larroy commented on a change in pull request #14286: Add examples of running MXNet with Horovod URL: https://github.com/apache/incubator-mxnet/pull/14286#discussion_r268352407
########## File path: example/distributed_training-horovod/README.md ########## @@ -0,0 +1,201 @@ +<!--- Licensed to the Apache Software Foundation (ASF) under one --> +<!--- or more contributor license agreements. See the NOTICE file --> +<!--- distributed with this work for additional information --> +<!--- regarding copyright ownership. The ASF licenses this file --> +<!--- to you 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. --> + +# Distributed Training using MXNet with Horovod +[Horovod](https://github.com/horovod/horovod) is a distributed training framework that demonstrates +excellent scaling efficiency for dense models running on a large number of nodes. It currently +supports mainstream deep learning frameworks such as MXNet, TensorFlow, Keras, and PyTorch. +It is created at Uber and currently hosted by the [Linux Foundation Deep Learning](https://lfdl.io)(LF DL). + +MXNet is supported in Horovod 0.16.0 [release](https://eng.uber.com/horovod-pyspark-apache-mxnet-support/). + +## What's New? +Compared with the standard distributed training script in MXNet which uses parameter server to +distribute and aggregate parameters, Horovod uses ring allreduce and/or tree-based allreduce algorithm +to communicate parameters between workers. There is no dedicated server and the communication data size +between workers does not depend on the number of workers. Therefore, it scales well in the case where +there are a large number of workers and network bandwidth is the bottleneck. + +# Install +## Install MXNet +```bash +$ pip install mxnet +``` +**Note**: There is a [known issue](https://github.com/horovod/horovod/issues/884) when running Horovod with MXNet on a Linux system with GCC version 5.X and above. We recommend users to build MXNet from source following this [guide](https://mxnet.incubator.apache.org/install/build_from_source.html) as a workaround for now. Also mxnet-mkl package in 1.4.0 release does not support Horovod. + +## Install Horovod +```bash +$ pip install horovod +``` + +This basic installation is good for laptops and for getting to know Horovod. +If you're installing Horovod on a server with GPUs, read the [Horovod on GPU](https://github.com/horovod/horovod/blob/master/docs/gpus.md) page. +If you want to use Docker, read the [Horovod in Docker](https://github.com/horovod/horovod/blob/master/docs/docker.md) page. + +## Install MPI +MPI is required to run distributed training with Horovod. Install [Open MPI](https://www.open-mpi.org/) or another MPI implementation. +Steps to install Open MPI are listed [here](https://www.open-mpi.org/faq/?category=building#easy-build). + +**Note**: Open MPI 3.1.3 has an issue that may cause hangs. It is recommended +to downgrade to Open MPI 3.1.2 or upgrade to Open MPI 4.0.0. + +# Usage + +To run MXNet with Horovod, make the following additions to your training script: + +1. Run `hvd.init()`. + +2. Pin the context to a processor using `hvd.local_rank()`. + Typically, each Horovod worker is associated with one process. The local rank is a unique ID specifically + for all processes running Horovod job on the same node. + +3. Scale the learning rate by number of workers. Effective batch size in synchronous distributed training is scaled by + the number of workers. An increase in learning rate compensates for the increased batch size. + +4. Wrap optimizer in `hvd.DistributedOptimizer`. The distributed optimizer delegates gradient computation + to the original optimizer, averages gradients using *allreduce* or *allgather*, and then applies those averaged + gradients. + +5. Add `hvd.broadcast_parameters` to broadcast initial variable states from rank 0 to all other processes. + This is necessary to ensure consistent initialization of all workers when training is started with random weights or + restored from a checkpoint. + +# Example + +Here we provide the building blocks to train a model using MXNet with Horovod. +The full examples are in [MNIST](gluon_mnist.py) and [ImageNet](resnet50_imagenet.py). + +## Gluon API +```python +from mxnet import autograd, gluon +import mxnet as mx +import horovod.mxnet as hvd + +# Initialize Horovod +hvd.init() + +# Set context to current process +context = mx.cpu(hvd.local_rank()) if args.no_cuda else mx.gpu(hvd.local_rank()) Review comment: I think args is not defined yet. Maybe context.num_gpus()? ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected] With regards, Apache Git Services
