
+    of the sample is from generator and dimension 1 is the probabblity
+    of the sample is from real data.
+    """
+    param_attr = ParamAttr(is_static=is_generator_training)
+    bias_attr = ParamAttr(is_static=is_generator_training,
+                          initial_mean=1.0,
+                          initial_std=0)
+
+    hidden = fc_layer(input=sample, name="dis_hidden", size=hidden_dim,
+                    bias_attr=bias_attr,
+                    param_attr=param_attr,
+                    act=ReluActivation())
+
+    hidden2 = fc_layer(input=hidden, name="dis_hidden2", size=hidden_dim,
+                    bias_attr=bias_attr,
+                    param_attr=param_attr,
+                    act=LinearActivation())
+    
+    hidden_bn = batch_norm_layer(hidden2, 
+                     act=ReluActivation(), 
+                     name="dis_hidden_bn", 
+                     bias_attr=bias_attr, 
+                     param_attr=ParamAttr(is_static=is_generator_training,
+                           initial_mean=1.0,
+                           initial_std=0.02),
+                     use_global_stats=False)
+    
+    return fc_layer(input=hidden_bn, name="dis_prob", size=2,
+                    bias_attr=bias_attr,
+                    param_attr=param_attr,
+                    act=SoftmaxActivation())
+
+def generator(noise):
+    """
+    generator generates a sample given noise
+    """
+    param_attr = ParamAttr(is_static=is_discriminator_training)
+    bias_attr = ParamAttr(is_static=is_discriminator_training,
+                           initial_mean=1.0,
+                           initial_std=0)
+    
+    hidden = fc_layer(input=noise,
+                    name="gen_layer_hidden",
+                    size=hidden_dim,
+                    bias_attr=bias_attr,
+                    param_attr=param_attr,
+                    act=ReluActivation())
+
+    hidden2 = fc_layer(input=hidden, name="gen_hidden2", size=hidden_dim,
+                    bias_attr=bias_attr,
+                    param_attr=param_attr,
+                    act=LinearActivation())
+    
+    hidden_bn = batch_norm_layer(hidden2, 
+                     act=ReluActivation(), 
+                     name="gen_layer_hidden_bn", 
+                     bias_attr=bias_attr, 
+                     param_attr=ParamAttr(is_static=is_discriminator_training,
+                           initial_mean=1.0,
+                           initial_std=0.02),
+                     use_global_stats=False)
+    
+    return fc_layer(input=hidden_bn,
+                    name="gen_layer1",
+                    size=sample_dim,
+                    bias_attr=bias_attr,
+                    param_attr=param_attr,
+                    act=LinearActivation())
+
+if is_generator_training:
+    noise = data_layer(name="noise", size=noise_dim)
+    sample = generator(noise)
+
+if is_discriminator_training:
+    sample = data_layer(name="sample", size=sample_dim)
+
+if is_generator_training or is_discriminator_training:
+    label = data_layer(name="label", size=1)
+    prob = discriminator(sample)
+    cost = cross_entropy(input=prob, label=label)
+    classification_error_evaluator(input=prob, label=label, name=mode+'_error')
+    outputs(cost)
+
+if is_generator:
+    noise = data_layer(name="noise", size=noise_dim)
+    outputs(generator(noise))

paddle/api/Arguments.cpp

@@ -27,11 +27,6 @@ Arguments* Arguments::createArguments(size_t slotNum) {
 
 void Arguments::resize(size_t slotNum) { m->outputs.resize(slotNum); }
 
-Matrix* Arguments::getSlotValue(size_t idx) const throw(RangeError) {
-  auto& a = m->getArg(idx);
-  return Matrix::createByPaddleMatrixPtr(&a.value);
-}
-
 Arguments::Arguments() : m(new ArgumentsPrivate()) {}
 
 Arguments::~Arguments() { delete m; }
@@ -43,6 +38,16 @@ Arguments* Arguments::createByPaddleArgumentVector(void* ptr) {
   return args;
 }
 
+Matrix* Arguments::getSlotValue(size_t idx) const throw(RangeError) {
+  auto& a = m->getArg(idx);
+  return Matrix::createByPaddleMatrixPtr(&a.value);
+}
+
+Matrix* Arguments::getSlotGrad(size_t idx) const throw(RangeError) {
+  auto& a = m->getArg(idx);
+  return Matrix::createByPaddleMatrixPtr(&a.grad);
+}
+
 IVector* Arguments::getSlotIds(size_t idx) const throw(RangeError) {
   auto& a = m->getArg(idx);
   return IVector::createByPaddleVectorPtr(&a.ids);
@@ -58,6 +63,11 @@ void Arguments::setSlotValue(size_t idx, Matrix* mat) throw(RangeError) {
   a.value = m->cast<paddle::Matrix>(mat->getSharedPtr());
 }
 
+void Arguments::setSlotGrad(size_t idx, Matrix* mat) throw(RangeError) {
+  auto& a = m->getArg(idx);
+  a.grad = m->cast<paddle::Matrix>(mat->getSharedPtr());
+}
+
 void Arguments::setSlotIn(size_t idx, Matrix* mat) throw(RangeError) {
   auto& a = m->getArg(idx);
   a.in = m->cast<paddle::Matrix>(mat->getSharedPtr());

paddle/api/Paddle.swig

@@ -193,5 +193,4 @@ namespace std {
 %ignore OptimizationConfigPrivate;
 %ignore ParameterTraverseCallbackPrivate;
 %include "utils/GlobalConstants.h"
-%include "api/PaddleAPI.h"
+%include "api/PaddleAPI.h"
-

paddle/api/PaddleAPI.h

@@ -156,12 +156,15 @@ public:
    *  @param dim1  dimension of data.
    *  @param dim2  dimension of data.
    *  @param copy  true if copy into a new matrix, false will create
-   *               matrix inplace.
+   *               matrix inplace. copy = false should be used with extreme
+   *               care because Matrix will share the memory with the given
+   *               numpy array. If the numpy array object is no longer valid,
+   *               the memory space will not be usable.
    */
   static Matrix* createCpuDenseFromNumpy(float* data,
                                          int dim1,
                                          int dim2,
-                                         bool copy = false);
+                                         bool copy = true);
 
   /// Create Gpu Dense Matrix from numpy matrix, dtype=float32
   static Matrix* createGpuDenseFromNumpy(float* data, int dim1, int dim2);
@@ -271,11 +274,18 @@ public:
    */
   static Vector* createCpuVectorFromNumpy(float* data,
                                           int dim,
-                                          bool copy = false);
+                                          bool copy = true);
 
   /// Create Gpu Vector from numpy array, which dtype=float32
   static Vector* createGpuVectorFromNumpy(float* data, int dim);
 
+  /**
+   * copy from another vector
+   * throw(RangeError) if size of src vector is different from size of this
+   * vector
+   */
+  void copyFrom(Vector* src) throw(RangeError);
+
   /// Cast to numpy array inplace.
   void toNumpyArrayInplace(float** view_data, int* dim1) throw(UnsupportError);
 
@@ -339,7 +349,7 @@ public:
    */
   static IVector* createCpuVectorFromNumpy(int* data,
                                            int dim,
-                                           bool copy = false);
+                                           bool copy = true);
   /**
    * Create Gpu IVector from numpy array, which dtype=int32
    */
@@ -418,6 +428,7 @@ public:
    * the param idx is the slot id
    */
   Matrix* getSlotValue(size_t idx) const throw(RangeError);
+  Matrix* getSlotGrad(size_t idx) const throw(RangeError);
   IVector* getSlotIds(size_t idx) const throw(RangeError);
   Matrix* getSlotIn(size_t idx) const throw(RangeError);
   IVector* getSlotSequenceStartPositions(size_t idx) const throw(RangeError);
@@ -434,6 +445,7 @@ public:
    * The other param is the input Matrix or vector.
    */
   void setSlotValue(size_t idx, Matrix* mat) throw(RangeError);
+  void setSlotGrad(size_t idx, Matrix* mat) throw(RangeError);
   void setSlotIn(size_t idx, Matrix* mat) throw(RangeError);
   void setSlotIds(size_t idx, IVector* vec) throw(RangeError);
   void setSlotSequenceStartPositions(size_t idx,
@@ -535,6 +547,7 @@ public:
   size_t getID() const;
 
   ParameterConfig* getConfig();
+  void setValueUpdated();
 
 private:
   static Parameter* createFromRawPtr(void* ptr);

paddle/api/Parameter.cpp

@@ -68,3 +68,5 @@ ParameterConfig* Parameter::getConfig() {
 }
 
 size_t Parameter::getID() const { return m->getPtr()->getID(); }
+
+void Parameter::setValueUpdated() { m->getPtr()->setValueUpdated(); }

paddle/api/Vector.cpp

@@ -281,6 +281,13 @@ FloatArray Vector::getData() const {
   }
 }
 
+void Vector::copyFrom(Vector* src) throw(RangeError) {
+  if (src->m->vec->getSize() !=  m->vec->getSize()) {
+    throw RangeError();
+  }
+  m->vec->copyFrom(*src->m->vec);
+}
+
 bool Vector::isGpu() const {
   return std::dynamic_pointer_cast<paddle::GpuVector>(m->vec) != nullptr;
 }

paddle/api/test/testMatrix.py

@@ -68,7 +68,7 @@ class TestMatrix(unittest.TestCase):
 
     def test_numpyCpu(self):
         numpy_mat = np.matrix([[1, 2], [3, 4], [5, 6]], dtype="float32")
-        m = swig_paddle.Matrix.createCpuDenseFromNumpy(numpy_mat)
+        m = swig_paddle.Matrix.createCpuDenseFromNumpy(numpy_mat, copy=False)
         self.assertEqual((int(m.getHeight()), int(m.getWidth())),
                          numpy_mat.shape)
 

paddle/api/test/testVector.py

@@ -43,7 +43,7 @@ class TestIVector(unittest.TestCase):
 
     def test_cpu_numpy(self):
         vec = np.array([1, 3, 4, 65, 78, 1, 4], dtype="int32")
-        iv = swig_paddle.IVector.createCpuVectorFromNumpy(vec)
+        iv = swig_paddle.IVector.createCpuVectorFromNumpy(vec, copy=False)
         self.assertEqual(vec.shape[0], int(iv.__len__()))
         vec[4] = 832
         for i in xrange(len(iv)):
@@ -107,7 +107,7 @@ class TestVector(unittest.TestCase):
 
     def testCpuNumpy(self):
         numpy_arr = np.array([1.2, 2.3, 3.4, 4.5], dtype="float32")
-        vec = swig_paddle.Vector.createCpuVectorFromNumpy(numpy_arr)
+        vec = swig_paddle.Vector.createCpuVectorFromNumpy(numpy_arr, copy=False)
         assert isinstance(vec, swig_paddle.Vector)
         numpy_arr[0] = 0.1
         for n, v in zip(numpy_arr, vec):
@@ -152,4 +152,4 @@ if __name__ == '__main__':
     unittest.TextTestRunner().run(suite)
     if swig_paddle.isGpuVersion():
         swig_paddle.setUseGpu(True)
-        unittest.main()
+        unittest.main()

paddle/api/test/util.py

@@ -24,7 +24,9 @@ def doubleEqual(a, b):
 
 def __readFromFile():
     for i in xrange(10002):
-        yield np.random.rand(784), random.randint(0, 9)
+        label = np.random.randint(0, 9)
+        sample = np.random.rand(784) + 0.1 * label
+        yield sample, label
 
 
 def loadMNISTTrainData(batch_size=100):

paddle/gserver/layers/BatchNormBaseLayer.cpp

@@ -68,10 +68,10 @@ void BatchNormBaseLayer::calFeatureMapSize() {
   } else {
     imageH_ = inputLayers_[0]->getOutput().getFrameHeight();
     imageW_ = inputLayers_[0]->getOutput().getFrameWidth();
+    getOutput().setFrameHeight(imageH_);
+    getOutput().setFrameWidth(imageW_);
   }
   imgPixels_ = imageH_ * imageW_;
-  getOutput().setFrameHeight(imageH_);
-  getOutput().setFrameWidth(imageW_);
 }
 
 }  // namespace paddle

paddle/gserver/tests/CMakeLists.txt

@@ -39,9 +39,17 @@ add_unittest_without_exec(test_ConvUnify
     test_ConvUnify.cpp
     LayerGradUtil.cpp
     TestUtil.cpp)
-
+    
 add_test(NAME test_ConvUnify
     COMMAND test_ConvUnify)
+################# test_BatchNorm #######################
+add_unittest_without_exec(test_BatchNorm
+    test_BatchNorm.cpp
+    LayerGradUtil.cpp
+    TestUtil.cpp)
+
+add_test(NAME test_BatchNorm
+    COMMAND test_BatchNorm)
 ################## test_Evaluator #######################
 add_unittest(test_Evaluator
     test_Evaluator.cpp

paddle/gserver/tests/test_BatchNorm.cpp

@@ -0,0 +1,120 @@
+/* Copyright (c) 2016 Baidu, Inc. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include <gtest/gtest.h>
+#include <vector>
+#include <string>
+#include "paddle/gserver/layers/DataLayer.h"
+#include "ModelConfig.pb.h"
+#include "paddle/trainer/Trainer.h"
+#include "paddle/utils/GlobalConstants.h"
+#include "paddle/gserver/layers/ExpandConvTransLayer.h"
+
+#include "TestUtil.h"
+#include "LayerGradUtil.h"
+
+using namespace paddle;  // NOLINT
+using namespace std;     // NOLINT
+
+P_DECLARE_bool(use_gpu);
+P_DECLARE_int32(gpu_id);
+P_DECLARE_double(checkgrad_eps);
+P_DECLARE_bool(thread_local_rand_use_global_seed);
+P_DECLARE_bool(prev_batch_state);
+
+// Test that the batchNormLayer can be followed by a ConvLayer
+TEST(Layer, batchNorm) {
+    FLAGS_use_gpu = false;
+    TestConfig configBN;
+    const int CHANNELS = 6272;
+    const int IMG_SIZE = 1;
+    configBN.layerConfig.set_type("batch_norm");
+    configBN.layerConfig.set_name("bn");
+    configBN.layerConfig.set_size(CHANNELS * IMG_SIZE * IMG_SIZE);
+    configBN.layerConfig.set_active_type("relu");
+    configBN.biasSize = CHANNELS;
+    configBN.inputDefs.push_back({INPUT_DATA, "layer_0",
+                                /* dim= */ IMG_SIZE * IMG_SIZE * CHANNELS,
+                                /* paraSize= */ CHANNELS});
+
+    configBN.inputDefs.push_back({INPUT_DATA, "layer_1_running_mean",
+                                    1, CHANNELS});
+    configBN.inputDefs.back().isStatic = true;
+    configBN.inputDefs.push_back({INPUT_DATA, "layer_2_running_var",
+                                    1, CHANNELS});
+    configBN.inputDefs.back().isStatic = true;
+
+    LayerInputConfig* input = configBN.layerConfig.add_inputs();
+    configBN.layerConfig.add_inputs();
+    configBN.layerConfig.add_inputs();
+
+    ImageConfig* img_conf = input->mutable_image_conf();
+    img_conf->set_channels(CHANNELS);
+    img_conf->set_img_size(IMG_SIZE);
+
+    // Setting up conv-layer config
+    TestConfig config;
+    config.biasSize = 64;
+    config.layerConfig.set_type("exconv");
+    config.layerConfig.set_num_filters(64);
+    config.layerConfig.set_partial_sum(1);
+    config.layerConfig.set_shared_biases(true);
+
+    config.inputDefs.push_back({INPUT_DATA, "bn", 6272, 204800});
+    input = config.layerConfig.add_inputs();
+    ConvConfig* conv = input->mutable_conv_conf();
+    conv->set_filter_size(5);
+    conv->set_filter_size_y(5);
+    conv->set_channels(128);
+    conv->set_padding(1);
+    conv->set_padding_y(1);
+    conv->set_stride(2);
+    conv->set_stride_y(2);
+    conv->set_groups(1);
+    conv->set_filter_channels(conv->channels() / conv->groups());
+    conv->set_img_size(7);
+    conv->set_output_x(3);
+    config.layerConfig.set_size(conv->output_x() * conv->output_x() *
+                                config.layerConfig.num_filters());
+    config.layerConfig.set_name("conv");
+
+    // data layer initialize
+    std::vector<DataLayerPtr> dataLayers;
+    LayerMap layerMap;
+    vector<Argument> datas;
+    initDataLayer(configBN, &dataLayers, &datas, &layerMap, "batch_norm",
+                  100, false, false);
+    // test layer initialize
+    std::vector<ParameterPtr> parameters;
+    LayerPtr bnLayer;
+    initTestLayer(configBN, &layerMap, &parameters, &bnLayer);
+
+    std::vector<ParameterPtr> parameters2;
+    LayerPtr convLayer;
+    initTestLayer(config, &layerMap, &parameters2, &convLayer);
+
+    bnLayer->forward(PASS_GC);
+    convLayer->forward(PASS_GC);
+
+    CHECK_EQ(convLayer->getOutputValue()->getHeight(), 100);
+    CHECK_EQ(convLayer->getOutputValue()->getWidth(), 576);
+}
+
+int main(int argc, char** argv) {
+  testing::InitGoogleTest(&argc, argv);
+  initMain(argc, argv);
+  FLAGS_thread_local_rand_use_global_seed = true;
+  srand(1);
+  return RUN_ALL_TESTS();
+}

paddle/py_paddle/util.py

@@ -559,10 +559,10 @@ def __monkey_patch_trainer__():
 
 
 def monkeypatches():
-    patches = [
+    patches = [__monkeypatch_init_paddle__,
-        __monkeypatch_init_paddle__, __monkeypatch_gradient_machine__,
+               __monkeypatch_gradient_machine__,
-        __monkey_patch_protobuf_objects__, __monkey_patch_parameter__,
+               __monkey_patch_protobuf_objects__,
-        __monkey_patch_trainer__
+               __monkey_patch_parameter__,
-    ]
+               __monkey_patch_trainer__]
     for patch in patches:
         patch()