Deep learning models are usually trained using GPUs because GPUs can do a lot more computations in parallel that CPUs. But even with the modern GPUs, it could take several days to train big models. Training can be done faster by using multiple GPUs like described in this tutorial. However only a certain number of GPUs can be attached to one host (typically 8 or 16). To make the training even faster, we can use multiple GPUs attached to multiple hosts.
In this tutorial, we will show how to train a model faster using multi-host distributed training.
We will use data parallelism to distribute the training which involves splitting the training data across GPUs attached to multiple hosts. Since the hosts are working with different subset of the training data in parallel, the training completes a lot faster.
In this tutorial, we will train a ResNet18 network using CIFAR-10 dataset using two hosts each having four GPUs.
Multihost distributed training involves working with three different types of processes - worker, parameter server and scheduler.
The parameters of the model needs to be shared with all hosts since multiple hosts are working together to train one model. To make this sharing efficient, the parameters are split across multiple hosts. A parameter server in each host stores a subset of parameters. In the figure above, parameters are split evenly between the two hosts. At the end of every iteration, each host communicates with every other host to update all parameters of the model.
Each host has a worker process which in each iteration fetches a batch of data, runs forward and backward pass on all GPUs in the host, computes the parameter updates and sends those updates to the parameter servers in each host. Since we have multiple workers to train the model, each worker only needs to process 1/N part of the training data where N is the number of workers.
Scheduler is responsible for scheduling the workers and parameter servers. There is only one scheduler in the entire cluster.
cifar10_dist.py contains code that trains a ResNet18 network using distributed training. In this section we'll walk through parts of that file that are unique to distributed training.
Like mentioned above, in distributed training, parameters are split into N parts and distributed across N hosts. This is done automatically by the distributed key-value store. User only needs to create the distributed kv store and ask the Trainer to use the created store.
store = mxnet.kv.create('dist')
It is the job of the trainer to take the gradients computed in the backward pass and update the parameters of the model. We'll tell the trainer to store and update the parameters in the distributed kv store we just created instead of doing it in GPU of CPU memory. For example,
trainer = gluon.Trainer(net.collect_params(), 'adam', {'learning_rate': .001}, kvstore=store)
In distributed training (using data parallelism), training data is split into equal parts across all workers and each worker uses its subset of the training data for training. For example, if we had two machines, each running a worker, each worker managing four GPUs we‘ll split the data like shown below. Note that we don’t split the data depending on the number of GPUs but split it depending on the number of workers.
Each worker can find out the total number of workers in the cluster and its own rank which is an integer between 0 and N-1 where N is the number of workers.
store = kv.create('dist') print("Total number of workers: %d" % store.num_workers) print("This worker's rank: %d" % store.rank)
Total number of workers: 2 This worker's rank: 0
Knowing the number of workers and a particular worker's rank, it is easy to split the dataset into partitions and pick one partition to train depending on the rank of the worker. Here is a sampler that does exactly that.
class SplitSampler(gluon.data.sampler.Sampler): """ Split the dataset into `num_parts` parts and sample from the part with index `part_index` Parameters ---------- length: int Number of examples in the dataset num_parts: int Partition the data into multiple parts part_index: int The index of the part to read from """ def __init__(self, length, num_parts=1, part_index=0): # Compute the length of each partition self.part_len = length // num_parts # Compute the start index for this partition self.start = self.part_len * part_index # Compute the end index for this partition self.end = self.start + self.part_len def __iter__(self): # Extract examples between `start` and `end`, shuffle and return them. indices = list(range(self.start, self.end)) random.shuffle(indices) return iter(indices) def __len__(self): return self.part_len
We can then create a DataLoader using the SplitSampler like shown below:
# Load the training data train_data = gluon.data.DataLoader(gluon.data.vision.CIFAR10(train=True).transform(transform), batch_size, sampler=SplitSampler(50000, store.num_workers, store.rank))
Note that we didn‘t split the dataset by the number of GPUs. We split it by the number of workers which usually translates to number of machines. It is the worker’s responsibility to split the partition it has across multiple GPUs it might have and run the training in parallel across multiple GPUs.
To train with multiple GPUs, we first need to specify the list of GPUs we want to use for training:
ctx = [mx.gpu(i) for i in range(gpus_per_machine)]
We can then train a batch like shown below:
# Train a batch using multiple GPUs def train_batch(batch, ctx, net, trainer): # Split and load data into multiple GPUs data = batch[0] data = gluon.utils.split_and_load(data, ctx) # Split and load label into multiple GPUs label = batch[1] label = gluon.utils.split_and_load(label, ctx) # Run the forward and backward pass forward_backward(net, data, label) # Update the parameters this_batch_size = batch[0].shape[0] trainer.step(this_batch_size)
Here is the code that runs the forward (computing loss) and backward (computing gradients) pass on multiple GPUs:
# We'll use cross entropy loss since we are doing multiclass classification loss = gluon.loss.SoftmaxCrossEntropyLoss() # Run one forward and backward pass on multiple GPUs def forward_backward(net, data, label): # Ask autograd to remember the forward pass with autograd.record(): # Compute the loss on all GPUs losses = [loss(net(X), Y) for X, Y in zip(data, label)] # Run the backward pass (calculate gradients) on all GPUs for l in losses: l.backward()
Given train_batch, training an epoch is simple:
for batch in train_data: # Train the batch using multiple GPUs train_batch(batch, ctx, net, trainer)
Note that there are several processes that needs to be launched on multiple machines to do distributed training. One worker and one parameter server needs to be launched on each host. Scheduler needs to be launched on one of the hosts. While this can be done manually, MXNet provides the launch.py tool to make this easy.
For example, the following command launches distributed training on two machines:
python ~/mxnet/tools/launch.py -n 2 -s 2 -H hosts \
--sync-dst-dir /home/ubuntu/cifar10_dist \
--launcher ssh \
"python /home/ubuntu/cifar10_dist/cifar10_dist.py"
-n 2 specifies the number of workers that must be launched-s 2 specifies the number of parameter servers that must be launched.--sync-dst-dir specifies a destination location where the contents of the current directory will be rsync'd--launcher ssh tells launch.py to use ssh to login on each machine in the cluster and launch processes."python /home/ubuntu/dist/dist.py" is the command that will get executed in each of the launched processes.-H hosts specifies the list of hosts in the cluster to be used for distributed training.Let's take a look at the hosts file.
~/dist$ cat hosts d1 d2
‘d1’ and ‘d2’ are the hostnames of the hosts we want to run distributed training using. launch.py should be able to ssh into these hosts by providing just the hostname on the command line. For example:
~/dist$ ssh d1
Welcome to Ubuntu 16.04.3 LTS (GNU/Linux 4.4.0-1049-aws x86_64)
* Documentation: https://help.ubuntu.com
* Management: https://landscape.canonical.com
* Support: https://ubuntu.com/advantage
Get cloud support with Ubuntu Advantage Cloud Guest:
http://www.ubuntu.com/business/services/cloud
0 packages can be updated.
0 updates are security updates.
Last login: Wed Jan 31 18:06:45 2018 from 72.21.198.67
Note that no authentication information was provided to login to the host. This can be done using multiple methods. One easy way is to specify the ssh certificates in ~/.ssh/config. Example:
~$ cat ~/.ssh/config
Host d1
HostName ec2-34-201-108-233.compute-1.amazonaws.com
port 22
user ubuntu
IdentityFile /home/ubuntu/my_key.pem
IdentitiesOnly yes
Host d2
HostName ec2-34-238-232-97.compute-1.amazonaws.com
port 22
user ubuntu
IdentityFile /home/ubuntu/my_key.pem
IdentitiesOnly yes
A better way is to use ssh agent forwarding. Check this article for more details.
Here is a sample output from running distributed training:
$ python ~/mxnet/tools/launch.py -n 2 -s 2 -H hosts --sync-dst-dir /home/ubuntu/cifar10_dist --launcher ssh "python /home/ubuntu/cifar10_dist/cifar10_dist.py" 2018-06-03 05:30:05,609 INFO rsync /home/ubuntu/cifar10_dist/ -> a1:/home/ubuntu/cifar10_dist 2018-06-03 05:30:05,879 INFO rsync /home/ubuntu/cifar10_dist/ -> a2:/home/ubuntu/cifar10_dist Epoch 0: Test_acc 0.467400 Epoch 0: Test_acc 0.466800 Epoch 1: Test_acc 0.568500 Epoch 1: Test_acc 0.571300 Epoch 2: Test_acc 0.586300 Epoch 2: Test_acc 0.594000 Epoch 3: Test_acc 0.659200 Epoch 3: Test_acc 0.653300 Epoch 4: Test_acc 0.681200 Epoch 4: Test_acc 0.687900
Note that the output from all hosts are merged and printed to the console.