sandeep-krishnamurthy closed pull request #11910: Improving documentation and error messages for Async distributed training with Gluon URL: https://github.com/apache/incubator-mxnet/pull/11910
This is a PR merged from a forked repository. As GitHub hides the original diff on merge, it is displayed below for the sake of provenance: As this is a foreign pull request (from a fork), the diff is supplied below (as it won't show otherwise due to GitHub magic): diff --git a/docs/faq/distributed_training.md b/docs/faq/distributed_training.md index 70078ba6095..d4fa72db23a 100644 --- a/docs/faq/distributed_training.md +++ b/docs/faq/distributed_training.md @@ -73,6 +73,23 @@ These can be passed as arguments to the iterator. You can look at [example/gluon/image_classification.py](https://github.com/apache/incubator-mxnet/blob/master/example/gluon/image_classification.py) to see an example usage. +### Updating weights +KVStore server supports two modes, one which aggregates the gradients and updates the weights using those gradients, and second where the server only aggregates gradients. In the latter case, when a worker process pulls from kvstore, it gets the aggregated gradients. The worker then uses these gradients and applies the weights locally. + +When using Gluon there is an option to choose between these modes by passing `update_on_kvstore` variable when you create the [Trainer](https://mxnet.incubator.apache.org/versions/master/api/python/gluon/gluon.html#mxnet.gluon.Trainer) object like this: + +``` +trainer = gluon.Trainer(net.collect_params(), optimizer='sgd', + optimizer_params={'learning_rate': opt.lr, + 'wd': opt.wd, + 'momentum': opt.momentum, + 'multi_precision': True}, + kvstore=kv, + update_on_kvstore=True) +``` + +When using the symbolic interface, it performs the weight updates on the server without the user having to do anything special. + ### Different Modes of Distributed Training Distributed training itself is enabled when kvstore creation string contains the word `dist`. @@ -86,9 +103,9 @@ In this mode, if a worker crashes, then it halts the progress of all workers. - `dist_async`: In asynchronous distributed training, the server receives gradients from one worker and immediately updates its store, which it uses to respond to any future pulls. This means that a worker who finishes processing a batch can pull the current parameters from server and start the next batch, even if other workers haven't finished processing the earlier batch. -This is faster than `dist_sync` but can take more epochs to converge. -In `async` mode, it is required to pass an optimizer because in the absence of an optimizer kvstore would replace the stored weights with received weights and this doesn't make sense for training in asynchronous mode. +This is faster than `dist_sync` because there is no cost of synchronization, but can take more epochs to converge. The update of weights is atomic, meaning no two updates happen on the same weight at the same time. However, the order of updates is not guaranteed. +In `async` mode, it is required to pass an optimizer because in the absence of an optimizer kvstore would replace the stored weights with received weights and this doesn't make sense for training in asynchronous mode. Hence, when using Gluon with `async` mode we need to set `update_on_kvstore` to `True`. - `dist_sync_device`: Same as `dist_sync` except that when there are multiple GPUs being used on each node, this mode aggregates gradients and updates weights on GPU while dist_sync does so on CPU memory. diff --git a/python/mxnet/gluon/trainer.py b/python/mxnet/gluon/trainer.py index b4263410a50..98a6878b94b 100644 --- a/python/mxnet/gluon/trainer.py +++ b/python/mxnet/gluon/trainer.py @@ -187,6 +187,11 @@ def _init_kvstore(self): arg_arrays = {param.name: param.data(self._contexts[0]) for param in self._params} kvstore, update_on_kvstore = _create_kvstore(config['kvstore'], len(self._contexts), arg_arrays) + if kvstore and 'async' in kvstore.type and config['update_on_kvstore'] is not None\ + and not config['update_on_kvstore']: + raise ValueError("Please set update_on_kvstore to true " + "when training in async mode.") + if config['update_on_kvstore'] is not None: update_on_kvstore = config['update_on_kvstore'] if kvstore: @@ -195,7 +200,8 @@ def _init_kvstore(self): self._distributed = 'dist' in kvstore.type if self._distributed: # kv.pull(row_sparse_grad) is not supported for dist kvstore - update_on_kvstore = self._contains_sparse_weight or self._contains_sparse_grad + update_on_kvstore = self._contains_sparse_weight or self._contains_sparse_grad \ + or 'async' in kvstore.type if update_on_kvstore: # optimizer preferably needs to be set before init for multiprecision kvstore.set_optimizer(self._optimizer) diff --git a/tests/nightly/dist_async_kvstore.py b/tests/nightly/dist_async_kvstore.py new file mode 100644 index 00000000000..3e400eafa04 --- /dev/null +++ b/tests/nightly/dist_async_kvstore.py @@ -0,0 +1,48 @@ +#!/usr/bin/env python + +# 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. + +# pylint: skip-file +import sys +sys.path.insert(0, "../../python/") +import mxnet as mx + +kv = mx.kv.create('dist_async') +my_rank = kv.rank +nworker = kv.num_workers + +def test_gluon_trainer_type(): + def check_trainer_kv_update(update_on_kv): + params = mx.gluon.ParameterDict() + x = params.get('x', shape=(10,1), lr_mult=1.0) + params.initialize(ctx=[mx.cpu(0), mx.cpu(1)], init='zeros') + try: + trainer = mx.gluon.Trainer(params, 'sgd', {'learning_rate': 0.1}, kvstore=kv, update_on_kvstore=update_on_kv) + trainer._init_kvstore() + assert trainer._kv_initialized + assert trainer._update_on_kvstore is True + except ValueError: + assert update_on_kv is False + + check_trainer_kv_update(False) + check_trainer_kv_update(True) + check_trainer_kv_update(None) + print('worker ' + str(my_rank) + ' passed test_gluon_trainer_type') + +if __name__ == "__main__": + test_gluon_trainer_type() \ No newline at end of file ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services