Files
MachineLearningNotebooks/training/51.distributed-tensorflow-with-parameter-server/51.distributed-tensorflow-with-parameter-server.ipynb
2018-09-14 15:14:43 -04:00

474 lines
18 KiB
Plaintext

{
"cells": [
{
"cell_type": "markdown",
"metadata": {},
"source": [
"Copyright (c) Microsoft Corporation. All rights reserved.\n",
"\n",
"Licensed under the MIT License."
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"# 51. Distributed TensorFlow using Parameter Server\n",
"In this tutorial we demonstrate how to use the Azure ML Training SDK to train Tensorflow model in a distributed manner using Parameter Server.\n",
"\n",
"# Prerequisites\n",
"\n",
"Make sure you go through the [00. Installation and Configuration](00.configuration.ipynb) Notebook first if you haven't."
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"# Check core SDK version number\n",
"import azureml.core\n",
"\n",
"print(\"SDK version:\", azureml.core.VERSION)"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"from azureml.core.workspace import Workspace\n",
"\n",
"ws = Workspace.from_config()\n",
"print('Workspace name: ' + ws.name, \n",
" 'Azure region: ' + ws.location, \n",
" 'Subscription id: ' + ws.subscription_id, \n",
" 'Resource group: ' + ws.resource_group, sep = '\\n')"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"import getpass\n",
"import os\n",
"from azureml.core.experiment import Experiment\n",
"\n",
"username = getpass.getuser().replace('-','')\n",
"\n",
"# choose a name for the run history container in the workspace\n",
"run_history_name = username + '-tf_ps'\n",
"\n",
"experiment = Experiment(ws, run_history_name)\n",
"\n",
"# project folder name\n",
"project_folder = './' + run_history_name\n",
"\n",
"print(project_folder)\n",
"os.makedirs(project_folder, exist_ok = True)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"This recipe is using a MLC-managed Batch AI cluster. "
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"from azureml.core.compute import BatchAiCompute\n",
"from azureml.core.compute import ComputeTarget\n",
"\n",
"batchai_cluster_name='gpucluster'\n",
"\n",
"\n",
"try:\n",
" # Check for existing cluster\n",
" compute_target = ComputeTarget(ws,batchai_cluster_name)\n",
" print('Found existing compute target')\n",
"except:\n",
" # Else, create new one\n",
" print('Creating a new compute target...')\n",
" provisioning_config = BatchAiCompute.provisioning_configuration(vm_size = \"STANDARD_NC6\", # NC6 is GPU-enabled\n",
" #vm_priority = 'lowpriority', # optional\n",
" autoscale_enabled = True,\n",
" cluster_min_nodes = 0, \n",
" cluster_max_nodes = 4)\n",
" compute_target = ComputeTarget.create(ws, batchai_cluster_name, provisioning_config)\n",
" # can poll for a minimum number of nodes and for a specific timeout. \n",
" # if no min node count is provided it will use the scale settings for the cluster\n",
" compute_target.wait_for_completion(show_output=True, min_node_count=None, timeout_in_minutes=20)\n",
"\n",
" # For a more detailed view of current BatchAI cluster status, use the 'status' property \n",
"print(compute_target.status.serialize())"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"%%writefile {project_folder}/mnist_replica.py\n",
"\n",
"# Copyright 2016 The TensorFlow Authors. All Rights Reserved.\n",
"#\n",
"# Licensed under the Apache License, Version 2.0 (the \"License\");\n",
"# you may not use this file except in compliance with the License.\n",
"# You may obtain a copy of the License at\n",
"#\n",
"# http://www.apache.org/licenses/LICENSE-2.0\n",
"#\n",
"# Unless required by applicable law or agreed to in writing, software\n",
"# distributed under the License is distributed on an \"AS IS\" BASIS,\n",
"# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n",
"# See the License for the specific language governing permissions and\n",
"# limitations under the License.\n",
"# ==============================================================================\n",
"\"\"\"Distributed MNIST training and validation, with model replicas.\n",
"A simple softmax model with one hidden layer is defined. The parameters\n",
"(weights and biases) are located on one parameter server (ps), while the ops\n",
"are executed on two worker nodes by default. The TF sessions also run on the\n",
"worker node.\n",
"Multiple invocations of this script can be done in parallel, with different\n",
"values for --task_index. There should be exactly one invocation with\n",
"--task_index, which will create a master session that carries out variable\n",
"initialization. The other, non-master, sessions will wait for the master\n",
"session to finish the initialization before proceeding to the training stage.\n",
"The coordination between the multiple worker invocations occurs due to\n",
"the definition of the parameters on the same ps devices. The parameter updates\n",
"from one worker is visible to all other workers. As such, the workers can\n",
"perform forward computation and gradient calculation in parallel, which\n",
"should lead to increased training speed for the simple model.\n",
"\"\"\"\n",
"\n",
"from __future__ import absolute_import\n",
"from __future__ import division\n",
"from __future__ import print_function\n",
"\n",
"import os\n",
"import math\n",
"import sys\n",
"import tempfile\n",
"import time\n",
"import json\n",
"\n",
"import tensorflow as tf\n",
"from tensorflow.examples.tutorials.mnist import input_data\n",
"from azureml.core.run import Run\n",
"\n",
"flags = tf.app.flags\n",
"flags.DEFINE_string(\"data_dir\", \"/tmp/mnist-data\",\n",
" \"Directory for storing mnist data\")\n",
"flags.DEFINE_boolean(\"download_only\", False,\n",
" \"Only perform downloading of data; Do not proceed to \"\n",
" \"session preparation, model definition or training\")\n",
"flags.DEFINE_integer(\"num_gpus\", 0, \"Total number of gpus for each machine.\"\n",
" \"If you don't use GPU, please set it to '0'\")\n",
"flags.DEFINE_integer(\"replicas_to_aggregate\", None,\n",
" \"Number of replicas to aggregate before parameter update \"\n",
" \"is applied (For sync_replicas mode only; default: \"\n",
" \"num_workers)\")\n",
"flags.DEFINE_integer(\"hidden_units\", 100,\n",
" \"Number of units in the hidden layer of the NN\")\n",
"flags.DEFINE_integer(\"train_steps\", 200,\n",
" \"Number of (global) training steps to perform\")\n",
"flags.DEFINE_integer(\"batch_size\", 100, \"Training batch size\")\n",
"flags.DEFINE_float(\"learning_rate\", 0.01, \"Learning rate\")\n",
"flags.DEFINE_boolean(\n",
" \"sync_replicas\", False,\n",
" \"Use the sync_replicas (synchronized replicas) mode, \"\n",
" \"wherein the parameter updates from workers are aggregated \"\n",
" \"before applied to avoid stale gradients\")\n",
"flags.DEFINE_boolean(\n",
" \"existing_servers\", False, \"Whether servers already exists. If True, \"\n",
" \"will use the worker hosts via their GRPC URLs (one client process \"\n",
" \"per worker host). Otherwise, will create an in-process TensorFlow \"\n",
" \"server.\")\n",
"\n",
"FLAGS = flags.FLAGS\n",
"\n",
"IMAGE_PIXELS = 28\n",
"\n",
"\n",
"def main(unused_argv):\n",
" data_root = os.path.join(\"outputs\", \"MNIST\")\n",
" mnist = None\n",
" tf_config = os.environ.get(\"TF_CONFIG\")\n",
" if not tf_config or tf_config == \"\":\n",
" raise ValueError(\"TF_CONFIG not found.\")\n",
" tf_config_json = json.loads(tf_config)\n",
" cluster = tf_config_json.get('cluster')\n",
" job_name = tf_config_json.get('task', {}).get('type')\n",
" task_index = tf_config_json.get('task', {}).get('index')\n",
" job_name = \"worker\" if job_name == \"master\" else job_name\n",
" sentinel_path = os.path.join(data_root, \"complete.txt\") \n",
" if job_name==\"worker\" and task_index==0:\n",
" mnist = input_data.read_data_sets(data_root, one_hot=True)\n",
" path = os.path.join(data_root, \"complete.txt\") \n",
" with open(sentinel_path, 'w+') as f:\n",
" f.write(\"download complete\")\n",
" else:\n",
" while not os.path.exists(sentinel_path):\n",
" time.sleep(0.01)\n",
" mnist = input_data.read_data_sets(data_root, one_hot=True)\n",
" \n",
" if FLAGS.download_only:\n",
" sys.exit(0)\n",
"\n",
" print(\"job name = %s\" % job_name)\n",
" print(\"task index = %d\" % task_index)\n",
" print(\"number of GPUs = %d\" % FLAGS.num_gpus)\n",
"\n",
" #Construct the cluster and start the server\n",
" cluster_spec = tf.train.ClusterSpec(cluster)\n",
" \n",
" # Get the number of workers.\n",
" num_workers = len(cluster_spec.task_indices(\"worker\"))\n",
"\n",
" if not FLAGS.existing_servers:\n",
" # Not using existing servers. Create an in-process server.\n",
" server = tf.train.Server(\n",
" cluster_spec, job_name=job_name, task_index=task_index)\n",
" if job_name == \"ps\":\n",
" server.join()\n",
"\n",
" is_chief = (task_index == 0)\n",
" if FLAGS.num_gpus > 0:\n",
" # Avoid gpu allocation conflict: now allocate task_num -> #gpu\n",
" # for each worker in the corresponding machine\n",
" gpu = (task_index % FLAGS.num_gpus)\n",
" worker_device = \"/job:worker/task:%d/gpu:%d\" % (task_index, gpu)\n",
" elif FLAGS.num_gpus == 0:\n",
" # Just allocate the CPU to worker server\n",
" cpu = 0\n",
" worker_device = \"/job:worker/task:%d/cpu:%d\" % (task_index, cpu)\n",
" # The device setter will automatically place Variables ops on separate\n",
" # parameter servers (ps). The non-Variable ops will be placed on the workers.\n",
" # The ps use CPU and workers use corresponding GPU\n",
" with tf.device(\n",
" tf.train.replica_device_setter(\n",
" worker_device=worker_device,\n",
" ps_device=\"/job:ps/cpu:0\",\n",
" cluster=cluster)):\n",
" global_step = tf.Variable(0, name=\"global_step\", trainable=False)\n",
"\n",
" # Variables of the hidden layer\n",
" hid_w = tf.Variable(\n",
" tf.truncated_normal(\n",
" [IMAGE_PIXELS * IMAGE_PIXELS, FLAGS.hidden_units],\n",
" stddev=1.0 / IMAGE_PIXELS),\n",
" name=\"hid_w\")\n",
" hid_b = tf.Variable(tf.zeros([FLAGS.hidden_units]), name=\"hid_b\")\n",
"\n",
" # Variables of the softmax layer\n",
" sm_w = tf.Variable(\n",
" tf.truncated_normal(\n",
" [FLAGS.hidden_units, 10],\n",
" stddev=1.0 / math.sqrt(FLAGS.hidden_units)),\n",
" name=\"sm_w\")\n",
" sm_b = tf.Variable(tf.zeros([10]), name=\"sm_b\")\n",
"\n",
" # Ops: located on the worker specified with task_index\n",
" x = tf.placeholder(tf.float32, [None, IMAGE_PIXELS * IMAGE_PIXELS])\n",
" y_ = tf.placeholder(tf.float32, [None, 10])\n",
"\n",
" hid_lin = tf.nn.xw_plus_b(x, hid_w, hid_b)\n",
" hid = tf.nn.relu(hid_lin)\n",
"\n",
" y = tf.nn.softmax(tf.nn.xw_plus_b(hid, sm_w, sm_b))\n",
" cross_entropy = -tf.reduce_sum(y_ * tf.log(tf.clip_by_value(y, 1e-10, 1.0)))\n",
"\n",
" opt = tf.train.AdamOptimizer(FLAGS.learning_rate)\n",
"\n",
" if FLAGS.sync_replicas:\n",
" if FLAGS.replicas_to_aggregate is None:\n",
" replicas_to_aggregate = num_workers\n",
" else:\n",
" replicas_to_aggregate = FLAGS.replicas_to_aggregate\n",
"\n",
" opt = tf.train.SyncReplicasOptimizer(\n",
" opt,\n",
" replicas_to_aggregate=replicas_to_aggregate,\n",
" total_num_replicas=num_workers,\n",
" name=\"mnist_sync_replicas\")\n",
"\n",
" train_step = opt.minimize(cross_entropy, global_step=global_step)\n",
"\n",
" if FLAGS.sync_replicas:\n",
" local_init_op = opt.local_step_init_op\n",
" if is_chief:\n",
" local_init_op = opt.chief_init_op\n",
"\n",
" ready_for_local_init_op = opt.ready_for_local_init_op\n",
"\n",
" # Initial token and chief queue runners required by the sync_replicas mode\n",
" chief_queue_runner = opt.get_chief_queue_runner()\n",
" sync_init_op = opt.get_init_tokens_op()\n",
"\n",
" init_op = tf.global_variables_initializer()\n",
" train_dir = tempfile.mkdtemp()\n",
"\n",
" if FLAGS.sync_replicas:\n",
" sv = tf.train.Supervisor(\n",
" is_chief=is_chief,\n",
" logdir=train_dir,\n",
" init_op=init_op,\n",
" local_init_op=local_init_op,\n",
" ready_for_local_init_op=ready_for_local_init_op,\n",
" recovery_wait_secs=1,\n",
" global_step=global_step)\n",
" else:\n",
" sv = tf.train.Supervisor(\n",
" is_chief=is_chief,\n",
" logdir=train_dir,\n",
" init_op=init_op,\n",
" recovery_wait_secs=1,\n",
" global_step=global_step)\n",
"\n",
" sess_config = tf.ConfigProto(\n",
" allow_soft_placement=True,\n",
" log_device_placement=False,\n",
" device_filters=[\"/job:ps\",\n",
" \"/job:worker/task:%d\" % task_index])\n",
"\n",
" # The chief worker (task_index==0) session will prepare the session,\n",
" # while the remaining workers will wait for the preparation to complete.\n",
" if is_chief:\n",
" print(\"Worker %d: Initializing session...\" % task_index)\n",
" else:\n",
" print(\"Worker %d: Waiting for session to be initialized...\" %\n",
" task_index)\n",
"\n",
" if FLAGS.existing_servers:\n",
" server_grpc_url = \"grpc://\" + worker_spec[task_index]\n",
" print(\"Using existing server at: %s\" % server_grpc_url)\n",
"\n",
" sess = sv.prepare_or_wait_for_session(server_grpc_url, config=sess_config)\n",
" else:\n",
" sess = sv.prepare_or_wait_for_session(server.target, config=sess_config)\n",
"\n",
" print(\"Worker %d: Session initialization complete.\" % task_index)\n",
"\n",
" if FLAGS.sync_replicas and is_chief:\n",
" # Chief worker will start the chief queue runner and call the init op.\n",
" sess.run(sync_init_op)\n",
" sv.start_queue_runners(sess, [chief_queue_runner])\n",
"\n",
" # Perform training\n",
" time_begin = time.time()\n",
" print(\"Training begins @ %f\" % time_begin)\n",
"\n",
" local_step = 0\n",
" while True:\n",
" # Training feed\n",
" batch_xs, batch_ys = mnist.train.next_batch(FLAGS.batch_size)\n",
" train_feed = {x: batch_xs, y_: batch_ys}\n",
"\n",
" _, step = sess.run([train_step, global_step], feed_dict=train_feed)\n",
" local_step += 1\n",
"\n",
" now = time.time()\n",
" print(\"%f: Worker %d: training step %d done (global step: %d)\" %\n",
" (now, task_index, local_step, step))\n",
"\n",
" if step >= FLAGS.train_steps:\n",
" break\n",
"\n",
" time_end = time.time()\n",
" print(\"Training ends @ %f\" % time_end)\n",
" training_time = time_end - time_begin\n",
" print(\"Training elapsed time: %f s\" % training_time)\n",
"\n",
" # Validation feed\n",
" val_feed = {x: mnist.validation.images, y_: mnist.validation.labels}\n",
" val_xent = sess.run(cross_entropy, feed_dict=val_feed)\n",
" print(\"After %d training step(s), validation cross entropy = %g\" %\n",
" (FLAGS.train_steps, val_xent))\n",
" if job_name==\"worker\" and task_index==0:\n",
" run = Run.get_submitted_run()\n",
" run.log(\"CrossEntropy\", val_xent)\n",
"\n",
"if __name__ == \"__main__\":\n",
" tf.app.run()"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"from azureml.train.dnn import *\n",
"tf_estimator = TensorFlow(source_directory=project_folder,\n",
" compute_target=compute_target,\n",
" entry_script='mnist_replica.py',\n",
" node_count=2,\n",
" worker_count=2,\n",
" parameter_server_count=1, \n",
" distributed_backend=\"ps\",\n",
" use_gpu=False)\n"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"run = experiment.submit(tf_estimator)\n",
"print(run)"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"from azureml.train.widgets import RunDetails\n",
"RunDetails(run).show()"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"run.wait_for_completion(show_output=True)"
]
}
],
"metadata": {
"kernelspec": {
"display_name": "Python 3",
"language": "python",
"name": "python3"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.6.5"
}
},
"nbformat": 4,
"nbformat_minor": 2
}