Merge branch 'develop' into merge_bn

benchmark/tensorflow/mnist.py

+#   Copyright (c) 2018 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 __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import argparse
+import time
+import numpy as np
+
+import tensorflow as tf
+import paddle.v2 as paddle
+
+DTYPE = tf.float32
+
+
+def parse_args():
+    parser = argparse.ArgumentParser("mnist model benchmark.")
+    parser.add_argument(
+        '--batch_size', type=int, default=128, help='The minibatch size.')
+    parser.add_argument(
+        '--iterations', type=int, default=35, help='The number of minibatches.')
+    parser.add_argument(
+        '--pass_num', type=int, default=5, help='The number of passes.')
+    parser.add_argument(
+        '--device',
+        type=str,
+        default='GPU',
+        choices=['CPU', 'GPU'],
+        help='The device type.')
+    args = parser.parse_args()
+    return args
+
+
+def run_benchmark(args):
+    def weight_variable(dtype, shape):
+        initial = tf.truncated_normal(shape, stddev=0.1, dtype=dtype)
+        return tf.Variable(initial)
+
+    def bias_variable(dtype, shape):
+        initial = tf.constant(0.1, shape=shape, dtype=dtype)
+        return tf.Variable(initial)
+
+    device = '/cpu:0' if args.device == 'CPU' else '/device:GPU:0'
+    with tf.device(device):
+        images = tf.placeholder(DTYPE, shape=(None, 28, 28, 1))
+        labels = tf.placeholder(tf.int64, shape=(None, ))
+
+        # conv1, relu, pool1
+        conv1_weights = weight_variable(DTYPE, [5, 5, 1, 20])
+        conv1_bias = bias_variable(DTYPE, [20])
+        conv1 = tf.nn.conv2d(
+            images, conv1_weights, strides=[1, 1, 1, 1], padding="VALID")
+        relu1 = tf.nn.relu(tf.nn.bias_add(conv1, conv1_bias))
+        pool1 = tf.nn.max_pool(
+            relu1, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding="VALID")
+
+        # conv2, relu, pool2
+        conv2_weights = weight_variable(DTYPE, [5, 5, 20, 50])
+        conv2_bias = bias_variable(DTYPE, [50])
+        conv2 = tf.nn.conv2d(
+            pool1, conv2_weights, strides=[1, 1, 1, 1], padding="VALID")
+        relu2 = tf.nn.relu(tf.nn.bias_add(conv2, conv2_bias))
+        pool2 = tf.nn.max_pool(
+            relu2, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding="VALID")
+
+        # FC 
+        pool_shape = pool2.get_shape().as_list()
+        hidden_dim = reduce(lambda a, b: a * b, pool_shape[1:], 1)
+        reshape = tf.reshape(pool2, shape=(tf.shape(pool2)[0], hidden_dim))
+        fc_weights = weight_variable(DTYPE, [hidden_dim, 10])
+        fc_bias = bias_variable(DTYPE, [10])
+        logits = tf.matmul(reshape, fc_weights) + fc_bias
+
+        # Get prediction
+        prediction = tf.nn.softmax(logits)
+
+        # Loss 
+        one_hot_labels = tf.one_hot(labels, depth=10)
+        cost = -tf.reduce_sum(tf.log(prediction) * one_hot_labels, [1])
+        avg_cost = tf.reduce_mean(cost)
+
+        # Get accuracy
+        correct = tf.equal(tf.argmax(prediction, 1), labels)
+        accuracy = tf.reduce_mean(tf.cast(correct, tf.float32))
+
+        # metrics, g_accuracy
+        with tf.variable_scope("reset_metrics_accuracy_scope") as scope:
+            g_accuracy = tf.metrics.accuracy(
+                labels, tf.argmax(
+                    prediction, axis=1))
+            vars = tf.contrib.framework.get_variables(
+                scope, collection=tf.GraphKeys.LOCAL_VARIABLES)
+            g_accuracy_reset_op = tf.variables_initializer(vars)
+
+        # Optimizer 
+        opt = tf.train.AdamOptimizer(
+            learning_rate=0.001, beta1=0.9, beta2=0.999)
+        train_op = opt.minimize(avg_cost)
+        # train_op = tf.train.AdamOptimizer(1e-4).minimize(avg_cost)
+
+    train_reader = paddle.batch(
+        paddle.dataset.mnist.train(), batch_size=args.batch_size)
+    test_reader = paddle.batch(
+        paddle.dataset.mnist.test(), batch_size=args.batch_size)
+
+    def eval_test():
+        sess.run(g_accuracy_reset_op)
+        for batch_id, data in enumerate(test_reader()):
+            images_data = np.array(
+                map(lambda x: np.transpose(x[0].reshape([1, 28, 28]), axes=[1,2,0]), data)).astype("float32")
+            labels_data = np.array(map(lambda x: x[1], data)).astype("int64")
+
+            loss, acc, g_acc = sess.run(
+                [avg_cost, accuracy, g_accuracy],
+                feed_dict={images: images_data,
+                           labels: labels_data})
+        return g_acc[1]
+
+    config = tf.ConfigProto(
+        intra_op_parallelism_threads=1, inter_op_parallelism_threads=1)
+    config.gpu_options.allow_growth = True
+
+    with tf.Session(config=config) as sess:
+        init_g = tf.global_variables_initializer()
+        init_l = tf.local_variables_initializer()
+        sess.run(init_g)
+        sess.run(init_l)
+        for pass_id in range(args.pass_num):
+            sess.run(g_accuracy_reset_op)
+
+            pass_start = time.time()
+            for batch_id, data in enumerate(train_reader()):
+                images_data = np.array(
+                    map(lambda x: np.transpose(x[0].reshape([1, 28, 28]), axes=[1,2,0]), data)).astype("float32")
+                labels_data = np.array(map(lambda x: x[1], data)).astype(
+                    "int64")
+
+                start = time.time()
+                _, loss, acc, g_acc = sess.run(
+                    [train_op, avg_cost, accuracy, g_accuracy],
+                    feed_dict={images: images_data,
+                               labels: labels_data})
+                end = time.time()
+
+                print("pass=%d, batch=%d, loss=%f, error=%f, elapse=%f" %
+                      (pass_id, batch_id, loss, 1 - acc, (end - start) / 1000))
+
+            pass_end = time.time()
+            test_avg_acc = eval_test()
+
+            print(
+                "pass=%d, training_avg_accuracy=%f, test_avg_acc=%f, elapse=%f"
+                % (pass_id, g_acc[1], test_avg_acc,
+                   (pass_end - pass_start) / 1000))
+
+
+def print_arguments(args):
+    print('-----------  Configuration Arguments -----------')
+    for arg, value in sorted(vars(args).iteritems()):
+        print('%s: %s' % (arg, value))
+    print('------------------------------------------------')
+
+
+if __name__ == '__main__':
+    args = parse_args()
+    print_arguments(args)
+    run_benchmark(args)

benchmark/tensorflow/resnet.py

+#   Copyright (c) 2018 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.
+"""
+based on https://github.com/tensorflow/models/blob/master/official/resnet/resnet_model.py
+
+Get help: python resnet.py --help
+See performance on flowers: python resnet.py
+Train on cifar10: python resnet.py --data=cifar10 --with_test
+"""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import argparse
+import time
+import numpy as np
+
+import paddle.v2 as paddle
+import tensorflow as tf
+
+DTYPE = tf.float32
+
+
+def parse_args():
+    parser = argparse.ArgumentParser('Convolution model benchmark.')
+    parser.add_argument(
+        '--model',
+        type=str,
+        choices=['resnet'],
+        default='resnet',
+        help='The model architecture.')
+    parser.add_argument(
+        '--batch_size', type=int, default=32, help='The minibatch size.')
+    parser.add_argument(
+        '--use_fake_data',
+        action='store_true',
+        help='use real data or fake data')
+    parser.add_argument(
+        '--skip_batch_num',
+        type=int,
+        default=5,
+        help='The first num of minibatch num to skip, for better performance test'
+    )
+    parser.add_argument(
+        '--iterations',
+        type=int,
+        default=105,
+        help='The number of minibatches.')
+    parser.add_argument(
+        '--pass_num', type=int, default=300, help='The number of passes.')
+    parser.add_argument(
+        '--order',
+        type=str,
+        default='NHWC',
+        choices=['NCHW', 'NHWC'],
+        help='The data order, now only support NCHW.')
+    parser.add_argument(
+        '--device',
+        type=str,
+        default='GPU',
+        choices=['CPU', 'GPU'],
+        help='The device type.')
+    parser.add_argument(
+        '--data',
+        type=str,
+        default='flowers102',
+        choices=['flowers102', 'cifar10'],
+        help='The kinds of data.')
+    parser.add_argument(
+        '--infer_only', action='store_true', help='If set, run forward only.')
+    parser.add_argument(
+        '--use_cprof', action='store_true', help='If set, use cProfile.')
+    parser.add_argument(
+        '--with_test',
+        action='store_true',
+        help='If set, test the testset during training.')
+    parser.add_argument(
+        '--use_nvprof',
+        action='store_true',
+        help='If set, use nvprof for CUDA.')
+    args = parser.parse_args()
+    return args
+
+
+def print_arguments(args):
+    vars(args)['use_nvprof'] = (vars(args)['use_nvprof'] and
+                                vars(args)['device'] == 'GPU')
+    vars(args)['iterations'] = vars(args)['pass_num'] * 1000 if vars(args)[
+        'with_test'] else vars(args)['iterations']
+    print('-----------  Configuration Arguments -----------')
+    for arg, value in sorted(vars(args).iteritems()):
+        print('%s: %s' % (arg, value))
+    print('------------------------------------------------')
+
+
+def fixed_padding(inputs, kernel_size, data_format):
+    """Pads the input along the spatial dimensions independently of input size.
+  Args:
+    inputs: A tensor of size [batch, channels, height_in, width_in] or
+      [batch, height_in, width_in, channels] depending on data_format.
+    kernel_size: The kernel to be used in the conv2d or max_pool2d operation.
+                 Should be a positive integer.
+    data_format: The input format ('channels_last' or 'channels_first').
+  Returns:
+    A tensor with the same format as the input with the data either intact
+    (if kernel_size == 1) or padded (if kernel_size > 1).
+  """
+    pad_total = kernel_size - 1
+    pad_beg = pad_total // 2
+    pad_end = pad_total - pad_beg
+
+    if data_format == 'channels_first':
+        padded_inputs = tf.pad(inputs, [[0, 0], [0, 0], [pad_beg, pad_end],
+                                        [pad_beg, pad_end]])
+    else:
+        padded_inputs = tf.pad(inputs, [[0, 0], [pad_beg, pad_end],
+                                        [pad_beg, pad_end], [0, 0]])
+    return padded_inputs
+
+
+def conv2d_fixed_padding(inputs, filters, kernel_size, strides, data_format):
+    """Strided 2-D convolution with explicit padding."""
+    # The padding is consistent and is based only on `kernel_size`, not on the
+    # dimensions of `inputs` (as opposed to using `tf.layers.conv2d` alone).
+    # This is consistent with PaddlePaddle.
+    # In addition, the calculation for output size in TensorFlow can refer: 
+    # https://github.com/tensorflow/tensorflow/blob/master/tensorflow/core/framework/common_shape_fns.cc
+    if strides > 1:
+        inputs = fixed_padding(inputs, kernel_size, data_format)
+
+    return tf.layers.conv2d(
+        inputs=inputs,
+        filters=filters,
+        kernel_size=kernel_size,
+        strides=strides,
+        padding=('SAME' if strides == 1 else 'VALID'),
+        use_bias=False,
+        kernel_initializer=tf.variance_scaling_initializer(),
+        data_format=data_format)
+
+
+def conv_bn(inputs,
+            filters,
+            kernel_size,
+            strides,
+            is_training,
+            data_format,
+            act=True):
+    # def conv2d_fixed_padding(inputs, filters, kernel_size, strides, data_format):
+    # set fused=True for a significant performance boost. See
+    # https://www.tensorflow.org/performance/performance_guide#common_fused_ops
+    inputs = conv2d_fixed_padding(
+        inputs=inputs,
+        filters=filters,
+        kernel_size=kernel_size,
+        strides=strides,
+        data_format=data_format)
+    inputs = tf.layers.batch_normalization(
+        inputs=inputs,
+        axis=1 if data_format == 'channels_first' else 3,
+        momentum=0.9,
+        epsilon=1e-05,
+        center=True,
+        scale=True,
+        training=is_training,
+        fused=True)
+    if act:
+        inputs = tf.nn.relu(inputs)
+    return inputs
+
+
+def basicblock(inputs, filters, is_training, projection_shortcut, strides,
+               data_format):
+    shortcut = inputs
+    if projection_shortcut is not None:
+        shortcut = projection_shortcut(inputs)
+    inputs = conv_bn(inputs, filters, 3, strides, is_training, data_format)
+    inputs = conv_bn(inputs, filters, 3, 1, is_training, data_format, act=False)
+    inputs = inputs + shortcut
+    inputs = tf.nn.relu(inputs)
+    return inputs
+
+
+def bottleneck(inputs, filters, is_training, projection_shortcut, strides,
+               data_format):
+    shortcut = inputs
+    if projection_shortcut is not None:
+        shortcut = projection_shortcut(inputs)
+    inputs = conv_bn(inputs, filters, 1, strides, is_training, data_format)
+    inputs = conv_bn(inputs, filters, 3, 1, is_training, data_format, act=False)
+    inputs = conv_bn(
+        inputs, filters * 4, 1, 1, is_training, data_format, act=False)
+    inputs = inputs + shortcut
+    inputs = tf.nn.relu(inputs)
+    return inputs
+
+
+def block_layer(inputs, filters, block_fn, blocks, strides, is_training, name,
+                data_format):
+    # Bottleneck blocks end with 4x the number of filters as they start with
+    filters_out = 4 * filters if block_fn is bottleneck else filters
+
+    def projection_shortcut(inputs):
+        return conv2d_fixed_padding(
+            inputs=inputs,
+            filters=filters_out,
+            kernel_size=1,
+            strides=strides,
+            data_format=data_format)
+
+    # Only the first block per block_layer uses projection_shortcut and strides
+    inputs = block_fn(inputs, filters, is_training, projection_shortcut,
+                      strides, data_format)
+
+    for _ in range(1, blocks):
+        inputs = block_fn(inputs, filters, is_training, None, 1, data_format)
+
+    return tf.identity(inputs, name)
+
+
+def resnet_imagenet(depth, class_dim, data_format):
+    """Returns the ResNet model for a given size and number of output classes."""
+
+    def resnet_generator(block_fn,
+                         layers,
+                         num_classes,
+                         data_format='channels_last'):
+        if data_format is None:
+            data_format = ('channels_first'
+                           if tf.test.is_built_with_cuda() else 'channels_last')
+
+        def model(inputs, is_training):
+            """Constructs the ResNet model given the inputs."""
+            if data_format == 'channels_first':
+                # Convert the inputs from channels_last (NHWC) to channels_first (NCHW).
+                # This provides a large performance boost on GPU. See
+                # https://www.tensorflow.org/performance/performance_guide#data_formats
+                inputs = tf.transpose(inputs, [0, 3, 1, 2])
+
+            inputs = conv_bn(inputs, 64, 7, 2, is_training, data_format)
+            inputs = tf.identity(inputs, 'initial_conv')
+            inputs = tf.layers.max_pooling2d(
+                inputs=inputs,
+                pool_size=3,
+                strides=2,
+                padding='SAME',
+                data_format=data_format)
+            inputs = tf.identity(inputs, 'initial_max_pool')
+            inputs = block_layer(inputs, 64, block_fn, layers[0], 1,
+                                 is_training, 'block_layer1', data_format)
+            inputs = block_layer(inputs, 128, block_fn, layers[1], 2,
+                                 is_training, 'block_layer2', data_format)
+            inputs = block_layer(inputs, 256, block_fn, layers[2], 2,
+                                 is_training, 'block_layer3', data_format)
+            inputs = block_layer(inputs, 512, block_fn, layers[3], 2,
+                                 is_training, 'block_layer4', data_format)
+            inputs = tf.layers.average_pooling2d(
+                inputs=inputs,
+                pool_size=7,
+                strides=1,
+                padding='VALID',
+                data_format=data_format)
+            inputs = tf.identity(inputs, 'final_avg_pool')
+            inputs = tf.reshape(inputs,
+                                [-1, 512 if block_fn is basicblock else 2048])
+            inputs = tf.layers.dense(inputs=inputs, units=num_classes)
+            inputs = tf.identity(inputs, 'final_dense')
+            return inputs
+
+        return model
+
+    model_params = {
+        18: {
+            'block': basicblock,
+            'layers': [2, 2, 2, 2]
+        },
+        34: {
+            'block': basicblock,
+            'layers': [3, 4, 6, 3]
+        },
+        50: {
+            'block': bottleneck,
+            'layers': [3, 4, 6, 3]
+        },
+        101: {
+            'block': bottleneck,
+            'layers': [3, 4, 23, 3]
+        },
+        152: {
+            'block': bottleneck,
+            'layers': [3, 8, 36, 3]
+        },
+        200: {
+            'block': bottleneck,
+            'layers': [3, 24, 36, 3]
+        }
+    }
+    if depth not in model_params:
+        raise ValueError('Not a valid depth:', depth)
+    params = model_params[depth]
+    return resnet_generator(params['block'], params['layers'], class_dim,
+                            data_format)
+
+
+def resnet_cifar10(depth, num_classes, data_format):
+    if depth % 6 != 2:
+        raise ValueError('depth must be 6n + 2:', depth)
+
+    num_blocks = (depth - 2) // 6
+
+    if data_format is None:
+        data_format = ('channels_first'
+                       if tf.test.is_built_with_cuda() else 'channels_last')
+
+    def model(inputs, is_training):
+        inputs = conv_bn(inputs, 16, 3, 1, is_training, data_format)
+        inputs = tf.identity(inputs, 'initial_conv')
+        inputs = block_layer(inputs, 16, basicblock, num_blocks, 1, is_training,
+                             'block_layer1', data_format)
+        inputs = block_layer(inputs, 32, basicblock, num_blocks, 2, is_training,
+                             'block_layer2', data_format)
+        inputs = block_layer(inputs, 64, basicblock, num_blocks, 2, is_training,
+                             'block_layer3', data_format)
+        inputs = tf.layers.average_pooling2d(
+            inputs=inputs,
+            pool_size=8,
+            strides=1,
+            padding='VALID',
+            data_format=data_format)
+        inputs = tf.identity(inputs, 'final_avg_pool')
+        inputs = tf.reshape(inputs, [-1, 64])
+        inputs = tf.layers.dense(inputs=inputs, units=num_classes)
+        inputs = tf.identity(inputs, 'final_dense')
+        return inputs
+
+    return model
+
+
+def run_benchmark(args, data_format='channels_last', device='/cpu:0'):
+    """Our model_fn for ResNet to be used with our Estimator."""
+
+    class_dim = 1000
+    dshape = (None, 224, 224, 3)
+
+    pdshape = (3, 224, 224)
+    if args.data == 'flowers102':
+        class_dim = 102
+        dshape = (None, 224, 224, 3)
+        pdshape = (3, 224, 224)
+    elif args.data == 'cifar10':
+        class_dim = 10
+        dshape = (None, 32, 32, 3)
+        pdshape = (3, 32, 32)
+
+    with tf.device(device):
+        images = tf.placeholder(DTYPE, shape=dshape)
+        labels = tf.placeholder(tf.int64, shape=(None, ))
+        is_training = tf.placeholder('bool')
+        onehot_labels = tf.one_hot(labels, depth=class_dim)
+
+        network = resnet_cifar10(
+            32, class_dim,
+            data_format) if args.data == 'cifar10' else resnet_imagenet(
+                50, class_dim, data_format)
+
+        logits = network(inputs=images, is_training=is_training)
+
+        cross_entropy = tf.losses.softmax_cross_entropy(
+            logits=logits, onehot_labels=onehot_labels)
+        avg_cost = tf.reduce_mean(cross_entropy)
+
+        correct = tf.equal(tf.argmax(logits, 1), labels)
+        accuracy = tf.reduce_mean(tf.cast(correct, tf.float32))
+
+        lr = 0.1 if args.data == 'cifar10' else 0.01
+        optimizer = tf.train.MomentumOptimizer(learning_rate=lr, momentum=0.9)
+
+        # Batch norm requires update_ops to be added as a train_op dependency.
+        update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
+        with tf.control_dependencies(update_ops):
+            train_op = optimizer.minimize(avg_cost)
+
+    train_reader = paddle.batch(
+        paddle.reader.shuffle(
+            paddle.dataset.cifar.train10()
+            if args.data == 'cifar10' else paddle.dataset.flowers.train(),
+            buf_size=5120),
+        batch_size=args.batch_size)
+    test_reader = paddle.batch(
+        paddle.dataset.cifar.test10()
+        if args.data == 'cifar10' else paddle.dataset.flowers.test(),
+        batch_size=100)
+
+    def test():
+        test_accs = []
+        for batch_id, data in enumerate(test_reader()):
+            test_images = np.array(
+                map(lambda x: np.transpose(x[0].reshape(pdshape),
+                axes=[1, 2, 0]), data)).astype("float32")
+            test_labels = np.array(map(lambda x: x[1], data)).astype('int64')
+            test_accs.append(
+                accuracy.eval(feed_dict={
+                    images: test_images,
+                    labels: test_labels,
+                    is_training: False
+                }))
+        print("Pass = %d, Train performance = %f imgs/s, Test accuracy = %f\n" %
+              (pass_id, num_samples / train_elapsed, np.mean(test_accs)))
+
+    config = tf.ConfigProto(
+        intra_op_parallelism_threads=1, inter_op_parallelism_threads=1)
+    config.gpu_options.allow_growth = True
+
+    with tf.Session(config=config) as sess:
+        init_g = tf.global_variables_initializer()
+        init_l = tf.local_variables_initializer()
+        sess.run(init_g)
+        sess.run(init_l)
+
+        if args.use_fake_data:
+            data = train_reader().next()
+            images_data = np.array(
+                    map(lambda x: np.transpose(x[0].reshape(pdshape),
+                    axes=[1, 2, 0]), data)).astype("float32")
+            labels_data = np.array(map(lambda x: x[1], data)).astype('int64')
+        iters, num_samples, start_time = 0, 0, 0.0
+        for pass_id in range(args.pass_num):
+            if iters == args.iterations:
+                break
+            train_accs = []
+            train_losses = []
+            for batch_id, data in enumerate(train_reader()):
+                if iters == args.skip_batch_num:
+                    start_time = time.time()
+                    num_samples = 0
+                if iters == args.iterations:
+                    break
+                if not args.use_fake_data:
+                    images_data = np.array(
+                        map(lambda x: np.transpose(x[0].reshape(pdshape),
+                        axes=[1, 2, 0]), data)).astype("float32")
+                    labels_data = np.array(map(lambda x: x[1], data)).astype(
+                        'int64')
+                _, loss, acc = sess.run([train_op, avg_cost, accuracy],
+                                        feed_dict={
+                                            images: images_data,
+                                            labels: labels_data,
+                                            is_training: True
+                                        })
+                iters += 1
+                train_accs.append(acc)
+                train_losses.append(loss)
+                num_samples += len(data)
+                print("Pass=%d, Iter=%d, Loss=%f, Accuray=%f\n" %
+                      (pass_id, iters, loss, acc))
+
+            train_elapsed = time.time() - start_time
+            print("Pass=%d, Loss=%f, Accuray=%f\n" %
+                  (pass_id, np.mean(train_losses), np.mean(train_accs)))
+
+            # evaluation
+            if args.with_test:
+                test()
+
+        if not args.with_test:
+            duration = time.time() - start_time
+            examples_per_sec = num_samples / duration
+            sec_per_batch = duration / (iters - args.skip_batch_num)
+
+            print('Total examples: %d, total time: %.5f' %
+                  (num_samples, duration))
+            print('%.5f examples/sec, %.5f sec/batch' %
+                  (examples_per_sec, sec_per_batch))
+
+
+if __name__ == '__main__':
+    args = parse_args()
+    print_arguments(args)
+    if tf.test.is_built_with_cuda():
+        device = '/device:GPU:0'
+        if args.order == 'NHWC':
+            data_format = 'channels_last'
+        else:
+            data_format = 'channels_first'
+    else:
+        device = '/cpu:0'
+        if args.order == 'NHWC':
+            data_format = 'channels_last'
+        else:
+            raise ValueError('Only support NHWC order in CPU mode')
+
+    run_benchmark(args, data_format, device)

benchmark/tensorflow/stacked_dynamic_lstm.py

+#   Copyright (c) 2018 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 __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+import argparse
+import time
+import tensorflow as tf
+
+import paddle.v2 as paddle
+
+
+def parse_args():
+    parser = argparse.ArgumentParser("LSTM model benchmark.")
+    parser.add_argument(
+        '--batch_size',
+        type=int,
+        default=32,
+        help='The sequence number of a batch data. (default: %(default)d)')
+    parser.add_argument(
+        '--stacked_num',
+        type=int,
+        default=5,
+        help='Number of lstm layers to stack. (default: %(default)d)')
+    parser.add_argument(
+        '--embedding_dim',
+        type=int,
+        default=512,
+        help='Dimension of embedding table. (default: %(default)d)')
+    parser.add_argument(
+        '--hidden_dim',
+        type=int,
+        default=512,
+        help='Hidden size of lstm unit. (default: %(default)d)')
+    parser.add_argument(
+        '--pass_num',
+        type=int,
+        default=10,
+        help='Epoch number to train. (default: %(default)d)')
+    parser.add_argument(
+        '--learning_rate',
+        type=float,
+        default=0.0002,
+        help='Learning rate used to train. (default: %(default)f)')
+    parser.add_argument(
+        '--infer_only', action='store_true', help='If set, run forward only.')
+    args = parser.parse_args()
+    return args
+
+
+def print_arguments(args):
+    print('-----------  Configuration Arguments -----------')
+    for arg, value in sorted(vars(args).iteritems()):
+        print('%s: %s' % (arg, value))
+    print('------------------------------------------------')
+
+
+def dynamic_lstm_model(dict_size,
+                       embedding_dim,
+                       hidden_dim,
+                       stacked_num,
+                       class_num=2,
+                       is_train=True):
+    word_idx = tf.placeholder(tf.int64, shape=[None, None])
+    sequence_length = tf.placeholder(tf.int64, shape=[None, ])
+
+    embedding_weights = tf.get_variable('word_embeddings',
+                                        [dict_size, embedding_dim])
+    embedding = tf.nn.embedding_lookup(embedding_weights, word_idx)
+
+    lstm_cell = tf.nn.rnn_cell.LSTMCell(
+        num_units=hidden_dim, use_peepholes=False)
+    stacked_cell = tf.nn.rnn_cell.MultiRNNCell([lstm_cell] * stacked_num)
+
+    # final_state [LSTMTuple(c, h), LSTMTuple(c, h) ...] total stacked_num LSTMTuples
+    _, final_state = tf.nn.dynamic_rnn(
+        cell=stacked_cell,
+        inputs=embedding,
+        dtype=tf.float32,
+        sequence_length=sequence_length)
+
+    w = tf.Variable(
+        tf.truncated_normal([hidden_dim, class_num]), dtype=tf.float32)
+    bias = tf.Variable(
+        tf.constant(
+            value=0.0, shape=[class_num], dtype=tf.float32))
+    prediction = tf.matmul(final_state[-1][1], w) + bias
+
+    if not is_train:
+        return (word_idx, sequence_length), tf.nn.softmax(prediction)
+
+    label = tf.placeholder(tf.int64, shape=[None, ])
+    loss = tf.nn.softmax_cross_entropy_with_logits(
+        labels=tf.one_hot(label, 2), logits=prediction)
+    avg_loss = tf.reduce_mean(loss)
+
+    correct_count = tf.equal(tf.argmax(prediction, 1), label)
+    acc = tf.reduce_mean(tf.cast(correct_count, tf.float32))
+
+    with tf.variable_scope("reset_metrics_accuracy_scope") as scope:
+        g_acc = tf.metrics.accuracy(label, tf.argmax(prediction, axis=1))
+        vars = tf.contrib.framework.get_variables(
+            scope, collection=tf.GraphKeys.LOCAL_VARIABLES)
+        reset_op = tf.variables_initializer(vars)
+
+    return (word_idx, sequence_length, label), avg_loss, acc, g_acc, reset_op
+
+
+def padding_data(data, padding_size, value):
+    data = data + [value] * padding_size
+    return data[:padding_size]
+
+
+def train(args):
+    word_dict = paddle.dataset.imdb.word_dict()
+    dict_size = len(word_dict)
+
+    feeding_list, avg_loss, acc, g_acc, reset_op = dynamic_lstm_model(
+        dict_size, args.embedding_dim, args.hidden_dim, args.stacked_num)
+
+    adam_optimizer = tf.train.AdamOptimizer(learning_rate=args.learning_rate)
+    train_op = adam_optimizer.minimize(avg_loss)
+
+    train_reader = paddle.batch(
+        paddle.reader.shuffle(
+            paddle.dataset.imdb.train(word_dict), buf_size=25000),
+        batch_size=args.batch_size)
+
+    test_reader = paddle.batch(
+        paddle.reader.shuffle(
+            paddle.dataset.imdb.test(word_dict), buf_size=25000),
+        batch_size=args.batch_size)
+
+    def do_validation(sess):
+        sess.run(reset_op)
+        for batch_id, data in enumerate(test_reader()):
+            word_idx = map(lambda x: x[0], data)
+            sequence_length = np.array(
+                [len(seq) for seq in word_idx]).astype('int64')
+            maxlen = np.max(sequence_length)
+            word_idx = [padding_data(seq, maxlen, 0) for seq in word_idx]
+            word_idx = np.array(word_idx).astype('int64')
+            label = np.array(map(lambda x: x[1], data)).astype('int64')
+
+            _, loss, fetch_acc, fetch_g_acc = sess.run(
+                [train_op, avg_loss, acc, g_acc],
+                feed_dict={
+                    feeding_list[0]: word_idx,
+                    feeding_list[1]: sequence_length,
+                    feeding_list[2]: label
+                })
+
+        return fetch_g_acc[1]
+
+    config = tf.ConfigProto(
+        intra_op_parallelism_threads=1, inter_op_parallelism_threads=1)
+    config.gpu_options.allow_growth = True
+    with tf.Session(config=config) as sess:
+        init_g = tf.global_variables_initializer()
+        init_l = tf.local_variables_initializer()
+        sess.run(init_l)
+        sess.run(init_g)
+
+        for pass_id in xrange(args.pass_num):
+            # clear accuracy local variable 
+            sess.run(reset_op)
+            pass_start_time = time.time()
+            words_seen = 0
+
+            for batch_id, data in enumerate(train_reader()):
+                word_idx = map(lambda x: x[0], data)
+                sequence_length = np.array(
+                    [len(seq) for seq in word_idx]).astype('int64')
+                words_seen += np.sum(sequence_length)
+                maxlen = np.max(sequence_length)
+                word_idx = [padding_data(seq, maxlen, 0) for seq in word_idx]
+                word_idx = np.array(word_idx).astype('int64')
+                label = np.array(map(lambda x: x[1], data)).astype('int64')
+
+                _, loss, fetch_acc, fetch_g_acc = sess.run(
+                    [train_op, avg_loss, acc, g_acc],
+                    feed_dict={
+                        feeding_list[0]: word_idx,
+                        feeding_list[1]: sequence_length,
+                        feeding_list[2]: label
+                    })
+
+                print("pass_id=%d, batch_id=%d, loss: %f, acc: %f, avg_acc: %f"
+                      % (pass_id, batch_id, loss, fetch_acc, fetch_g_acc[1]))
+
+            pass_end_time = time.time()
+            time_consumed = pass_end_time - pass_start_time
+            words_per_sec = words_seen / time_consumed
+            test_acc = do_validation(sess)
+            print("pass_id=%d, test_acc: %f, words/s: %f, sec/pass: %f" %
+                  (pass_id, test_acc, words_per_sec, time_consumed))
+
+
+if __name__ == '__main__':
+    args = parse_args()
+    print_arguments(args)
+
+    if args.infer_only:
+        pass
+    else:
+        train(args)

benchmark/tensorflow/vgg.py

+#   Copyright (c) 2018 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.
+"""VGG16 benchmark in TensorFlow"""
+import tensorflow as tf
+import paddle.v2 as paddle
+import numpy as np
+import argparse
+import time
+
+parser = argparse.ArgumentParser(description=__doc__)
+parser.add_argument(
+    '--batch_size', type=int, default=128, help="Batch size for training.")
+parser.add_argument(
+    '--skip_batch_num',
+    type=int,
+    default=5,
+    help='The first num of minibatch num to skip, for better performance test')
+parser.add_argument(
+    '--iterations', type=int, default=80, help='The number of minibatches.')
+parser.add_argument(
+    '--learning_rate',
+    type=float,
+    default=1e-3,
+    help="Learning rate for training.")
+parser.add_argument('--num_passes', type=int, default=50, help="No. of passes.")
+parser.add_argument(
+    '--device',
+    type=str,
+    default='GPU',
+    choices=['CPU', 'GPU'],
+    help="The device type.")
+parser.add_argument(
+    '--data_format',
+    type=str,
+    default='NHWC',
+    choices=['NCHW', 'NHWC'],
+    help='The data order, NCHW=[batch, channels, height, width].'
+    'Only support NHWC right now.')
+parser.add_argument(
+    '--data_set',
+    type=str,
+    default='cifar10',
+    choices=['cifar10', 'flowers'],
+    help='Optional dataset for benchmark.')
+args = parser.parse_args()
+
+
+class VGG16Model(object):
+    def __init__(self):
+        self.parameters = []
+
+    def batch_norm_relu(self, inputs, is_training):
+        """Performs a batch normalization followed by a ReLU."""
+        # We set fused=True for a significant speed boost. See
+        # https://www.tensorflow.org/speed/speed_guide#common_fused_ops
+        inputs = tf.layers.batch_normalization(
+            inputs=inputs,
+            axis=1 if args.data_format == 'NCHW' else -1,
+            momentum=0.9,
+            epsilon=1e-05,
+            center=True,
+            scale=True,
+            training=is_training,
+            fused=True)
+        inputs = tf.nn.relu(inputs)
+        return inputs
+
+    def conv_bn_layer(self,
+                      name,
+                      images,
+                      kernel_shape,
+                      is_training,
+                      drop_rate=0.0):
+        with tf.name_scope(name) as scope:
+            kernel = tf.Variable(
+                tf.truncated_normal(
+                    kernel_shape, dtype=tf.float32, stddev=1e-1),
+                name='weights')
+            conv = tf.nn.conv2d(
+                images,
+                kernel, [1, 1, 1, 1],
+                data_format=args.data_format,
+                padding='SAME')
+            biases = tf.Variable(
+                tf.constant(
+                    0.0, shape=[kernel_shape[-1]], dtype=tf.float32),
+                trainable=True,
+                name='biases')
+            out = tf.nn.bias_add(conv, biases)
+            out = self.batch_norm_relu(out, is_training)
+            out = tf.layers.dropout(out, rate=drop_rate, training=is_training)
+            return out
+
+    def fc_layer(self, name, inputs, shape):
+        with tf.name_scope(name) as scope:
+            fc_w = tf.Variable(
+                tf.truncated_normal(
+                    shape, dtype=tf.float32, stddev=1e-1),
+                name='weights')
+            fc_b = tf.Variable(
+                tf.constant(
+                    0.0, shape=[shape[-1]], dtype=tf.float32),
+                trainable=True,
+                name='biases')
+            out = tf.nn.bias_add(tf.matmul(inputs, fc_w), fc_b)
+            return out
+
+    def network(self, images, class_dim, is_training):
+        """ VGG16 model structure.
+
+            TODO(kuke): enable this network to support the 'NCHW' data format
+        """
+
+        # conv1
+        conv1_1 = self.conv_bn_layer(
+            'conv1_1', images, [3, 3, 3, 64], is_training, drop_rate=0.3)
+        conv1_2 = self.conv_bn_layer(
+            'conv1_2', conv1_1, [3, 3, 64, 64], is_training, drop_rate=0.0)
+        # pool1
+        pool1 = tf.nn.max_pool(
+            conv1_2,
+            ksize=[1, 2, 2, 1],
+            strides=[1, 2, 2, 1],
+            padding='SAME',
+            name='pool1')
+        # conv2
+        conv2_1 = self.conv_bn_layer(
+            'conv2_1', pool1, [3, 3, 64, 128], is_training, drop_rate=0.4)
+        conv2_2 = self.conv_bn_layer(
+            'conv2_2', conv2_1, [3, 3, 128, 128], is_training, drop_rate=0.0)
+        # pool2
+        pool2 = tf.nn.max_pool(
+            conv2_2,
+            ksize=[1, 2, 2, 1],
+            strides=[1, 2, 2, 1],
+            padding='SAME',
+            name='pool2')
+        # conv3
+        conv3_1 = self.conv_bn_layer(
+            'conv3_1', pool2, [3, 3, 128, 256], is_training, drop_rate=0.4)
+        conv3_2 = self.conv_bn_layer(
+            'conv3_2', conv3_1, [3, 3, 256, 256], is_training, drop_rate=0.4)
+        conv3_3 = self.conv_bn_layer(
+            'conv3_3', conv3_2, [3, 3, 256, 256], is_training, drop_rate=0.0)
+        # pool3
+        pool3 = tf.nn.max_pool(
+            conv3_3,
+            ksize=[1, 2, 2, 1],
+            strides=[1, 2, 2, 1],
+            padding='SAME',
+            name='pool3')
+        # conv4
+        conv4_1 = self.conv_bn_layer(
+            'conv4_1', pool3, [3, 3, 256, 512], is_training, drop_rate=0.4)
+        conv4_2 = self.conv_bn_layer(
+            'conv4_2', conv4_1, [3, 3, 512, 512], is_training, drop_rate=0.4)
+        conv4_3 = self.conv_bn_layer(
+            'conv4_3', conv4_2, [3, 3, 512, 512], is_training, drop_rate=0.0)
+        # pool4
+        pool4 = tf.nn.max_pool(
+            conv4_3,
+            ksize=[1, 2, 2, 1],
+            strides=[1, 2, 2, 1],
+            padding='SAME',
+            name='pool4')
+        # conv5
+        conv5_1 = self.conv_bn_layer(
+            'conv5_1', pool4, [3, 3, 512, 512], is_training, drop_rate=0.4)
+        conv5_2 = self.conv_bn_layer(
+            'conv5_2', conv5_1, [3, 3, 512, 512], is_training, drop_rate=0.4)
+        conv5_3 = self.conv_bn_layer(
+            'conv5_3', conv5_2, [3, 3, 512, 512], is_training, drop_rate=0.0)
+        # pool5
+        pool5 = tf.nn.max_pool(
+            conv5_3,
+            ksize=[1, 2, 2, 1],
+            strides=[1, 2, 2, 1],
+            padding='SAME',
+            name='pool4')
+        # flatten
+        shape = int(np.prod(pool5.get_shape()[1:]))
+        pool5_flat = tf.reshape(pool5, [-1, shape])
+        # fc1
+        drop = tf.layers.dropout(pool5_flat, rate=0.5, training=is_training)
+        fc1 = self.fc_layer('fc1', drop, [shape, 512])
+        # fc2
+        bn = self.batch_norm_relu(fc1, is_training)
+        drop = tf.layers.dropout(bn, rate=0.5, training=is_training)
+        fc2 = self.fc_layer('fc2', drop, [512, 512])
+
+        fc3 = self.fc_layer('fc3', fc2, [512, class_dim])
+
+        return fc3
+
+
+def run_benchmark():
+    """Run benchmark on cifar10 or flowers."""
+
+    if args.data_set == "cifar10":
+        class_dim = 10
+        raw_shape = (3, 32, 32)
+        dat_shape = (None, 32, 32, 3) if args.data_format == 'NHWC' else (
+            None, 3, 32, 32)
+    else:
+        class_dim = 102
+        raw_shape = (3, 224, 224)
+        dat_shape = (None, 224, 224, 3) if args.data_format == 'NHWC' else (
+            None, 3, 224, 224)
+
+    device = '/cpu:0' if args.device == 'CPU' else '/device:GPU:0'
+
+    with tf.device(device):
+        images = tf.placeholder(tf.float32, shape=dat_shape)
+        labels = tf.placeholder(tf.int64, shape=(None, ))
+        is_training = tf.placeholder('bool')
+        onehot_labels = tf.one_hot(labels, depth=class_dim)
+
+        vgg16 = VGG16Model()
+        logits = vgg16.network(images, class_dim, is_training)
+        loss = tf.losses.softmax_cross_entropy(
+            onehot_labels=onehot_labels, logits=logits)
+        avg_loss = tf.reduce_mean(loss)
+
+        correct = tf.equal(tf.argmax(logits, 1), labels)
+        accuracy = tf.reduce_mean(tf.cast(correct, tf.float32))
+
+        optimizer = tf.train.AdamOptimizer(learning_rate=args.learning_rate)
+        update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
+        with tf.control_dependencies(update_ops):
+            train_op = optimizer.minimize(avg_loss)
+
+    # data reader
+    train_reader = paddle.batch(
+        paddle.reader.shuffle(
+            paddle.dataset.cifar.train10()
+            if args.data_set == 'cifar10' else paddle.dataset.flowers.train(),
+            buf_size=5120),
+        batch_size=args.batch_size)
+    test_reader = paddle.batch(
+        paddle.reader.shuffle(
+            paddle.dataset.cifar.test10()
+            if args.data_set == 'cifar10' else paddle.dataset.flowers.test(),
+            buf_size=5120),
+        batch_size=args.batch_size)
+
+    # test
+    def test():
+        test_accs = []
+        for batch_id, data in enumerate(test_reader()):
+            test_images = np.array(
+         map(lambda x: np.transpose(x[0].reshape(raw_shape),
+         axes=[1, 2, 0]) if args.data_format == 'NHWC' else x[0], data)).astype("float32")
+            test_labels = np.array(map(lambda x: x[1], data)).astype('int64')
+            test_accs.append(
+                accuracy.eval(feed_dict={
+                    images: test_images,
+                    labels: test_labels,
+                    is_training: False
+                }))
+        return np.mean(test_accs)
+
+    config = tf.ConfigProto(
+        intra_op_parallelism_threads=1, inter_op_parallelism_threads=1)
+    config.gpu_options.allow_growth = True
+
+    with tf.Session(config=config) as sess:
+        init_g = tf.global_variables_initializer()
+        init_l = tf.local_variables_initializer()
+        sess.run(init_g)
+        sess.run(init_l)
+        iters, num_samples, start_time = 0, 0, time.time()
+        for pass_id in range(args.num_passes):
+            # train
+            num_samples = 0
+            start_time = time.time()
+            for batch_id, data in enumerate(train_reader()):
+                if iters == args.skip_batch_num:
+                    start_time = time.time()
+                    num_samples = 0
+                if iters == args.iterations:
+                    break
+                train_images = np.array(
+                    map(lambda x: np.transpose(x[0].reshape(raw_shape),
+                    axes=[1, 2, 0]) if args.data_format == 'NHWC' else x[0], data)).astype("float32")
+                train_labels = np.array(map(lambda x: x[1], data)).astype(
+                    'int64')
+                _, loss, acc = sess.run([train_op, avg_loss, accuracy],
+                                        feed_dict={
+                                            images: train_images,
+                                            labels: train_labels,
+                                            is_training: True
+                                        })
+                iters += 1
+                num_samples += len(data)
+                print("Pass = %d, Iters = %d, Loss = %f, Accuracy = %f" %
+                      (pass_id, iters, loss, acc))
+            train_elapsed = time.time() - start_time
+            # test
+            pass_test_acc = test()
+            print("Pass = %d, Train speed = %f imgs/s, Test accuracy = %f\n" %
+                  (pass_id, num_samples / train_elapsed, pass_test_acc))
+
+
+def print_arguments():
+    print('-----------  Configuration Arguments -----------')
+    for arg, value in sorted(vars(args).iteritems()):
+        print('%s: %s' % (arg, value))
+    print('------------------------------------------------')
+
+
+if __name__ == '__main__':
+    print_arguments()
+    run_benchmark()

cmake/external/mkldnn.cmake

@@ -36,7 +36,8 @@ MESSAGE(STATUS "Set ${MKLDNN_INSTALL_DIR}/lib to runtime path")
 SET(CMAKE_INSTALL_RPATH_USE_LINK_PATH TRUE)
 SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${MKLDNN_INSTALL_DIR}/lib")
 
-INCLUDE_DIRECTORIES(${MKLDNN_INC_DIR})
+INCLUDE_DIRECTORIES(${MKLDNN_INC_DIR}) # For MKLDNN code to include internal headers.
+INCLUDE_DIRECTORIES(${THIRD_PARTY_PATH}/install) # For Paddle code to include mkldnn.h
 
 IF(${CBLAS_PROVIDER} STREQUAL "MKLML")
     SET(MKLDNN_DEPENDS   ${MKLML_PROJECT})

doc/fluid/design/concepts/index_cn.rst

@@ -16,3 +16,4 @@
   block.md
   scope.md
   executor.md
+  parallel_executor.md

doc/fluid/design/concepts/index_en.rst

@@ -16,3 +16,4 @@ Core Concepts
   block.md
   scope.md
   executor.md
+  parallel_executor.md

doc/fluid/design/muti_devices/kernel_hint_design.md

-# Problem
+# Kernel Hint Design
+
+## Problem
 In PaddlePaddle's [Design](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/switch_kernel.md), one Operator may have multiple kernels. Users may have some personal preference to choose a certain type of kernel for an operator, such as `force_cpu` to choose a CPU kernel, `use_cudnn` to choose a CUDNN kernel, we need to provide a way for users to do this.
 
 In the current design, we use KernelType to describe one kernel.

doc/fluid/design/muti_devices/kernel_selection.md

-# Background
+# Kernel Selection
+
+## Background
 Every operator has many kernels because there are multiple data types, places, data layout, library type that Fluid supports. We use the `OpKernelType ` to describe kernel types that operators can hold.
 
 The `OpKernelType ` is as follows:

doc/v2/build_and_install/index_en.rst

-Install and Build
-=================
+install and Compile
+==========
 
 .. _install_steps:
 
-Install Steps
-++++++++
+PaddlePaddle provides various methods of installation for many different users
 
-You can choose either pip or Docker to complete your install:
+Focus on Deep Learning Model Development
+-----------------
+
+PaddlePaddle provides lots of packages of python wheel , that pip can install:
 
 .. toctree::
 	:maxdepth: 1
 
 	pip_install_en.rst
+
+This is the most convenient way of installation. Please choose the right installation package with machine configure and system.
+
+Follow the Bottom Frame
+----------
+
+PaddlePaddle also supports installation using Docker. Please refer to the tutorial below:
+
+.. toctree::
+	:maxdepth: 1
+
 	docker_install_en.rst
 
-Build from Source
------------------
+We recommend running PaddlePaddle in Docker. This method has the following advantages：
 
-..  warning::
+- Does not require installation of third-party dependencies. 
+- Easy to share runtime environment. 
 
-    We recommend to directly install via above installation steps, you'll only need to build PaddlePaddle from source when you need a modifed binary.
+Lastly, users can also compile and install PaddlePaddle from source code. The instructions are below:
 
 .. toctree::
     :maxdepth: 1
 
-    build_from_source_en.md
+    build_from_source_en.rst
+
+.. warning::
+
+	One caveat with this approach is that developers will have to download, compile and install all third-party dependencies. Thus this process of installation is more time consuming.
+
 
 FAQ
-++++++++++
+-----------
+
+For any problems during installation, please refer to the page below for answers:
+
+:ref:`常见问题解答 <install_faq>`
+
+If the problem still persists, you are welcome to seek assistance from the PaddlePaddle community：
 
-`FAQ <http://www.paddlepaddle.org/docs/develop/documentation/zh/faq/build_and_install/index_en.html>`_
+`创建issue <https://github.com/PaddlePaddle/Paddle/issues/new>`_

paddle/.gitignore

+.timestamp
 *.o
 *.a
 .svn

paddle/fluid/framework/data_device_transform_test.cu

@@ -105,7 +105,7 @@ static void BuildVar(const std::string& param_name,
 TEST(Operator, CPUtoGPU) {
   using namespace paddle::framework;
   using namespace paddle::platform;
-  InitDevices();
+  InitDevices(true);
 
   paddle::framework::Scope scope;
   paddle::platform::CPUPlace cpu_place;

paddle/fluid/framework/details/multi_devices_graph_builder.cc

@@ -59,7 +59,11 @@ std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
   auto graph = new SSAGraph();
   SSAGraph &result = *graph;
   std::unordered_set<std::string> og_has_been_broadcast;
-  result.vars_.resize(places_.size());
+
+  // We cannot invoke resize. It is a bug of GCC 4.8
+  result.vars_ = std::vector<
+      std::unordered_map<std::string, std::vector<std::unique_ptr<VarHandle>>>>(
+      places_.size());
 
   bool is_forwarding = true;
   for (auto *op : program.Block(0).AllOps()) {
@@ -147,15 +151,16 @@ std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
             if (vars.empty()) {  // This device has no data. continue.
               continue;
             }
-            auto *prev_grad = &vars[vars.size() - 1];
-            op_handle->AddInput(prev_grad);
+            auto &prev_grad = vars[vars.size() - 1];
+            op_handle->AddInput(prev_grad.get());
 
-            auto &var = vars[vars.size()];
-            var.place_ = p;
-            var.name_ = og;
-            var.version_ = vars.size() - 1;
+            vars.emplace_back(new VarHandle);
+            auto &var = vars.back();
+            var->place_ = p;
+            var->name_ = og;
+            var->version_ = vars.size() - 1;
 
-            op_handle->AddOutput(&var);
+            op_handle->AddOutput(var.get());
           }
 #else
           PADDLE_ENFORCE("Not implemented");

paddle/fluid/framework/details/ssa_graph.h

@@ -16,6 +16,8 @@
 
 #include <map>
 #include <string>
+#include <vector>
+
 #include "paddle/fluid/framework/details/op_handle_base.h"
 #include "paddle/fluid/framework/details/var_handle.h"
 
@@ -24,7 +26,9 @@ namespace framework {
 namespace details {
 
 struct SSAGraph {
-  std::vector<std::unordered_map<std::string, std::map<int, VarHandle>>> vars_;
+  std::vector<
+      std::unordered_map<std::string, std::vector<std::unique_ptr<VarHandle>>>>
+      vars_;
   // aux variables to represent dependency. Useful to resolve data hazard.
   std::unordered_set<std::unique_ptr<VarHandleBase>> dep_vars_;
   std::vector<std::unique_ptr<OpHandleBase>> ops_;

paddle/fluid/framework/details/ssa_graph_builder.cc

@@ -27,8 +27,8 @@ void SSAGraphBuilder::PolishGraphToSupportDataHazards(SSAGraph *graph) {
       auto it_old = name_pair.second.rbegin();
       ++it_old;
       for (; it_old != name_pair.second.rend(); it_new = it_old, ++it_old) {
-        auto *write_op = it_new->second.generated_op_;
-        auto &read_ops = it_old->second.pending_ops_;
+        auto *write_op = (*it_new)->generated_op_;
+        auto &read_ops = (*it_old)->pending_ops_;
 
         for (auto *read_op : read_ops) {
           // Manually add a dependency var from read_op to write_op;
@@ -54,14 +54,15 @@ VarHandle *SSAGraphBuilder::CreateOrGetLatestVarHandle(
   auto &var_holder = var_holders[each_var_name];
   VarHandle *var = nullptr;
   if (var_holder.empty()) {
+    var_holder.emplace_back(new VarHandle);
     auto &init_var = var_holder[0];
-    init_var.place_ = place;
-    init_var.name_ = each_var_name;
-    init_var.generated_op_ = nullptr;
-    init_var.version_ = 0;
-    var = &init_var;
+    init_var->place_ = place;
+    init_var->name_ = each_var_name;
+    init_var->generated_op_ = nullptr;
+    init_var->version_ = 0;
+    var = init_var.get();
   } else {
-    var = &var_holder.rbegin()->second;
+    var = var_holder.rbegin()->get();
   }
   return var;
 }
@@ -72,11 +73,12 @@ void SSAGraphBuilder::CreateOpOutput(SSAGraph *graph, OpHandleBase *op_handle,
                                      size_t place_offset) {
   auto &vars = graph->vars_[place_offset][each_var_name];
   size_t version = vars.size();
-  auto &var = vars[version];
-  var.version_ = version;
-  var.name_ = each_var_name;
-  var.place_ = place;
-  op_handle->AddOutput(&var);
+  vars.emplace_back(new VarHandle());
+  auto &var = vars.back();
+  var->version_ = version;
+  var->name_ = each_var_name;
+  var->place_ = place;
+  op_handle->AddOutput(var.get());
 }
 
 template <typename Callback>
@@ -84,7 +86,7 @@ void IterAllVar(const SSAGraph &graph, Callback callback) {
   for (auto &each : graph.vars_) {
     for (auto &pair1 : each) {
       for (auto &pair2 : pair1.second) {
-        callback(pair2.second);
+        callback(*pair2);
       }
     }
   }

paddle/fluid/framework/details/threaded_ssa_graph_executor.cc

@@ -69,7 +69,7 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
   for (auto &var_map : graph_->vars_) {
     for (auto &name_pair : var_map) {
       for (auto &version_pair : name_pair.second) {
-        InsertPendingVar(version_pair.second);
+        InsertPendingVar(*version_pair);
       }
     }
   }
@@ -95,7 +95,7 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
     for (auto &var_map : graph_->vars_) {
       auto it = var_map.find(fetch_var_name);
       if (it != var_map.end()) {
-        fetched_vars[fetch_var_name].push_back(&it->second.rbegin()->second);
+        fetched_vars[fetch_var_name].push_back(it->second.rbegin()->get());
       }
     }
   }

paddle/fluid/framework/init.cc

@@ -64,7 +64,7 @@ void InitP2P(int count) {
 #endif
 }
 
-void InitDevices() {
+void InitDevices(bool init_p2p) {
   /*Init all avaiable devices by default */
 
   std::vector<platform::Place> places;
@@ -85,7 +85,9 @@ void InitDevices() {
   for (int i = 0; i < count; ++i) {
     places.emplace_back(platform::CUDAPlace(i));
   }
+  if (init_p2p) {
     InitP2P(count);
+  }
   platform::DeviceContextPool::Init(places);
 }
 

paddle/fluid/framework/init.h

@@ -24,7 +24,7 @@ void InitGflags(std::vector<std::string> &argv);
 
 void InitGLOG(const std::string &prog_name);
 
-void InitDevices();
+void InitDevices(bool init_p2p);
 
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/init_test.cc

@@ -21,7 +21,7 @@ TEST(InitDevices, CPU) {
   using paddle::platform::DeviceContextPool;
 
 #ifndef PADDLE_WITH_CUDA
-  InitDevices();
+  InitDevices(true);
   DeviceContextPool& pool = DeviceContextPool::Instance();
   ASSERT_EQ(pool.size(), 1U);
 #endif
@@ -33,7 +33,7 @@ TEST(InitDevices, CUDA) {
 
 #ifdef PADDLE_WITH_CUDA
   int count = paddle::platform::GetCUDADeviceCount();
-  InitDevices();
+  InitDevices(true);
   DeviceContextPool& pool = DeviceContextPool::Instance();
   ASSERT_EQ(pool.size(), 1U + static_cast<unsigned>(count));
 #endif

paddle/fluid/framework/lod_tensor_test.cu

@@ -30,7 +30,7 @@ __global__ void test(size_t* a, int size) {
 }
 
 TEST(LoD, data) {
-  paddle::framework::InitDevices();
+  paddle::framework::InitDevices(true);
 
   paddle::framework::LoD lod{{0, 1, 2}};
   lod.push_back({0, 2, 4, 5});
@@ -46,7 +46,7 @@ TEST(LoD, data) {
 }
 
 TEST(LoDTensor, LoDInGPU) {
-  paddle::framework::InitDevices();
+  paddle::framework::InitDevices(true);
 
   paddle::framework::LoDTensor lod_tensor;
   paddle::platform::CUDAPlace place(0);

paddle/fluid/framework/operator_test.cc

@@ -72,7 +72,7 @@ REGISTER_OP_WITHOUT_GRADIENT(test_operator,
                              paddle::framework::OpWithoutKernelCheckerMaker);
 
 TEST(OperatorBase, all) {
-  paddle::framework::InitDevices();
+  paddle::framework::InitDevices(true);
   paddle::framework::proto::OpDesc op_desc;
   op_desc.set_type("test_operator");
   BuildVar("input", {"IN1"}, op_desc.add_inputs());
@@ -198,7 +198,7 @@ REGISTER_OP_CPU_KERNEL(op_with_kernel,
 
 // test with single input
 TEST(OpKernel, all) {
-  paddle::framework::InitDevices();
+  paddle::framework::InitDevices(true);
   paddle::framework::proto::OpDesc op_desc;
   op_desc.set_type("op_with_kernel");
   BuildVar("x", {"IN1"}, op_desc.add_inputs());
@@ -228,7 +228,7 @@ REGISTER_OP_CPU_KERNEL(op_multi_inputs_with_kernel,
 TEST(OpKernel, multi_inputs) {
   using namespace paddle::framework;
 
-  paddle::framework::InitDevices();
+  paddle::framework::InitDevices(true);
   proto::OpDesc op_desc;
 
   op_desc.set_type("op_multi_inputs_with_kernel");
@@ -269,7 +269,7 @@ class OperatorClone : public paddle::framework::OperatorBase {
 };
 
 TEST(Operator, Clone) {
-  paddle::framework::InitDevices();
+  paddle::framework::InitDevices(true);
   OperatorClone a("ABC", paddle::framework::VariableNameMap{},
                   paddle::framework::VariableNameMap{},
                   paddle::framework::AttributeMap{});

paddle/fluid/framework/program_desc.cc

@@ -85,9 +85,9 @@ ProgramDesc::ProgramDesc(const std::string &binary_str) {
 }
 
 const std::vector<std::string> ProgramDesc::GetFeedTargetNames() {
-  BlockDesc *global_block = blocks_[0].get();
+  auto &global_block = Block(0);
   std::vector<std::string> feed_target_names;
-  for (auto *op : global_block->AllOps()) {
+  for (auto *op : global_block.AllOps()) {
     if (op->Type() == kFeedOpType) {
       feed_target_names.insert(feed_target_names.begin(), op->Output("Out")[0]);
     }
@@ -96,9 +96,9 @@ const std::vector<std::string> ProgramDesc::GetFeedTargetNames() {
 }
 
 const std::vector<std::string> ProgramDesc::GetFetchTargetNames() {
-  BlockDesc *global_block = blocks_[0].get();
+  auto &global_block = Block(0);
   std::vector<std::string> fetch_target_names;
-  for (auto *op : global_block->AllOps()) {
+  for (auto *op : global_block.AllOps()) {
     if (op->Type() == kFetchOpType) {
       fetch_target_names.push_back(op->Input("X")[0]);
     }
@@ -106,5 +106,43 @@ const std::vector<std::string> ProgramDesc::GetFetchTargetNames() {
   return fetch_target_names;
 }
 
+void ProgramDesc::SetFeedHolderName(const std::string &feed_holder_name) {
+  auto *global_block = MutableBlock(0);
+  int index = 0;
+  for (auto *op : global_block->AllOps()) {
+    if (op->Type() == kFeedOpType) {
+      // Unify the input's name of all feed_ops to feed_holder_name
+      global_block->RemoveVar(op->Input("X")[0]);
+      op->SetInput("X", {feed_holder_name});
+      op->SetAttr("col", {index});
+      op->CheckAttrs();
+      index++;
+    }
+  }
+
+  auto *feed_holder = global_block->Var(feed_holder_name);
+  feed_holder->SetType(proto::VarType::FEED_MINIBATCH);
+  feed_holder->SetPersistable(true);
+}
+
+void ProgramDesc::SetFetchHolderName(const std::string &fetch_holder_name) {
+  auto *global_block = MutableBlock(0);
+  int index = 0;
+  for (auto *op : global_block->AllOps()) {
+    if (op->Type() == kFetchOpType) {
+      // Unify the output's name of all fetch_ops to fetch_holder_name
+      global_block->RemoveVar(op->Output("Out")[0]);
+      op->SetOutput("Out", {fetch_holder_name});
+      op->SetAttr("col", {index});
+      op->CheckAttrs();
+      index++;
+    }
+  }
+
+  auto *fetch_holder = global_block->Var(fetch_holder_name);
+  fetch_holder->SetType(proto::VarType::FETCH_LIST);
+  fetch_holder->SetPersistable(true);
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/program_desc.h

@@ -15,6 +15,7 @@ limitations under the License. */
 #pragma once
 
 #include <memory>
+#include <string>
 #include <vector>
 #include "paddle/fluid/framework/block_desc.h"
 #include "paddle/fluid/framework/framework.pb.h"
@@ -52,9 +53,26 @@ class ProgramDesc {
 
   proto::ProgramDesc *Proto();
 
+  // The output variable of feed_op is referenced as feed_target.
+  // This function is used to collect the output variable's name of all
+  // feed_ops.
   const std::vector<std::string> GetFeedTargetNames();
+
+  // The input variable of fetch_op is referenced as fetch_target.
+  // This function is used to collect the input variable's name of all
+  // fetch_ops.
   const std::vector<std::string> GetFetchTargetNames();
 
+  // The input variable of feed_op that holds input Tensor provided by users is
+  // referenced as feed_holder.
+  // This function is used to change or unify the feed_holder variables' name.
+  void SetFeedHolderName(const std::string &feed_holder_name);
+
+  // The output variable of fetch_op that holds output Tensor needed by users is
+  // referenced as fetch_holder.
+  // This function is used to change or unify the fetch_holder variables' name.
+  void SetFetchHolderName(const std::string &fetch_holder_name);
+
  private:
   proto::ProgramDesc desc_;
 

paddle/fluid/inference/tests/book/test_inference_fit_a_line.cc

@@ -12,6 +12,7 @@ limitations under the License. */
 #include "gflags/gflags.h"
 #include "gtest/gtest.h"
 #include "paddle/fluid/inference/tests/test_helper.h"
+#include "paddle/fluid/inference/tests/test_multi_thread_helper.h"
 
 DEFINE_string(dirname, "", "Directory of the inference model.");
 
@@ -26,32 +27,63 @@ TEST(inference, fit_a_line) {
   // 0. Call `paddle::framework::InitDevices()` initialize all the devices
   // In unittests, this is done in paddle/testing/paddle_gtest_main.cc
 
-  paddle::framework::LoDTensor input;
+  for (int num_threads : {1, 2}) {
+    std::vector<std::vector<paddle::framework::LoDTensor*>> cpu_feeds;
+    cpu_feeds.resize(num_threads);
+    for (int i = 0; i < num_threads; ++i) {
+      auto* input = new paddle::framework::LoDTensor();
       // The second dim of the input tensor should be 13
       // The input data should be >= 0
       int64_t batch_size = 10;
-  SetupTensor<float>(&input, {batch_size, 13}, static_cast<float>(0),
+      SetupTensor<float>(input, {batch_size, 13}, static_cast<float>(0),
                          static_cast<float>(10));
-  std::vector<paddle::framework::LoDTensor*> cpu_feeds;
-  cpu_feeds.push_back(&input);
+      cpu_feeds[i].push_back(input);
+    }
 
-  paddle::framework::LoDTensor output1;
-  std::vector<paddle::framework::LoDTensor*> cpu_fetchs1;
-  cpu_fetchs1.push_back(&output1);
+    std::vector<std::vector<paddle::framework::LoDTensor*>> cpu_fetchs1;
+    cpu_fetchs1.resize(num_threads);
+    for (int i = 0; i < num_threads; ++i) {
+      auto* output = new paddle::framework::LoDTensor();
+      cpu_fetchs1[i].push_back(output);
+    }
 
     // Run inference on CPU
-  TestInference<paddle::platform::CPUPlace>(dirname, cpu_feeds, cpu_fetchs1);
-  LOG(INFO) << output1.dims();
+    LOG(INFO) << "--- CPU Runs (num_threads: " << num_threads << "): ---";
+    if (num_threads == 1) {
+      TestInference<paddle::platform::CPUPlace>(dirname, cpu_feeds[0],
+                                                cpu_fetchs1[0]);
+    } else {
+      TestMultiThreadInference<paddle::platform::CPUPlace>(
+          dirname, cpu_feeds, cpu_fetchs1, num_threads);
+    }
 
 #ifdef PADDLE_WITH_CUDA
-  paddle::framework::LoDTensor output2;
-  std::vector<paddle::framework::LoDTensor*> cpu_fetchs2;
-  cpu_fetchs2.push_back(&output2);
+    std::vector<std::vector<paddle::framework::LoDTensor*>> cpu_fetchs2;
+    cpu_fetchs2.resize(num_threads);
+    for (int i = 0; i < num_threads; ++i) {
+      auto* output = new paddle::framework::LoDTensor();
+      cpu_fetchs2[i].push_back(output);
+    }
 
     // Run inference on CUDA GPU
-  TestInference<paddle::platform::CUDAPlace>(dirname, cpu_feeds, cpu_fetchs2);
-  LOG(INFO) << output2.dims();
+    LOG(INFO) << "--- GPU Runs (num_threads: " << num_threads << "): ---";
+    if (num_threads == 1) {
+      TestInference<paddle::platform::CUDAPlace>(dirname, cpu_feeds[0],
+                                                 cpu_fetchs2[0]);
+    } else {
+      TestMultiThreadInference<paddle::platform::CUDAPlace>(
+          dirname, cpu_feeds, cpu_fetchs2, num_threads);
+    }
 
-  CheckError<float>(output1, output2);
+    for (int i = 0; i < num_threads; ++i) {
+      CheckError<float>(*cpu_fetchs1[i][0], *cpu_fetchs2[i][0]);
+      delete cpu_fetchs2[i][0];
+    }
 #endif
+
+    for (int i = 0; i < num_threads; ++i) {
+      delete cpu_feeds[i][0];
+      delete cpu_fetchs1[i][0];
+    }
+  }  // num_threads-loop
 }

paddle/fluid/inference/tests/test_helper.h

@@ -25,7 +25,8 @@ limitations under the License. */
 template <typename T>
 void SetupTensor(paddle::framework::LoDTensor* input,
                  paddle::framework::DDim dims, T lower, T upper) {
-  std::mt19937 rng(100);  // An arbitrarily chosen but fixed seed.
+  static unsigned int seed = 100;
+  std::mt19937 rng(seed++);
   std::uniform_real_distribution<double> uniform_dist(0, 1);
 
   T* input_ptr = input->mutable_data<T>(dims, paddle::platform::CPUPlace());

paddle/fluid/inference/tests/test_multi_thread_helper.h

+/* Copyright (c) 2018 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. */
+
+#pragma once
+
+#include <map>
+#include <string>
+#include <thread>  // NOLINT
+#include <vector>
+#include "paddle/fluid/framework/lod_tensor.h"
+#include "paddle/fluid/inference/io.h"
+
+void ThreadedRunInference(
+    const std::unique_ptr<paddle::framework::ProgramDesc>& inference_program,
+    paddle::framework::Executor* executor, paddle::framework::Scope* scope,
+    const int thread_id,
+    const std::vector<paddle::framework::LoDTensor*>& cpu_feeds,
+    const std::vector<paddle::framework::LoDTensor*>& cpu_fetchs) {
+  auto copy_program = std::unique_ptr<paddle::framework::ProgramDesc>(
+      new paddle::framework::ProgramDesc(*inference_program));
+
+  std::string feed_holder_name = "feed_" + paddle::string::to_string(thread_id);
+  std::string fetch_holder_name =
+      "fetch_" + paddle::string::to_string(thread_id);
+  copy_program->SetFeedHolderName(feed_holder_name);
+  copy_program->SetFetchHolderName(fetch_holder_name);
+
+  // 3. Get the feed_target_names and fetch_target_names
+  const std::vector<std::string>& feed_target_names =
+      copy_program->GetFeedTargetNames();
+  const std::vector<std::string>& fetch_target_names =
+      copy_program->GetFetchTargetNames();
+
+  // 4. Prepare inputs: set up maps for feed targets
+  std::map<std::string, const paddle::framework::LoDTensor*> feed_targets;
+  for (size_t i = 0; i < feed_target_names.size(); ++i) {
+    // Please make sure that cpu_feeds[i] is right for feed_target_names[i]
+    feed_targets[feed_target_names[i]] = cpu_feeds[i];
+  }
+
+  // 5. Define Tensor to get the outputs: set up maps for fetch targets
+  std::map<std::string, paddle::framework::LoDTensor*> fetch_targets;
+  for (size_t i = 0; i < fetch_target_names.size(); ++i) {
+    fetch_targets[fetch_target_names[i]] = cpu_fetchs[i];
+  }
+
+  // 6. Run the inference program
+  executor->Run(*copy_program, scope, feed_targets, fetch_targets, true,
+                feed_holder_name, fetch_holder_name);
+}
+
+template <typename Place>
+void TestMultiThreadInference(
+    const std::string& dirname,
+    const std::vector<std::vector<paddle::framework::LoDTensor*>>& cpu_feeds,
+    const std::vector<std::vector<paddle::framework::LoDTensor*>>& cpu_fetchs,
+    const int num_threads) {
+  // 1. Define place, executor, scope
+  auto place = Place();
+  auto executor = paddle::framework::Executor(place);
+  auto* scope = new paddle::framework::Scope();
+
+  // 2. Initialize the inference_program and load parameters
+  std::unique_ptr<paddle::framework::ProgramDesc> inference_program =
+      paddle::inference::Load(executor, *scope, dirname);
+
+  std::vector<std::thread*> threads;
+  for (int i = 0; i < num_threads; ++i) {
+    threads.push_back(new std::thread(
+        ThreadedRunInference, std::ref(inference_program), &executor, scope, i,
+        std::ref(cpu_feeds[i]), std::ref(cpu_fetchs[i])));
+  }
+  for (int i = 0; i < num_threads; ++i) {
+    threads[i]->join();
+    delete threads[i];
+  }
+
+  delete scope;
+}

paddle/fluid/operators/activation_op.cc

@@ -662,14 +662,3 @@ REGISTER_OP(swish, ops::ActivationOp, ops::SwishOpMaker, swish_grad,
                                 ops::grad_functor<double>>);
 
 FOR_EACH_KERNEL_FUNCTOR(REGISTER_ACTIVATION_CPU_KERNEL);
-
-REGISTER_OP_CPU_KERNEL(relu,
-                       ops::ActivationKernel<paddle::platform::CPUDeviceContext,
-                                             ops::ReluFunctor<float>>,
-                       ops::ActivationKernel<paddle::platform::CPUDeviceContext,
-                                             ops::ReluFunctor<double>>);
-REGISTER_OP_CPU_KERNEL(
-    relu_grad, ops::ActivationGradKernel<paddle::platform::CPUDeviceContext,
-                                         ops::ReluGradFunctor<float>>,
-    ops::ActivationGradKernel<paddle::platform::CPUDeviceContext,
-                              ops::ReluGradFunctor<double>>);

paddle/fluid/operators/activation_op.cu

 /* 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.
@@ -17,31 +14,19 @@ limitations under the License. */
 #include "paddle/fluid/platform/float16.h"
 
 namespace ops = paddle::operators;
+namespace plat = paddle::platform;
 
 #define REGISTER_ACTIVATION_CUDA_KERNEL(act_type, functor, grad_functor)    \
   REGISTER_OP_CUDA_KERNEL(                                                  \
-      act_type, ops::ActivationKernel<paddle::platform::CUDADeviceContext, \
-                                      ops::functor<float>>,                \
-      ops::ActivationKernel<paddle::platform::CUDADeviceContext,           \
-                            ops::functor<double>>);                        \
+      act_type,                                                             \
+      ops::ActivationKernel<plat::CUDADeviceContext, ops::functor<float>>,  \
+      ops::ActivationKernel<plat::CUDADeviceContext, ops::functor<double>>, \
+      ops::ActivationKernel<plat::CUDADeviceContext,                        \
+                            ops::functor<plat::float16>>);                  \
   REGISTER_OP_CUDA_KERNEL(                                                  \
-      act_type##_grad,                                                     \
-      ops::ActivationGradKernel<paddle::platform::CUDADeviceContext,       \
+      act_type##_grad, ops::ActivationGradKernel<plat::CUDADeviceContext,   \
                                                  ops::grad_functor<float>>, \
-      ops::ActivationGradKernel<paddle::platform::CUDADeviceContext,       \
+      ops::ActivationGradKernel<plat::CUDADeviceContext,                    \
                                 ops::grad_functor<double>>);
 
 FOR_EACH_KERNEL_FUNCTOR(REGISTER_ACTIVATION_CUDA_KERNEL);
-
-REGISTER_OP_CUDA_KERNEL(
-    relu, ops::ActivationKernel<paddle::platform::CUDADeviceContext,
-                                ops::ReluFunctor<float>>,
-    ops::ActivationKernel<paddle::platform::CUDADeviceContext,
-                          ops::ReluFunctor<double>>,
-    ops::ActivationKernel<paddle::platform::CUDADeviceContext,
-                          ops::ReluFunctor<paddle::platform::float16>>);
-REGISTER_OP_CUDA_KERNEL(
-    relu_grad, ops::ActivationGradKernel<paddle::platform::CUDADeviceContext,
-                                         ops::ReluGradFunctor<float>>,
-    ops::ActivationGradKernel<paddle::platform::CUDADeviceContext,
-                              ops::ReluGradFunctor<double>>);

paddle/fluid/operators/activation_op.h

 /* Copyright (c) 2018 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.
@@ -13,9 +10,13 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
+#include <utility>
+#include <vector>
+
 #include "paddle/fluid/framework/eigen.h"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/operators/detail/safe_ref.h"
+#include "paddle/fluid/platform/float16.h"
 
 #ifdef PADDLE_WITH_MKLDNN
 #include "paddle/fluid/platform/mkldnn_helper.h"
@@ -336,11 +337,25 @@ struct Sine {
   HOSTDEVICE T operator()(const T& val) const { return sin(val); }
 };
 
+template <>
+struct Sine<platform::float16> {
+  HOSTDEVICE platform::float16 operator()(const platform::float16& val) const {
+    return platform::float16(sin(static_cast<float>(val)));
+  }
+};
+
 template <typename T>
 struct Cosine {
   HOSTDEVICE T operator()(const T& val) const { return cos(val); }
 };
 
+template <>
+struct Cosine<platform::float16> {
+  HOSTDEVICE platform::float16 operator()(const platform::float16& val) const {
+    return platform::float16(cos(static_cast<float>(val)));
+  }
+};
+
 // cosine'(x) = -sin(x)
 template <typename T>
 struct CosGradFunctor : public BaseActivationFunctor<T> {
@@ -824,6 +839,7 @@ struct SwishGradFunctor : public BaseActivationFunctor<T> {
   __macro(sigmoid, SigmoidFunctor, SigmoidGradFunctor);              \
   __macro(logsigmoid, LogSigmoidFunctor, LogSigmoidGradFunctor);     \
   __macro(exp, ExpFunctor, ExpGradFunctor);                          \
+  __macro(relu, ReluFunctor, ReluGradFunctor);                       \
   __macro(tanh, TanhFunctor, TanhGradFunctor);                       \
   __macro(softshrink, SoftShrinkFunctor, SoftShrinkGradFunctor);     \
   __macro(sqrt, SqrtFunctor, SqrtGradFunctor);                       \

paddle/fluid/operators/adagrad_op.cc

@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/operators/adagrad_op.h"
+#include <vector>
 
 #include <cmath>
 

paddle/fluid/operators/array_operator.h

@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
+#include <string>
 #include "paddle/fluid/framework/lod_tensor_array.h"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/platform/device_context.h"

paddle/fluid/operators/assign_value_op.cc

@@ -13,6 +13,8 @@
 // limitations under the License.
 
 #include "paddle/fluid/operators/assign_value_op.h"
+#include <string>
+#include <vector>
 
 namespace paddle {
 namespace operators {

paddle/fluid/operators/assign_value_op.h

@@ -14,6 +14,7 @@
 
 #pragma once
 
+#include <vector>
 #include "paddle/fluid/framework/eigen.h"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/platform/enforce.h"

paddle/fluid/operators/auc_op.cc

@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/operators/auc_op.h"
+#include <string>
 
 namespace paddle {
 namespace operators {

paddle/fluid/operators/auc_op.h

@@ -13,6 +13,8 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
+#include <string>
+#include <vector>
 #include "paddle/fluid/framework/eigen.h"
 #include "paddle/fluid/framework/op_registry.h"
 
@@ -40,7 +42,7 @@ class AucKernel : public framework::OpKernel<T> {
     std::vector<float> thresholds_list;
     thresholds_list.reserve(num_thresholds);
     for (int i = 1; i < num_thresholds - 1; i++) {
-      thresholds_list[i] = (float)i / (num_thresholds - 1);
+      thresholds_list[i] = static_cast<float>(i) / (num_thresholds - 1);
     }
     const float kEpsilon = 1e-7;
     thresholds_list[0] = 0.0f - kEpsilon;
@@ -105,11 +107,12 @@ class AucKernel : public framework::OpKernel<T> {
     float* fp_rate_data = fp_rate.mutable_data<float>(ctx.GetPlace());
     float* rec_rate_data = rec_rate.mutable_data<float>(ctx.GetPlace());
     for (int i = 0; i < num_thresholds; i++) {
-      tp_rate_data[i] =
-          ((float)tp_data[i] + epsilon) / (tp_data[i] + fn_data[i] + epsilon);
-      fp_rate_data[i] = (float)fp_data[i] / (fp_data[i] + tn_data[i] + epsilon);
-      rec_rate_data[i] =
-          ((float)tp_data[i] + epsilon) / (tp_data[i] + fp_data[i] + epsilon);
+      tp_rate_data[i] = (static_cast<float>(tp_data[i]) + epsilon) /
+                        (tp_data[i] + fn_data[i] + epsilon);
+      fp_rate_data[i] =
+          static_cast<float>(fp_data[i]) / (fp_data[i] + tn_data[i] + epsilon);
+      rec_rate_data[i] = (static_cast<float>(tp_data[i]) + epsilon) /
+                         (tp_data[i] + fp_data[i] + epsilon);
     }
     *auc_data = 0.0f;
     if (curve == "ROC") {

paddle/fluid/operators/average_accumulates_op.cc

@@ -19,15 +19,15 @@ namespace operators {
 
 template <>
 void GetAccumulators<paddle::platform::CPUDeviceContext>(
-    const framework::ExecutionContext& ctx, int64_t& num_updates_,
-    int64_t& num_accumulates_, int64_t& old_num_accumulates_) {
+    const framework::ExecutionContext& ctx, int64_t* num_updates_,
+    int64_t* num_accumulates_, int64_t* old_num_accumulates_) {
   auto* in_old_num_accumulates = ctx.Input<Tensor>("in_old_num_accumulates");
   auto* in_num_accumulates = ctx.Input<Tensor>("in_num_accumulates");
   auto* in_num_updates = ctx.Input<Tensor>("in_num_updates");
 
-  old_num_accumulates_ = in_old_num_accumulates->data<int64_t>()[0];
-  num_accumulates_ = in_num_accumulates->data<int64_t>()[0];
-  num_updates_ = in_num_updates->data<int64_t>()[0];
+  *old_num_accumulates_ = in_old_num_accumulates->data<int64_t>()[0];
+  *num_accumulates_ = in_num_accumulates->data<int64_t>()[0];
+  *num_updates_ = in_num_updates->data<int64_t>()[0];
 }
 
 template <>

paddle/fluid/operators/average_accumulates_op.cu

@@ -19,18 +19,18 @@ namespace paddle {
 namespace operators {
 template <>
 void GetAccumulators<paddle::platform::CUDADeviceContext>(
-    const framework::ExecutionContext& ctx, int64_t& num_updates_,
-    int64_t& num_accumulates_, int64_t& old_num_accumulates_) {
+    const framework::ExecutionContext& ctx, int64_t* num_updates_,
+    int64_t* num_accumulates_, int64_t* old_num_accumulates_) {
   auto* in_old_num_accumulates = ctx.Input<Tensor>("in_old_num_accumulates");
   auto* in_num_accumulates = ctx.Input<Tensor>("in_num_accumulates");
   auto* in_num_updates = ctx.Input<Tensor>("in_num_updates");
   auto stream = ctx.cuda_device_context().stream();
-  memory::Copy(platform::CPUPlace(), &old_num_accumulates_,
+  memory::Copy(platform::CPUPlace(), old_num_accumulates_,
                platform::CUDAPlace(), in_old_num_accumulates->data<int64_t>(),
                sizeof(int64_t), stream);
-  memory::Copy(platform::CPUPlace(), &num_accumulates_, platform::CUDAPlace(),
+  memory::Copy(platform::CPUPlace(), num_accumulates_, platform::CUDAPlace(),
                in_num_accumulates->data<int64_t>(), sizeof(int64_t), stream);
-  memory::Copy(platform::CPUPlace(), &num_updates_, platform::CUDAPlace(),
+  memory::Copy(platform::CPUPlace(), num_updates_, platform::CUDAPlace(),
                in_num_updates->data<int64_t>(), sizeof(int64_t), stream);
 }
 

paddle/fluid/operators/average_accumulates_op.h

@@ -29,8 +29,8 @@ using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
 
 template <typename DeviceContext>
 void GetAccumulators(const framework::ExecutionContext& ctx,
-                     int64_t& num_updates, int64_t& num_accumulates,
-                     int64_t& old_num_accumulates);
+                     int64_t* num_updates, int64_t* num_accumulates,
+                     int64_t* old_num_accumulates);
 
 template <typename DeviceContext>
 void SetAccumulators(const framework::ExecutionContext& ctx,
@@ -47,8 +47,8 @@ class AverageAccumulatesKernel : public framework::OpKernel<T> {
     int64_t num_updates = 0;
     int64_t num_accumulates = 0;
     int64_t old_num_accumulates = 0;
-    GetAccumulators<DeviceContext>(ctx, num_updates, num_accumulates,
-                                   old_num_accumulates);
+    GetAccumulators<DeviceContext>(ctx, &num_updates, &num_accumulates,
+                                   &old_num_accumulates);
 
     // Get attrs
     float average_window = ctx.Attr<float>("average_window");

paddle/fluid/operators/spp_op.cc

@@ -13,6 +13,8 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/operators/spp_op.h"
+#include <string>
+#include <vector>
 namespace paddle {
 namespace operators {
 

paddle/fluid/operators/spp_op.h

@@ -13,6 +13,8 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
+#include <string>
+#include <vector>
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/operators/math/math_function.h"
 #include "paddle/fluid/operators/math/pooling.h"

paddle/fluid/operators/sum_op.cc

@@ -10,6 +10,8 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/operators/sum_op.h"
+#include <algorithm>
+#include <string>
 #include <vector>
 #include "paddle/fluid/framework/var_type_inference.h"
 #include "paddle/fluid/operators/detail/safe_ref.h"

paddle/fluid/operators/sum_op.h

@@ -10,6 +10,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
+#include <vector>
 #include "paddle/fluid/framework/eigen.h"
 #include "paddle/fluid/framework/lod_tensor_array.h"
 #include "paddle/fluid/framework/op_registry.h"

paddle/fluid/operators/top_k_op.h

@@ -15,6 +15,8 @@ limitations under the License. */
 #pragma once
 #include <algorithm>
 #include <iostream>
+#include <utility>
+#include <vector>
 #include "paddle/fluid/framework/eigen.h"
 #include "paddle/fluid/framework/op_registry.h"
 

paddle/fluid/operators/transpose_op.cc

@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/operators/transpose_op.h"
+#include <vector>
 
 namespace paddle {
 namespace operators {

paddle/fluid/operators/transpose_op.h

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #pragma once
 
+#include <vector>
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/operators/math/math_function.h"
 

paddle/fluid/operators/unpool_op.cc

@@ -13,6 +13,8 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/operators/unpool_op.h"
+#include <string>
+#include <vector>
 namespace paddle {
 namespace operators {
 

paddle/fluid/operators/unpool_op.h

@@ -14,6 +14,8 @@ limitations under the License. */
 
 #pragma once
 
+#include <string>
+#include <vector>
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/operators/math/math_function.h"
 #include "paddle/fluid/operators/math/unpooling.h"

paddle/fluid/operators/warpctc_op.h

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #pragma once
 
+#include <vector>
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/operators/math/math_function.h"
 #include "paddle/fluid/operators/math/sequence_padding.h"

paddle/fluid/platform/call_once.h

-/* 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. */
-
-#pragma once
-
-#include <mutex>
-
-namespace paddle {
-namespace platform {
-
-/*
- The current implementation of std::call_once has a bug described in
- https://stackoverflow.com/questions/41717579/stdcall-once-hangs-on-second-call-after-callable-threw-on-first-call.
- This is likely caused by a deeper bug of pthread_once, which is discussed in
- https://patchwork.ozlabs.org/patch/482350/
-
- This wrap is a hack to avoid this bug.
-*/
-template <typename Callable, typename... Args>
-inline void call_once(std::once_flag& flag, Callable&& f, Args&&... args) {
-  bool good = true;
-  std::exception ex;
-  try {
-    std::call_once(flag,
-                   [&](Args&&... args) {
-                     try {
-                       f(args...);
-                     } catch (const std::exception& e) {
-                       ex = e;
-                       good = false;
-                     } catch (...) {
-                       ex = std::runtime_error("excption caught in call_once");
-                       good = false;
-                     }
-                   },
-                   args...);
-  } catch (std::system_error& x) {
-    throw std::runtime_error("call once failed");
-  }
-  if (!good) {
-    throw std::exception(ex);
-  }
-}
-
-}  // namespace platform
-}  // namespace paddle

paddle/fluid/platform/device_context.cc

@@ -8,10 +8,14 @@ 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 "paddle/fluid/platform/device_context.h"
+
+#include <string>
 #include <unordered_set>
+#include <vector>
+
 #include "paddle/fluid/memory/memory.h"
+
 namespace paddle {
 namespace platform {
 

paddle/fluid/platform/device_context.h

@@ -8,11 +8,12 @@ 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. */
-
 #pragma once
 
 #include <memory>
+#include <string>
 #include <unordered_map>
+#include <vector>
 
 #ifdef PADDLE_WITH_CUDA
 #include "paddle/fluid/platform/dynload/cublas.h"

paddle/fluid/platform/device_context_test.cu

@@ -11,11 +11,12 @@ 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 "paddle/fluid/platform/device_context.h"
 
+#include <vector>
+
 #include "glog/logging.h"
+#include "gtest/gtest.h"
 
 TEST(Device, Init) {
   using paddle::platform::DeviceContext;

paddle/fluid/platform/device_tracer.cc

@@ -11,15 +11,19 @@ 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 "paddle/fluid/platform/device_tracer.h"
-#include <google/protobuf/text_format.h>
+
+#include <deque>
 #include <fstream>
 #include <map>
-#include <mutex>
+#include <mutex>  // NOLINT
 #include <numeric>
-#include <thread>
+#include <string>
+#include <thread>  // NOLINT
+#include <vector>
+
 #include "glog/logging.h"
+#include "google/protobuf/text_format.h"
 #include "paddle/fluid/framework/block_desc.h"
 #include "paddle/fluid/string/printf.h"
 
@@ -123,7 +127,7 @@ void DisableActivity() {
 
 void CUPTIAPI bufferRequested(uint8_t **buffer, size_t *size,
                               size_t *maxNumRecords) {
-  uint8_t *buf = (uint8_t *)malloc(kBufSize + kAlignSize);
+  uint8_t *buf = reinterpret_cast<uint8_t *>(malloc(kBufSize + kAlignSize));
   *size = kBufSize;
   *buffer = ALIGN_BUFFER(buf, kAlignSize);
   *maxNumRecords = 0;

paddle/fluid/platform/device_tracer.h

@@ -11,8 +11,10 @@ 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. */
-
 #pragma once
+
+#include <string>
+
 #include "paddle/fluid/platform/dynload/cupti.h"
 #include "paddle/fluid/platform/profiler.pb.h"
 

paddle/fluid/platform/dynload/nccl.h

@@ -18,7 +18,6 @@ limitations under the License. */
 
 #include <mutex>  // NOLINT
 
-#include "paddle/fluid/platform/call_once.h"
 #include "paddle/fluid/platform/dynload/dynamic_loader.h"
 
 namespace paddle {

paddle/fluid/platform/float16.h

@@ -1003,6 +1003,46 @@ HOSTDEVICE inline float16 exp(const float16& a) {
   return float16(::expf(static_cast<float>(a)));
 }
 
+template <>
+HOSTDEVICE inline float16 log(const float16& a) {
+  return float16(::logf(static_cast<float>(a)));
+}
+
+template <>
+HOSTDEVICE inline float16 tanh(const float16& a) {
+  return float16(::tanhf(static_cast<float>(a)));
+}
+
+template <>
+HOSTDEVICE inline float16 sqrt(const float16& a) {
+  return float16(::sqrtf(static_cast<float>(a)));
+}
+
+template <>
+HOSTDEVICE inline float16 ceil(const float16& a) {
+  return float16(::ceilf(static_cast<float>(a)));
+}
+
+template <>
+HOSTDEVICE inline float16 floor(const float16& a) {
+  return float16(::floorf(static_cast<float>(a)));
+}
+
+template <>
+HOSTDEVICE inline float16 round(const float16& a) {
+  return float16(::roundf(static_cast<float>(a)));
+}
+
+template <>
+HOSTDEVICE inline float16 pow(const float16& a, const float16& b) {
+  return float16(::powf(static_cast<float>(a), static_cast<float>(b)));
+}
+
+template <>
+HOSTDEVICE inline float16 abs(const float16& a) {
+  return float16(::fabs(static_cast<float>(a)));
+}
+
 }  // namespace numext
 
 }  // namespace Eigen

paddle/fluid/platform/mkldnn_helper.h

@@ -11,11 +11,11 @@ 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. */
-
 #pragma once
 
-#include <mkldnn.hpp>
+#include <vector>
 
+#include "mkldnn/include/mkldnn.hpp"
 #include "paddle/fluid/framework/operator.h"
 
 namespace paddle {

paddle/fluid/platform/profiler.cc

@@ -15,8 +15,11 @@ limitations under the License. */
 #include "paddle/fluid/platform/profiler.h"
 #include <sys/time.h>
 #include <time.h>
+#include <algorithm>
 #include <iomanip>
 #include <map>
+#include <mutex>  // NOLINT
+#include <string>
 #ifdef PADDLE_WITH_CUDA
 #include <cuda.h>
 #endif  // PADDLE_WITH_CUDA
@@ -28,10 +31,10 @@ limitations under the License. */
 namespace paddle {
 namespace platform {
 
+struct EventList;
+
 // The profiler state, the initial value is ProfilerState::kDisabled
 static ProfilerState g_state = ProfilerState::kDisabled;
-// To record which timer the profiler used, CUDA or CPU.
-static std::string g_profiler_place = "";
 // The thread local event list only can be accessed by the specific thread
 // The thread index of each thread
 static thread_local int32_t g_thread_id;
@@ -45,6 +48,39 @@ static std::list<std::shared_ptr<EventList>> g_all_event_lists;
 // The thread local event list only can be accessed by the specific thread
 static thread_local std::shared_ptr<EventList> g_event_list;
 
+struct EventList {
+  constexpr static size_t kMB = 1024 * 1024;
+  constexpr static size_t kEventBlockSize = 16 * kMB;
+  constexpr static size_t kEventSize = sizeof(Event);
+  constexpr static size_t kEventAlign = alignof(Event);
+  constexpr static size_t kNumBlock =
+      kEventBlockSize /
+      ((kEventSize + kEventAlign - 1) / kEventAlign * kEventAlign);
+
+  template <typename... Args>
+  void Record(Args&&... args) {
+    if (event_blocks.empty() || event_blocks.front().size() == kNumBlock) {
+      event_blocks.emplace_front();
+      event_blocks.front().reserve(kNumBlock);
+    }
+    event_blocks.front().emplace_back(std::forward<Args>(args)...);
+  }
+
+  std::vector<Event> Reduce() {
+    std::vector<Event> result;
+    for (auto& block : event_blocks) {
+      result.insert(result.begin(), std::make_move_iterator(block.begin()),
+                    std::make_move_iterator(block.end()));
+    }
+    event_blocks.clear();
+    return result;
+  }
+
+  void Clear() { event_blocks.clear(); }
+
+  std::forward_list<std::vector<Event>> event_blocks;
+};
+
 inline uint64_t GetTimeInNsec() {
   using clock = std::conditional<std::chrono::high_resolution_clock::is_steady,
                                  std::chrono::high_resolution_clock,
@@ -60,9 +96,9 @@ inline uint64_t PosixInNsec() {
   return 1000 * (static_cast<uint64_t>(tv.tv_sec) * 1000000 + tv.tv_usec);
 }
 
-Event::Event(EventKind kind, std::string name, uint32_t thread_id,
+Event::Event(EventType type, std::string name, uint32_t thread_id,
              const DeviceContext* dev_ctx)
-    : kind_(kind), name_(name), thread_id_(thread_id), has_cuda_(false) {
+    : type_(type), name_(name), thread_id_(thread_id), has_cuda_(false) {
 #ifdef PADDLE_WITH_CUDA
   has_cuda_ = dev_ctx ? platform::is_gpu_place(dev_ctx->GetPlace()) : false;
   if (has_cuda_) {
@@ -76,17 +112,7 @@ Event::Event(EventKind kind, std::string name, uint32_t thread_id,
   cpu_ns_ = GetTimeInNsec();
 }
 
-std::string Event::kind() const {
-  switch (kind_) {
-    case EventKind::kMark:
-      return "mark";
-    case EventKind::kPushRange:
-      return "push";
-    case EventKind::kPopRange:
-      return "pop";
-  }
-  PADDLE_THROW("Unknown EventKind.");
-}
+const EventType& Event::type() const { return type_; }
 
 double Event::CpuElapsedMs(const Event& e) const {
   return (e.cpu_ns_ - cpu_ns_) / (1000000.0);
@@ -129,15 +155,15 @@ inline EventList& GetEventList() {
 }
 
 void Mark(const std::string& name, const DeviceContext* dev_ctx) {
-  GetEventList().Record(EventKind::kMark, name, g_thread_id, dev_ctx);
+  GetEventList().Record(EventType::kMark, name, g_thread_id, dev_ctx);
 }
 
 void PushEvent(const std::string& name, const DeviceContext* dev_ctx) {
-  GetEventList().Record(EventKind::kPushRange, name, g_thread_id, dev_ctx);
+  GetEventList().Record(EventType::kPushRange, name, g_thread_id, dev_ctx);
 }
 
 void PopEvent(const std::string& name, const DeviceContext* dev_ctx) {
-  GetEventList().Record(EventKind::kPopRange, name, g_thread_id, dev_ctx);
+  GetEventList().Record(EventType::kPopRange, name, g_thread_id, dev_ctx);
 }
 
 RecordEvent::RecordEvent(const std::string& name, const DeviceContext* dev_ctx)
@@ -197,12 +223,7 @@ void EnableProfiler(ProfilerState state) {
                  "The profiling state should be disabled when calling ",
                  "EnableProfiler.");
   g_state = state;
-  if (g_state == ProfilerState::kCUDA) {
-    g_profiler_place = "CUDA";
-  } else if (g_state == ProfilerState::kCPU) {
-    g_profiler_place = "CPU";
-  } else {
-    g_profiler_place = "All";
+  if (g_state == ProfilerState::kAll) {
     GetDeviceTracer()->Enable();
   }
 #ifdef PADDLE_WITH_CUDA
@@ -240,27 +261,63 @@ std::vector<std::vector<Event>> GetAllEvents() {
   return result;
 }
 
-void DisableProfiler(EventSortingKey sorted_key,
-                     const std::string& profile_path) {
-  PADDLE_ENFORCE(g_state != ProfilerState::kDisabled,
-                 "Can't disable profiling, since it's not starting.");
-  // Mark the profiling stop.
-  Mark("_stop_profiler_", nullptr);
-  g_state = ProfilerState::kDisabled;
+// The information of each event given in the profiling report
+struct EventItem {
+  std::string name;
+  int calls;
+  double total_time;
+  double min_time;
+  double max_time;
+  double ave_time;
+};
+
+// Print results
+void PrintProfiler(const std::vector<std::vector<EventItem>>& events_table,
+                   const std::string& sorted_domain, const size_t name_width,
+                   const size_t data_width) {
+  // Output header information
+  std::cout << "\n------------------------->"
+            << "     Profiling Report     "
+            << "<-------------------------\n\n";
+  std::string place;
+  if (g_state == ProfilerState::kCPU) {
+    place = "CPU";
+  } else if (g_state == ProfilerState::kCUDA) {
+    place = "CUDA";
+  } else if (g_state == ProfilerState::kAll) {
+    place = "All";
+  } else {
+    PADDLE_THROW("Invalid profiler state");
+  }
 
-  std::vector<std::vector<Event>> all_events = GetAllEvents();
-  ParseEvents(all_events, sorted_key);
-  ResetProfiler();
-  DeviceTracer* tracer = GetDeviceTracer();
-  if (g_profiler_place == "All" && tracer && tracer->IsEnabled()) {
-    tracer->Disable();
-    tracer->GenProfile(profile_path);
+  std::cout << "Place: " << place << std::endl;
+  std::cout << "Time unit: ms" << std::endl;
+  std::cout << "Sorted by " << sorted_domain
+            << " in descending order in the same thread\n\n";
+  // Output events table
+  std::cout.setf(std::ios::left);
+  std::cout << std::setw(name_width) << "Event" << std::setw(data_width)
+            << "Calls" << std::setw(data_width) << "Total"
+            << std::setw(data_width) << "Min." << std::setw(data_width)
+            << "Max." << std::setw(data_width) << "Ave." << std::endl;
+  for (size_t i = 0; i < events_table.size(); ++i) {
+    for (size_t j = 0; j < events_table[i].size(); ++j) {
+      const EventItem& event_item = events_table[i][j];
+      std::cout << std::setw(name_width) << event_item.name
+                << std::setw(data_width) << event_item.calls
+                << std::setw(data_width) << event_item.total_time
+                << std::setw(data_width) << event_item.min_time
+                << std::setw(data_width) << event_item.max_time
+                << std::setw(data_width) << event_item.ave_time << std::endl;
+    }
   }
+  std::cout << std::endl;
 }
 
-void ParseEvents(std::vector<std::vector<Event>>& events,
-                 EventSortingKey sorted_by) {
-  if (g_profiler_place == "") return;
+// Parse the event list and output the profiling report
+void ParseEvents(const std::vector<std::vector<Event>>& events,
+                 EventSortingKey sorted_by = EventSortingKey::kDefault) {
+  if (g_state == ProfilerState::kDisabled) return;
 
   std::string sorted_domain;
   std::function<bool(const EventItem&, const EventItem&)> sorted_func;
@@ -307,9 +364,9 @@ void ParseEvents(std::vector<std::vector<Event>>& events,
     std::unordered_map<std::string, int> event_idx;
 
     for (size_t j = 0; j < events[i].size(); j++) {
-      if (events[i][j].kind() == "push") {
+      if (events[i][j].type() == EventType::kPushRange) {
         pushed_events.push_back(events[i][j]);
-      } else if (events[i][j].kind() == "pop") {
+      } else if (events[i][j].type() == EventType::kPopRange) {
         std::list<Event>::reverse_iterator rit = pushed_events.rbegin();
         while (rit != pushed_events.rend() &&
                rit->name() != events[i][j].name()) {
@@ -317,8 +374,8 @@ void ParseEvents(std::vector<std::vector<Event>>& events,
         }
 
         if (rit != pushed_events.rend()) {
-          double event_time =
-              (g_profiler_place == "CUDA" || g_profiler_place == "All")
+          double event_time = (g_state == ProfilerState::kCUDA ||
+                               g_state == ProfilerState::kAll)
                                   ? rit->CudaElapsedMs(events[i][j])
                                   : rit->CpuElapsedMs(events[i][j]);
 
@@ -376,35 +433,22 @@ void ParseEvents(std::vector<std::vector<Event>>& events,
   PrintProfiler(events_table, sorted_domain, max_name_width + 4, 12);
 }
 
-void PrintProfiler(std::vector<std::vector<EventItem>>& events_table,
-                   std::string& sorted_domain, const size_t name_width,
-                   const size_t data_width) {
-  // Output header information
-  std::cout << "\n------------------------->"
-            << "     Profiling Report     "
-            << "<-------------------------\n\n";
-  std::cout << "Place: " << g_profiler_place << std::endl;
-  std::cout << "Time unit: ms" << std::endl;
-  std::cout << "Sorted by " << sorted_domain
-            << " in descending order in the same thread\n\n";
-  // Output events table
-  std::cout.setf(std::ios::left);
-  std::cout << std::setw(name_width) << "Event" << std::setw(data_width)
-            << "Calls" << std::setw(data_width) << "Total"
-            << std::setw(data_width) << "Min." << std::setw(data_width)
-            << "Max." << std::setw(data_width) << "Ave." << std::endl;
-  for (size_t i = 0; i < events_table.size(); ++i) {
-    for (size_t j = 0; j < events_table[i].size(); ++j) {
-      EventItem& event_item = events_table[i][j];
-      std::cout << std::setw(name_width) << event_item.name
-                << std::setw(data_width) << event_item.calls
-                << std::setw(data_width) << event_item.total_time
-                << std::setw(data_width) << event_item.min_time
-                << std::setw(data_width) << event_item.max_time
-                << std::setw(data_width) << event_item.ave_time << std::endl;
-    }
+void DisableProfiler(EventSortingKey sorted_key,
+                     const std::string& profile_path) {
+  PADDLE_ENFORCE(g_state != ProfilerState::kDisabled,
+                 "Can't disable profiling, since it's not starting.");
+  // Mark the profiling stop.
+  Mark("_stop_profiler_", nullptr);
+
+  std::vector<std::vector<Event>> all_events = GetAllEvents();
+  ParseEvents(all_events, sorted_key);
+  ResetProfiler();
+  DeviceTracer* tracer = GetDeviceTracer();
+  if (g_state == ProfilerState::kAll && tracer && tracer->IsEnabled()) {
+    tracer->Disable();
+    tracer->GenProfile(profile_path);
   }
-  std::cout << std::endl;
+  g_state = ProfilerState::kDisabled;
 }
 
 }  // namespace platform

paddle/fluid/platform/profiler.h

@@ -15,7 +15,7 @@ limitations under the License. */
 #pragma once
 #include <forward_list>
 #include <list>
-#include <mutex>
+#include <string>
 #include <vector>
 #include "paddle/fluid/platform/device_context.h"
 #include "paddle/fluid/platform/profiler.pb.h"
@@ -23,16 +23,16 @@ limitations under the License. */
 namespace paddle {
 namespace platform {
 
-enum EventKind { kMark, kPushRange, kPopRange };
+enum EventType { kMark, kPushRange, kPopRange };
 
 class Event {
  public:
   // The DeviceContext is used to get the cuda stream.
   // If CPU profiling mode, can pass nullptr.
-  Event(EventKind kind, std::string name, uint32_t thread_id,
+  Event(EventType type, std::string name, uint32_t thread_id,
         const DeviceContext* dev_ctx);
 
-  std::string kind() const;
+  const EventType& type() const;
   std::string name() const { return name_; }
   uint32_t thread_id() const { return thread_id_; }
   bool has_cuda() const { return has_cuda_; }
@@ -46,7 +46,7 @@ class Event {
   double CudaElapsedMs(const Event& e) const;
 
  private:
-  EventKind kind_;
+  EventType type_;
   std::string name_;
   uint32_t thread_id_;
   int64_t cpu_ns_;
@@ -57,39 +57,6 @@ class Event {
 #endif
 };
 
-struct EventList {
-  constexpr static size_t kMB = 1024 * 1024;
-  constexpr static size_t kEventBlockSize = 16 * kMB;
-  constexpr static size_t kEventSize = sizeof(Event);
-  constexpr static size_t kEventAlign = alignof(Event);
-  constexpr static size_t kNumBlock =
-      kEventBlockSize /
-      ((kEventSize + kEventAlign - 1) / kEventAlign * kEventAlign);
-
-  template <typename... Args>
-  void Record(Args&&... args) {
-    if (event_blocks.empty() || event_blocks.front().size() == kNumBlock) {
-      event_blocks.emplace_front();
-      event_blocks.front().reserve(kNumBlock);
-    }
-    event_blocks.front().emplace_back(std::forward<Args>(args)...);
-  }
-
-  std::vector<Event> Reduce() {
-    std::vector<Event> result;
-    for (auto& block : event_blocks) {
-      result.insert(result.begin(), std::make_move_iterator(block.begin()),
-                    std::make_move_iterator(block.end()));
-    }
-    event_blocks.clear();
-    return result;
-  }
-
-  void Clear() { event_blocks.clear(); }
-
-  std::forward_list<std::vector<Event>> event_blocks;
-};
-
 enum ProfilerState {
   kDisabled,  // disabled state
   kCPU,       // CPU profiling state
@@ -136,16 +103,6 @@ struct RecordThread {
 // event_lists, event_lists[i][j] represents the j-th Event of i-th thread.
 std::vector<std::vector<Event>> GetAllEvents();
 
-// The information of each event given in the profiling report
-struct EventItem {
-  std::string name;
-  int calls;
-  double total_time;
-  double min_time;
-  double max_time;
-  double ave_time;
-};
-
 // Candidate keys to sort the profiling report
 enum EventSortingKey { kDefault, kCalls, kTotal, kMin, kMax, kAve };
 
@@ -158,14 +115,5 @@ void ResetProfiler();
 void DisableProfiler(EventSortingKey sorted_key,
                      const std::string& profile_path);
 
-// Parse the event list and output the profiling report
-void ParseEvents(std::vector<std::vector<Event>>&,
-                 EventSortingKey sorted_by = EventSortingKey::kDefault);
-
-// Print results
-void PrintProfiler(std::vector<std::vector<EventItem>>& events_table,
-                   std::string& sorted_domain, const size_t name_width,
-                   const size_t data_width);
-
 }  // namespace platform
 }  // namespace paddle

paddle/fluid/platform/profiler_test.cc

@@ -13,22 +13,23 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/platform/profiler.h"
+#include <string>
 #ifdef PADDLE_WITH_CUDA
-#include "cuda_runtime.h"
+#include <cuda_runtime.h>
 #endif
 #include "gtest/gtest.h"
 
 TEST(Event, CpuElapsedTime) {
   using paddle::platform::Event;
-  using paddle::platform::EventKind;
+  using paddle::platform::EventType;
 
-  Event start_event(EventKind::kPushRange, "test", 0, nullptr);
+  Event start_event(EventType::kPushRange, "test", 0, nullptr);
   EXPECT_TRUE(start_event.has_cuda() == false);
   int counter = 0;
   while (counter != 1000) {
     counter++;
   }
-  Event stop_event(EventKind::kPopRange, "test", 0, nullptr);
+  Event stop_event(EventType::kPopRange, "test", 0, nullptr);
   EXPECT_GT(start_event.CpuElapsedMs(stop_event), 0);
 }
 
@@ -38,16 +39,16 @@ TEST(Event, CudaElapsedTime) {
   using paddle::platform::CUDADeviceContext;
   using paddle::platform::CUDAPlace;
   using paddle::platform::Event;
-  using paddle::platform::EventKind;
+  using paddle::platform::EventType;
 
   DeviceContext* dev_ctx = new CUDADeviceContext(CUDAPlace(0));
-  Event start_event(EventKind::kPushRange, "test", 0, dev_ctx);
+  Event start_event(EventType::kPushRange, "test", 0, dev_ctx);
   EXPECT_TRUE(start_event.has_cuda() == true);
   int counter = 0;
   while (counter != 1000) {
     counter++;
   }
-  Event stop_event(EventKind::kPopRange, "test", 0, dev_ctx);
+  Event stop_event(EventType::kPopRange, "test", 0, dev_ctx);
   EXPECT_GT(start_event.CudaElapsedMs(stop_event), 0);
 }
 #endif
@@ -55,7 +56,7 @@ TEST(Event, CudaElapsedTime) {
 TEST(RecordEvent, RecordEvent) {
   using paddle::platform::DeviceContext;
   using paddle::platform::Event;
-  using paddle::platform::EventKind;
+  using paddle::platform::EventType;
   using paddle::platform::RecordEvent;
   using paddle::platform::ProfilerState;
   using paddle::platform::EventSortingKey;

paddle/fluid/pybind/.gitignore

+pybind.h

paddle/fluid/pybind/pybind.cc

@@ -465,7 +465,8 @@ All parameter, weight, gradient are variables in Paddle.
 
   m.def("init_gflags", framework::InitGflags);
   m.def("init_glog", framework::InitGLOG);
-  m.def("init_devices", &framework::InitDevices);
+  m.def("init_devices",
+        [](bool init_p2p) { framework::InitDevices(init_p2p); });
 
   m.def("is_compiled_with_cuda", IsCompiledWithCUDA);
 #ifdef PADDLE_WITH_CUDA

paddle/testing/paddle_gtest_main.cc

@@ -41,6 +41,6 @@ int main(int argc, char** argv) {
   paddle::memory::Used(paddle::platform::CUDAPlace(0));
 #endif
 
-  paddle::framework::InitDevices();
+  paddle::framework::InitDevices(true);
   return RUN_ALL_TESTS();
 }

python/.gitignore

 *pyc
 build
 dist
+paddlepaddle.egg-info
 paddle.egg-info
 paddlepaddle_gpu.egg-info
 .idea

python/paddle/.gitignore

+version.py

python/paddle/fluid/__init__.py

@@ -86,6 +86,8 @@ def __bootstrap__():
     import core
     import os
 
+    in_test = 'unittest' in sys.modules
+
     try:
         num_threads = int(os.getenv('OMP_NUM_THREADS', '1'))
     except ValueError:
@@ -110,8 +112,11 @@ def __bootstrap__():
     core.init_gflags([sys.argv[0]] +
                      ["--tryfromenv=" + ",".join(read_env_flags)])
     core.init_glog(sys.argv[0])
-    core.init_devices()
+    # don't init_p2p when in unittest to save time.
+    core.init_devices(not in_test)
 
 
+# TODO(panyx0718): Avoid doing complex initialization logic in __init__.py.
+# Consider paddle.init(args) or paddle.main(args)
 layers.monkey_patch_variable()
 __bootstrap__()

python/paddle/fluid/distribute_transpiler.py

@@ -102,6 +102,8 @@ def split_dense_variable(var_list,
         the parameter server side can gain better performance. By default
         minimum block size is 1024. The max block size is used to prevent
         very large blocks that may cause send error.
+        :return: A list of VarBlocks. Each VarBlock specifies a shard of
+           the var.
     """
     blocks = []
     for var in var_list:
@@ -192,22 +194,24 @@ class DistributeTranspiler:
         self.trainer_id = trainer_id
         pserver_endpoints = pservers.split(",")
 
-        # step1
+        # step1: For large parameters and gradients, split them into smaller
+        # blocks.
         param_list = [pg[0] for pg in params_grads]
         grad_list = [pg[1] for pg in params_grads]
         grad_blocks = split_dense_variable(grad_list, len(pserver_endpoints))
         param_blocks = split_dense_variable(param_list, len(pserver_endpoints))
-        # step2
+        # step2: Create new vars for the parameters and gradients blocks and
+        # add ops to do the split.
         grad_var_mapping = self._append_split_op(program, grad_blocks)
-        # step3
+        param_var_mapping = self._create_vars_from_blocklist(program,
+                                                             param_blocks)
+        # step3: Add gradients as send op inputs and parameters as send
+        # op outputs.
         send_inputs = []
         send_outputs = []
         for b in grad_blocks:  # append by order
             varname, block_id, _ = b.split(":")
             send_inputs.append(grad_var_mapping[varname][int(block_id)])
-
-        param_var_mapping = self._create_vars_from_blocklist(program,
-                                                             param_blocks)
         for b in param_blocks:
             varname, block_id, _ = b.split(":")
             send_outputs.append(param_var_mapping[varname][int(block_id)])
@@ -237,7 +241,7 @@ class DistributeTranspiler:
                      "RPCClient": rpc_client_var},
             attrs={"endpoints": pserver_endpoints,
                    "epmap": eplist})
-        # step4
+        # step4: Concat the parameters splits together after recv.
         for varname, splited_var in param_var_mapping.iteritems():
             if len(splited_var) <= 1:
                 continue
@@ -258,13 +262,14 @@ class DistributeTranspiler:
     def get_pserver_program(self, endpoint):
         """
         Get pserver side program using the endpoint.
+        TODO(panyx0718): Revisit this assumption. what if #blocks > #pservers.
         NOTE: assume blocks of the same variable is not distributed
         on the same pserver, only change param/grad varnames for
         trainers to fetch.
         """
         # step1
         pserver_program = Program()
-        # step2
+        # step2: Create vars to receive vars at parameter servers.
         recv_inputs = []
         for v in self.param_grad_ep_mapping[endpoint]["params"]:
             self._clone_var(pserver_program.global_block(), v)
@@ -278,12 +283,6 @@ class DistributeTranspiler:
                 orig_var_name = v.name[:suff_idx]
             else:
                 orig_var_name = v.name
-            single_trainer_var = pserver_program.global_block().create_var(
-                name=orig_var_name,
-                persistable=True,
-                type=v.type,
-                dtype=v.dtype,
-                shape=v.shape)
             if self.trainers > 1:
                 for trainer_id in xrange(self.trainers):
                     var = pserver_program.global_block().create_var(
@@ -294,6 +293,12 @@ class DistributeTranspiler:
                         shape=v.shape)
                     recv_inputs.append(var)
             else:
+                single_trainer_var = pserver_program.global_block().create_var(
+                    name=orig_var_name,
+                    persistable=True,
+                    type=v.type,
+                    dtype=v.dtype,
+                    shape=v.shape)
                 recv_inputs.append(single_trainer_var)
 
         # step3
@@ -344,7 +349,7 @@ class DistributeTranspiler:
                 self._append_pserver_non_opt_ops(block, op)
 
         append_block = optimize_block
-        # append lr decay ops to the child block if exits
+        # append lr decay ops to the child block if exists
         lr_ops = self._get_lr_ops()
         if len(lr_ops) > 0:
             for _, op in enumerate(lr_ops):
@@ -447,8 +452,10 @@ class DistributeTranspiler:
                                     block_list,
                                     add_trainer_suffix=False):
         """
+        Create vars for each split.
         NOTE: only grads need to be named for different trainers, use
               add_trainer_suffix to rename the grad vars.
+        :return: A dict mapping from original var name to each var split.
         """
         block_map = dict()
         var_mapping = dict()
@@ -615,6 +622,7 @@ class DistributeTranspiler:
                         type="sum",
                         inputs={"X": vars2merge},
                         outputs={"Out": merged_var})
+                    # TODO(panyx0718): What if it's SELECTED_ROWS.
                     if not merged_var.type == core.VarDesc.VarType.SELECTED_ROWS:
                         optimize_block.append_op(
                             type="scale",
@@ -638,7 +646,7 @@ class DistributeTranspiler:
                     shape=param_block.shape)
                 new_inputs[key] = tmpvar
             elif key == "LearningRate":
-                # leraning rate variable has already be created by non-optimize op,
+                # learning rate variable has already be created by non-optimize op,
                 # don't create it once again.
                 lr_varname = opt_op.input(key)[0]
                 if pserver_block.vars.has_key(lr_varname):
@@ -773,6 +781,7 @@ class DistributeTranspiler:
         return False
 
     def _get_input_map_from_op(self, varmap, op):
+        """Returns a dict from op input name to the vars in varmap."""
         iomap = dict()
         for key in op.input_names:
             vars = []
@@ -785,6 +794,7 @@ class DistributeTranspiler:
         return iomap
 
     def _get_output_map_from_op(self, varmap, op):
+        """Returns a dict from op output name to the vars in varmap."""
         iomap = dict()
         for key in op.output_names:
             vars = []
@@ -812,6 +822,7 @@ class DistributeTranspiler:
                 find_ops.append(op)
         # make a union find struct by the ops in default_main_program
         ufind = UnionFind(block.ops)
+
         for op1 in block.ops:
             for op2 in block.ops:
                 # NOTE: we need to skip all optimize ops, since it is connected

python/setup.py.in

@@ -102,7 +102,7 @@ if '${WITH_FLUID_ONLY}'== 'OFF':
     package_data['py_paddle']=['*.py','_swig_paddle.so']
 
 package_dir={
-    '': '${CMAKE_CURRENT_SOURCE_DIR}',
+    '': '${PADDLE_BINARY_DIR}/python',
     # The paddle.fluid.proto will be generated while compiling.
     # So that package points to other directory.
     'paddle.fluid.proto.profiler': '${PADDLE_BINARY_DIR}/paddle/fluid/platform',