From 1bf276cc0816ad0a334f166a82cb0d7eeaeb3dfa Mon Sep 17 00:00:00 2001 From: Robert Nishihara Date: Wed, 8 Nov 2017 23:40:51 -0800 Subject: [PATCH] Basic parameter server example. (#1198) * Basic parameter server example. * Consolidate files. * Whitespace. * Add documentation. --- doc/source/example-parameter-server.rst | 64 +++++++ doc/source/index.rst | 1 + examples/parameter_server/model.py | 168 ++++++++++++++++++ examples/parameter_server/parameter_server.py | 83 +++++++++ 4 files changed, 316 insertions(+) create mode 100644 doc/source/example-parameter-server.rst create mode 100644 examples/parameter_server/model.py create mode 100644 examples/parameter_server/parameter_server.py diff --git a/doc/source/example-parameter-server.rst b/doc/source/example-parameter-server.rst new file mode 100644 index 000000000..b9aad1f47 --- /dev/null +++ b/doc/source/example-parameter-server.rst @@ -0,0 +1,64 @@ +Parameter Server +================ + +This document walks through how to implement a simple parameter server example +using actors. To run the application, first install some dependencies. + +.. code-block:: bash + + pip install tensorflow + +You can view the `code for this example`_. + +.. _`code for this example`: https://github.com/ray-project/ray/tree/master/examples/parameter_server + +The example can be run as follows. + +.. code-block:: bash + + python ray/examples/parameter_server/parameter_server.py --num-workers=4 + +Note that this examples uses distributed actor handles, which are still +considered experimental. + +The parameter server itself is implemented as an actor, which exposes the +methods ``push`` and ``pull``. + +.. code-block:: python + + @ray.remote + class ParameterServer(object): + def __init__(self, keys, values): + values = [value.copy() for value in values] + self.weights = dict(zip(keys, values)) + + def push(self, keys, values): + for key, value in zip(keys, values): + self.weights[key] += value + + def pull(self, keys): + return [self.weights[key] for key in keys] + +We then define a worker task, which take a parameter server as an argument and +submits tasks to it. The structure of the code looks as follows. + +.. code-block:: python + + @ray.remote + def worker_task(ps): + while True: + # Get the latest weights from the parameter server. + weights = ray.get(ps.pull.remote(keys)) + + # Compute an update. + ... + + # Push the update to the parameter server. + ps.push.remote(keys, update) + +Then we can create a parameter server and initiate training as follows. + +.. code-block:: python + + ps = ParameterServer.remote(keys, initial_values) + worker_tasks = [worker_task.remote(ps) for _ in range(4)] diff --git a/doc/source/index.rst b/doc/source/index.rst index 1110324a8..aba48550a 100644 --- a/doc/source/index.rst +++ b/doc/source/index.rst @@ -52,6 +52,7 @@ Example Program example-hyperopt.rst example-rl-pong.rst example-policy-gradient.rst + example-parameter-server.rst example-resnet.rst example-a3c.rst example-lbfgs.rst diff --git a/examples/parameter_server/model.py b/examples/parameter_server/model.py new file mode 100644 index 000000000..cdae2698d --- /dev/null +++ b/examples/parameter_server/model.py @@ -0,0 +1,168 @@ +# Most of the tensorflow code is adapted from Tensorflow's tutorial on using +# CNNs to train MNIST +# https://www.tensorflow.org/get_started/mnist/pros#build-a-multilayer-convolutional-network. # noqa: E501 + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import ray +import tensorflow as tf + + +class SimpleCNN(object): + def __init__(self): + with tf.Graph().as_default(): + + # Create the model + self.x = tf.placeholder(tf.float32, [None, 784]) + + # Define loss and optimizer + self.y_ = tf.placeholder(tf.float32, [None, 10]) + + # Build the graph for the deep net + self.y_conv, self.keep_prob = deepnn(self.x) + + with tf.name_scope('loss'): + cross_entropy = tf.nn.softmax_cross_entropy_with_logits( + labels=self.y_, logits=self.y_conv) + self.cross_entropy = tf.reduce_mean(cross_entropy) + + with tf.name_scope('adam_optimizer'): + self.optimizer = tf.train.AdamOptimizer(1e-4) + self.train_step = self.optimizer.minimize( + self.cross_entropy) + + with tf.name_scope('accuracy'): + correct_prediction = tf.equal(tf.argmax(self.y_conv, 1), + tf.argmax(self.y_, 1)) + correct_prediction = tf.cast(correct_prediction, tf.float32) + self.accuracy = tf.reduce_mean(correct_prediction) + + self.sess = tf.Session(config=tf.ConfigProto( + intra_op_parallelism_threads=1, + inter_op_parallelism_threads=1)) + self.sess.run(tf.global_variables_initializer()) + + # Helper values. + + self.variables = ray.experimental.TensorFlowVariables( + self.cross_entropy, self.sess) + + self.grads = self.optimizer.compute_gradients( + self.cross_entropy) + + def compute_update(self, x, y): + # TODO(rkn): Computing the weights before and after the training step + # and taking the diff is awful. + weights = self.get_weights()[1] + self.sess.run(self.train_step, feed_dict={self.x: x, + self.y_: y, + self.keep_prob: 0.5}) + new_weights = self.get_weights()[1] + return [x - y for x, y in zip(new_weights, weights)] + + def compute_gradients(self, x, y): + return self.sess.run([grad[0] for grad in self.grads], + feed_dict={self.x: x, + self.y_: y, + self.keep_prob: 0.5}) + + def compute_accuracy(self, x, y): + return self.sess.run(self.accuracy, + feed_dict={self.x: x, + self.y_: y, + self.keep_prob: 1.0}) + + def set_weights(self, variable_names, weights): + self.variables.set_weights(dict(zip(variable_names, weights))) + + def get_weights(self): + weights = self.variables.get_weights() + return list(weights.keys()), list(weights.values()) + + +def deepnn(x): + """deepnn builds the graph for a deep net for classifying digits. + + Args: + x: an input tensor with the dimensions (N_examples, 784), where 784 is + the number of pixels in a standard MNIST image. + + Returns: + A tuple (y, keep_prob). y is a tensor of shape (N_examples, 10), with + values equal to the logits of classifying the digit into one of 10 + classes (the digits 0-9). keep_prob is a scalar placeholder for the + probability of dropout. + """ + # Reshape to use within a convolutional neural net. + # Last dimension is for "features" - there is only one here, since images + # are grayscale -- it would be 3 for an RGB image, 4 for RGBA, etc. + with tf.name_scope('reshape'): + x_image = tf.reshape(x, [-1, 28, 28, 1]) + + # First convolutional layer - maps one grayscale image to 32 feature maps. + with tf.name_scope('conv1'): + W_conv1 = weight_variable([5, 5, 1, 32]) + b_conv1 = bias_variable([32]) + h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1) + + # Pooling layer - downsamples by 2X. + with tf.name_scope('pool1'): + h_pool1 = max_pool_2x2(h_conv1) + + # Second convolutional layer -- maps 32 feature maps to 64. + with tf.name_scope('conv2'): + W_conv2 = weight_variable([5, 5, 32, 64]) + b_conv2 = bias_variable([64]) + h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2) + + # Second pooling layer. + with tf.name_scope('pool2'): + h_pool2 = max_pool_2x2(h_conv2) + + # Fully connected layer 1 -- after 2 round of downsampling, our 28x28 image + # is down to 7x7x64 feature maps -- maps this to 1024 features. + with tf.name_scope('fc1'): + W_fc1 = weight_variable([7 * 7 * 64, 1024]) + b_fc1 = bias_variable([1024]) + + h_pool2_flat = tf.reshape(h_pool2, [-1, 7 * 7 * 64]) + h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1) + + # Dropout - controls the complexity of the model, prevents co-adaptation of + # features. + with tf.name_scope('dropout'): + keep_prob = tf.placeholder(tf.float32) + h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob) + + # Map the 1024 features to 10 classes, one for each digit + with tf.name_scope('fc2'): + W_fc2 = weight_variable([1024, 10]) + b_fc2 = bias_variable([10]) + + y_conv = tf.matmul(h_fc1_drop, W_fc2) + b_fc2 + return y_conv, keep_prob + + +def conv2d(x, W): + """conv2d returns a 2d convolution layer with full stride.""" + return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME') + + +def max_pool_2x2(x): + """max_pool_2x2 downsamples a feature map by 2X.""" + return tf.nn.max_pool(x, ksize=[1, 2, 2, 1], + strides=[1, 2, 2, 1], padding='SAME') + + +def weight_variable(shape): + """weight_variable generates a weight variable of a given shape.""" + initial = tf.truncated_normal(shape, stddev=0.1) + return tf.Variable(initial) + + +def bias_variable(shape): + """bias_variable generates a bias variable of a given shape.""" + initial = tf.constant(0.1, shape=shape) + return tf.Variable(initial) diff --git a/examples/parameter_server/parameter_server.py b/examples/parameter_server/parameter_server.py new file mode 100644 index 000000000..dd6192e17 --- /dev/null +++ b/examples/parameter_server/parameter_server.py @@ -0,0 +1,83 @@ +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import argparse +from tensorflow.examples.tutorials.mnist import input_data +import time + +import ray + +import model + +parser = argparse.ArgumentParser(description="Run the parameter server " + "example.") +parser.add_argument("--num-workers", default=4, type=int, + help="The number of workers to use.") +parser.add_argument("--redis-address", default=None, type=str, + help="The Redis address of the cluster.") + + +@ray.remote +class ParameterServer(object): + def __init__(self, keys, values): + # These values will be mutated, so we must create a copy that is not + # backed by the object store. + values = [value.copy() for value in values] + self.weights = dict(zip(keys, values)) + + def push(self, keys, values): + for key, value in zip(keys, values): + self.weights[key] += value + + def pull(self, keys): + return [self.weights[key] for key in keys] + + +@ray.remote +def worker_task(ps): + # Download MNIST. + mnist = input_data.read_data_sets("MNIST_data", one_hot=True) + batch_size = 50 + + # Initialize the model. + net = model.SimpleCNN() + keys = net.get_weights()[0] + + while True: + # Get the current weights from the parameter server. + weights = ray.get(ps.pull.remote(keys)) + net.set_weights(keys, weights) + + # Compute an update and push it to the parameter server. + xs, ys = mnist.train.next_batch(batch_size) + gradients = net.compute_update(xs, ys) + ps.push.remote(keys, gradients) + + +if __name__ == '__main__': + args = parser.parse_args() + + ray.init(redis_address=args.redis_address) + + # Create a parameter server with some random weights. + net = model.SimpleCNN() + all_keys, all_values = net.get_weights() + ps = ParameterServer.remote(all_keys, all_values) + + # Start some training tasks. + worker_tasks = [worker_task.remote(ps) for _ in range(args.num_workers)] + + # Download MNIST. + mnist = input_data.read_data_sets("MNIST_data", one_hot=True) + + i = 0 + while True: + # Get and evaluate the current model. + current_weights = ray.get(ps.pull.remote(all_keys)) + net.set_weights(all_keys, current_weights) + test_xs, test_ys = mnist.test.next_batch(1000) + accuracy = net.compute_accuracy(test_xs, test_ys) + print("Iteration {}: accuracy is {}".format(i, accuracy)) + i += 1 + time.sleep(1)