What strategies and forms of parallelization are feasible and available for training and serving a neural network?:
inside a machine across cores (e.g. GPU / TPU / CPU)
across machines on a network or a rack
I'm also looking for evidence for how they may also be used in e.g. TensorFlow, PyTorch or MXNet.
Training
To my knowledge, when training large neural networks on large datasets, one could at least have:
Different cores or machines operate on different parts of the graph ("graph splitting"). E.g. backpropagation through the graph itself can be parallelized e.g. by having different layers hosted on different machines since (I think?) the autodiff graph is always a DAG.
Different cores or machines operate on different samples of data ("data splitting"). In SGD, the computation of gradients across batches or samples can also be parallelized (e.g. the gradients can be combined after computing them independently on different batches). I believe this is also called gradient accumulation (?).
When is each strategy better for what type of problem or neural network? Which modes are supported by modern libraries? and can one combine all four (2x2) strategies?
On top of that, I have read about:
Asynchronous training
Synchronous training
but I don't know what exactly that refers to, e.g. is it the computation of gradients on different data batches or the computation of gradients on different subgraphs? Or perhaps it refers to something else altogether?
Serving
If the network is huge, prediction / inference may also be slow, and the model may not fit on a single machine in memory at serving time. Are there any known multi-core and multi-node prediction solutions that work that can handle such models?
Training
In general, there are two strategies of parallelizing model training: data parallelism and model parallelism.
1. Data parallelism
This strategy splits training data into N partitions, each of which will be trained on different “devices” (different CPU cores, GPUs, or even machines). In contrast to training without data parallelism which produces one gradient per minibatch, we now have N gradients for each minibatch step. The next question is how we should combine these N gradients.
One way to do it is by averaging all the N gradients and then updating the model parameters once based on the average. This technique is called synchronous distributed SGD. By doing the average, we have a more accurate gradient, but with a cost of waiting all the devices to finish computing its own local gradient.
Another way is by not combining the gradients — each gradient will instead be used to update the model parameters independently. So, there will be N parameter updates for each minibatch step, in contrast to only one for the previous technique. This technique is called asynchronous distributed SGD. Because it doesn't have to wait other devices to finish, the async approach will take less time to complete a minibatch step than the sync approach will do. However, the async approach will produce a more noisy gradient, so it might need to complete more minibatch steps to catch up with the performance (in terms of loss) of the sync approach.
There are many papers proposing some improvements and optimizations on either approach, but the main idea is generally the same as described above.
In the literature there's been some disagreement on which technique is better in practice. At the end most people now settle on the synchronous approach.
Data Parallelism in PyTorch
To do synchronous SGD, we can wrap our model with torch.nn.parallel.DistributedDataParallel:
from torch.nn.parallel import DistributedDataParallel as DDP
# `model` is the model we previously initialized
model = ...
# `rank` is a device number starting from 0
model = model.to(rank)
ddp_model = DDP(model, device_ids=[rank])
Then we can train it similarly. For more details, you can refer to the official tutorial.
For doing asynchronous SGD in PyTorch, we need to implement it more manually since there is no wrapper similar to DistributedDataParallel for it.
Data Parallelism in TensorFlow/Keras
For synchronous SGD, we can use tf.distribute.MirroredStrategy to wrap the model initalization:
import tensorflow as tf
strategy = tf.distribute.MirroredStrategy()
with strategy.scope():
model = Model(...)
model.compile(...)
Then we can train it as usual. For more details, you can refer to the official guides on Keras website and TensorFlow website.
For asynchronous SGD, we can use tf.distribute.experimental.ParameterServerStrategy similarly.
2. Model Parallelism
This strategy splits the model into N parts, each of which will be computed on different devices. A common way to split the model is based on layers: different sets of layers are placed on different devices. But we can also split it more intricately depending on the model architecture.
Model Parallelism in TensorFlow and PyTorch
To implement model parallelism in either TensorFlow or PyTorch, the idea is the same: to move some model parameters into a different device.
In PyTorch we can use torch.nn.Module.to method to move a module into a different device. For example, suppose we want to create two linear layers each of which is placed on a different GPU:
import torch.nn as nn
linear1 = nn.Linear(16, 8).to('cuda:0')
linear2 = nn.Linear(8, 4).to('cuda:1')
In TensorFlow we can use tf.device to place an operation into a specific device. To implement the PyTorch example above in TensorFlow:
import tensorflow as tf
from tensorflow.keras import layers
with tf.device('/GPU:0'):
linear1 = layers.Dense(8, input_dim=16)
with tf.device('/GPU:1'):
linear2 = layers.Dense(4, input_dim=8)
For more details you can refer to the official PyTorch tutorial; or if you use TensorFlow you can even use a more high-level library like mesh.
3. Hybrid: Data and Model Parallelism
Recall that data parallelism only splits the training data, whereas model parallelism only splits the model structures. If we have a model so large that even after using either parallelism strategy it still doesn't fit in the memory, we can always do both.
In practice most people prefer data parallelism to model parallelism since the former is more decoupled (in fact, independent) from the model architecture than the latter. That is, by using data parallelism they can change the model architecture as they like, without worrying which part of the model should be parallelized.
Model Inference / Serving
Parallelizing model serving is easier than parallelizing model training since the model parameters are already fixed and each request can be processed independently. Similar to scaling a regular Python web service, we can scale model serving by spawning more processes (to workaround Python's GIL) in a single machine, or even spawning more machine instances.
When we use a GPU to serve the model, though, we need to do more work to scale it. Because of how concurrency is handled differently by a GPU compared to a CPU, in order to maximize the performance, we need to do inference request batching. The idea is when a request comes, instead of immediately processing it, we wait some timeout duration for other requests to come. When the timeout is up, even if the number of requests is only one, we batch them all to be processed on the GPU.
In order to minimize the average request latency, we need to find the optimal timeout duration. To find it we need to observe that there is a trade-off between minimizing the timeout duration and maximizing the number of batch size. If the timeout is too low, the batch size will be small, so the GPU will be underutilized. But if the timeout is too high, the requests that come early will wait too long before they get processed. So, the optimal timeout duration depends on the model complexity (hence, the inference duration) and the average requests per second to receive.
Implementing a scheduler to do request batching is not a trivial task, so instead of doing it manually, we'd better use TensorFlow Serving or PyTorch Serve which already supports it.
To learn more about parallel and distributed learning, you can read this review paper.
As the question is quite broad, I'll try to shed a little different light and touch on different topics than what was shown in
#Daniel's in-depth answer.
Training
Data parallelization vs model parallelization
As mentioned by #Daniel data parallelism is used way more often and is easier to do correctly. Major caveat of model parallelism is the need to wait for part of neural network and synchronization between them.
Say you have a simple feedforward 5 layer neural network spread across 5 different GPUs, each layer for one device. In this case, during each forward pass each device has to wait for computations from the previous layers. In this simplistic case, copying data between devices and synchronization would take a lot longer and won't bring benefits.
On the other hand, there are models better suited for model parallelization like Inception networks, see picture below:
Here you can see 4 independent paths from previous layer which could go in parallel and only 2 synchronization points (Filter concatenation and Previous Layer).
Questions
E.g. backpropagation through the graph itself can be parallelized e.g.
by having different layers hosted on different machines since (I
think?) the autodiff graph is always a DAG.
It's not that easy. Gradients are calculated based on the loss value (usually) and you need to know gradients of deeper layers to calculate gradients for the more shallow ones. As above, if you have independent paths it's easier and may help, but it's way easier on a single device.
I believe this is also called gradient accumulation (?)
No, it's actually reduction across multiple devices. You can see some of that in PyTorch tutorial. Gradient accumulation is when you run your forward pass (either on single or multiple devices) N times and backpropagate (the gradient is kept in the graph and the values are added during each pass) and optimizer only makes a single step to change neural network's weights (and clears the gradient). In this case, loss is usually divided by the number of steps without optimizer. This is used for more reliable gradient estimation, usually when you are unable to use large batches.
Reduction across devices looks like this:
This is all-reduce in data parallelization, each device calculates the values which are send to all other devices and backpropagated there.
When is each strategy better for what type of problem or neural
network?
Described above, data parallel is almost always fine if you have enough of data and the samples are big (up to 8k samples or more can be done at once without very big struggle).
Which modes are supported by modern libraries?
tensorflow and pytorch both support either, most modern and maintained libraries have those functionalities implemented one way or another
can one combine all four (2x2) strategies
Yes, you can parallelize both model and data across and within machines.
synchronous vs asynchronous
asynchronous
Described by #Daniel in brief, but it's worth mentioning updates are not totally separate. That would make little sense, as we would essentially train N different models based on their batches.
Instead, there is a global parameter space, where each replica is supposed to share calculated updates asynchronously (so forward pass, backward, calculate update with optimizer and share this update to global params).
This approach has one problem though: there is no guarantee that when one worker calculated forward pass another worker updated the parameters, so the update is calculated with respect to old set of params and this is called stale gradients. Due to this, convergence might be hurt.
Other approach is to calculate N steps and updates for each worker and synchronize them afterwards, though it's not used as often.
This part was based on great blogpost and you should definitely read it if interested (there is more about staleness and some solutions).
synchronous
Mostly described previously, there are different approaches but PyTorch gathers output from network and backpropagates on them (torch.nn.parallel.DistributedDataParallel)[https://pytorch.org/docs/stable/nn.html#torch.nn.parallel.DistributedDataParallel]. BTW. You should solely this (no torch.nn.DataParallel) as it overcomes Python's GIL problem.
Takeaways
Data parallelization is always almost used when going for speed up as you "only" have to replicate neural network on each device (either over the network or within single machine), run part of batch on each during the forward pass, concatenate them into a single batch (synchronization) on one device and backpropagate on said.
There are multiple ways to do data parallelization, already introduced by #Daniel
Model parallelization is done when the model is too large to fit on single machine (OpenAI's GPT-3 would be an extreme case) or when the architecture is suited for this task, but both are rarely the case AFAIK.
The more and the longer parallel paths the model has (synchronization points), the better it might be suited for model parallelization
It's important to start workers at similar times with similar loads in order not to way for synchronization processes in synchronous approach or not to get stale gradients in asynchronous (though in the latter case it's not enough).
Serving
Small models
As you are after large models I won't delve into options for smaller ones, just a brief mention.
If you want to serve multiple users over the network you need some way to scale your architecture (usually cloud like GCP or AWS). You could do that using Kubernetes and it's PODs or pre-allocate some servers to handle requests, but that approach would be inefficient (small number of users and running servers would generate pointless costs, while large numbers of users may halt the infrastructure and take too long to process resuests).
Other way is to use autoscaling based on serverless approach. Resources will be provided based on each request so it has large scaling abilities + you don't pay when the traffic is low. You can see Azure Functions as they are on the path to improve it for ML/DL tasks, or torchlambda for PyTorch (disclaimer, I'm the author) for smaller models.
Large models
As mentioned previously, you could use Kubernetes with your custom code or ready to use tools.
In the first case, you can spread the model just the same as for training, but only do forward pass. In this way even giant models can be put up on the network (once again, GPT-3 with 175B parameters), but requires a lot of work.
In the second, #Daniel provided two possibilities. Others worth mentioning could be (read respective docs as those have a lot of functionalities):
KubeFlow - multiple frameworks, based on Kubernetes (so auto-scaling, multi-node), training, serving and what not, connects with other things like MLFlow below
AWS SageMaker - training and serving with Python API, supported by Amazon
MLFlow - multiple frameworks, for experiment handling and serving
BentoML - multiple frameworks, training and serving
For PyTorch, you could read more here, while tensorflow has a lot of serving functionality out of the box via Tensorflow EXtended (TFX).
Questions from OP's comment
Are there any forms of parallelism that are better within a machine vs
across machines
The best for of parallelism would probably be within one giant computer as to minimize transfer between devices.
Additionally, there are different backends (at least in PyTorch) one can choose from (mpi, gloo, nccl) and not all of them support direct sending, receiving, reducing etc. data between devices (some may support CPU to CPU, others GPU to GPU). If there is no direct link between devices, those have to be first copied to another device and copied again to target device (e.g. GPU on other machine -> CPU on host -> GPU on host). See pytorch info.
The more data and the bigger network, the more profitable it should be to parallelize computations. If whole dataset can be fit on a single device there is no need for parallelization. Additionally, one should take into account things like internet transfer speed, network reliability etc. Those costs may outweigh benefits.
In general, go for data parallelization if you have lots of of data (say ImageNet with 1.000.000 images) or big samples (say images 2000x2000). If possible, within a single machine as to minimize between-machines transfer. Distribute model only if there is no way around it (e.g. it doesn't fit on GPU). Don't otherwise (there is little to no point to parallelize when training MNIST as the whole dataset will easily fit in RAM and the read will be fastest from it).
why bother build custom ML-specific hardware such as TPUs?
CPUs are not the best suited for highly parallel computations (e.g. matrices multiplication) + CPU may be occupied with many other tasks (like data loading), hence it makes sense to use GPU.
As GPU was created with graphics in mind (so algebraic transformation), it can take some of CPU duties and can be specialized (many more cores when compared to CPU but simpler ones, see V100 for example).
Now, TPUs are tailored specificially for tensor computations (so deep learning mainly) and originated in Google, still WIP when compared to GPUs. Those are suited for certain types of models (mainly convolutional neural networks) and can bring speedups in this case. Additionally one should use the largest batches with this device (see here), best to be divisible by 128. You can compare that to NVidia's Tensor Cores technology (GPU) where you are fine with batches (or layer sizes) divisible by 16 or 8 (float16 precision and int8 respectively) for good utilization (although the more the better and depends on number of cores, exact graphic card and many other stuff, see some guidelines here).
On the other hand, TPUs support still isn't the best, although two major frameworks support it (tensorflow officially, while PyTorch with torch_xla package).
In general, GPU is a good default choice in deep learning right now, TPUs for convolution heavy architectures, though might give some headache tbh. Also (once again thanks #Daniel), TPUs are more power effective, hence should be cheaper when comparing single floating point operation cost.
Let's say I trained a model with a very complex computational graph tailored for training. After a lot of training, the best model was saved to a checkpoint file. Now, I want to use the learned parameters of this best model for inference. However, the computational graph used for training is not exactly the same as the one I intend to use for inference. Concretely, there is a module in the graph with several layers in charge of outputting embedding vectors for items (recommender system context). However, for the sake of computational performance, during inference time I would like to have all the item embedding vectors precomputed in advance, so that the only computation required per request would just involve a couple of hidden layers.
Therefore, what I would like to know how to do is:
How to just restore the part of the network that outputs item embedding vectors, in order to precompute these vectors for all items (this would happen in some pre-processing script off-line)
Once all item embedding vectors are precomputed, during on-line inference time how to just restore the hidden layers in the later parts of the network and make them receive the precomputed item embedding vectors instead.
How can the points above be accomplished? I think point 1. is easier to get done. But my biggest concern is with point 2. In the computational graph used for training, in order to evaluate any layer I would have to provide values for the input placeholders. However, during on-line inference these placeholders would be obsolete because a lot of stuff would be precomputed and I don't know how to tell hidden layers in the later parts of the network that they should no longer depend on these obsolete placeholders but depend on the precomputed stuff instead.
When I was learning tensorflow, one basic concept of tensorflow was computational graphs, and the graphs was said to be static.
And I found in Pytorch, the graphs was said to be dynamic.
What's the difference of static Computational Graphs in tensorflow and dynamic Computational Graphs in Pytorch?
Both frameworks operate on tensors and view any model as a directed acyclic graph (DAG), but they differ drastically on how you can define them.
TensorFlow follows ‘data as code and code is data’ idiom. In TensorFlow you define graph statically before a model can run. All communication with outer world is performed via tf.Session object and tf.Placeholder which are tensors that will be substituted by external data at runtime.
In PyTorch things are way more imperative and dynamic: you can define, change and execute nodes as you go, no special session interfaces or placeholders. Overall, the framework is more tightly integrated with Python language and feels more native most of the times. When you write in TensorFlow sometimes you feel that your model is behind a brick wall with several tiny holes to communicate over. Anyways, this still sounds like a matter of taste more or less.
However, those approaches differ not only in a software engineering perspective: there are several dynamic neural network architectures that can benefit from the dynamic approach. Recall RNNs: with static graphs, the input sequence length will stay constant. This means that if you develop a sentiment analysis model for English sentences you must fix the sentence length to some maximum value and pad all smaller sequences with zeros. Not too convenient, huh. And you will get more problems in the domain of recursive RNNs and tree-RNNs. Currently Tensorflow has limited support for dynamic inputs via Tensorflow Fold. PyTorch has it by-default.
Reference:
https://medium.com/towards-data-science/pytorch-vs-tensorflow-spotting-the-difference-25c75777377b
https://www.reddit.com/r/MachineLearning/comments/5w3q74/d_so_pytorch_vs_tensorflow_whats_the_verdict_on/
Both TensorFlow and PyTorch allow specifying new computations at any point in time. However, TensorFlow has a "compilation" steps which incurs performance penalty every time you modify the graph. So TensorFlow optimal performance is achieved when you specify the computation once, and then flow new data through the same sequence of computations.
It's similar to interpreters vs. compilers -- the compilation step makes things faster, but also discourages people from modifying the program too often.
To make things concrete, when you modify the graph in TensorFlow (by appending new computations using regular API, or removing some computation using tf.contrib.graph_editor), this line is triggered in session.py. It will serialize the graph, and then the underlying runtime will rerun some optimizations which can take extra time, perhaps 200usec. In contrast, running an op in previously defined graph, or in numpy/PyTorch can be as low as 1 usec.
In tensorflow you first have to define the graph, then you execute it.
Once defined you graph is immutable: you can't add/remove nodes at runtime.
In pytorch, instead, you can change the structure of the graph at runtime: you can thus add/remove nodes at runtime, dynamically changing its structure.
I am trying to train a very deep model on Cloud ML however i am having serious memory issues that i am not managing to go around. The model is a very deep convolutional neural network to auto-tag music.
The model for this can be found in the image below. A batch of 20 with a tensor of 12x38832x1 is inserted in the network.
The music was originally 465894x1 samples which was then split into 12 windows. Hence, 12x38832x1. When using the map_fn function each loop would have the seperate 38832x1 samples (conv1d).
Processing windows at a time yields better results than the whole music using one CNN. This was split prior to storing the data in TFRecords in order to minimise the needed processing during training. This is loaded in a queue with maximum queue size of 200 samples (ie 10 batches).
Once dequeue, it is transposed to have the 12 dimension first which then can be used in the map_fn function for processing of the windows. This is not transposed prior to being queued as the first dimension needs to match the batch dimension of the output which is [20, 50]. Where 20 is the batch size as the data and 50 are the different tags.
For each window, the data is processed and the results of each map_fn are superpooled using a smaller network. The processing of the windows is done by a very deep neural network which is giving me problems to keep as all the config options i am giving are giving me out of memory errors.
As a model i am using one similar to Census Tensorflow Model.
First and foremost, i am not sure if this is the best option since for evaluation a separate graph is built and not shared variables. This would require double the amount of parameters.
Secondly, as a cluster setup, i have been using one complex_l master, 3 complex_l workers and 3 large_model parameter servers. I do not know if am underestimating the amount of memory needed here.
My model has previously worked with a much smaller network. However, increasing it in size started giving me bad out of memory errors.
My questions are:
The memory requirement is big, but i am sure it can be processed on cloud ml. Am i underestimating the amount of memory needed? What are your suggestions about the cluster for such a network?
When using a train.server in the dispatch function, do you need to pass on the cluster_spec so it is used in the replica_device setter? Or does it allocate on it's own? When not using it, and setting tf.configProto of log placement, all the variables seem to be on the master worker. On the Census Example in the task.py this is not passed on. I can assume this is correct?
How does one calculate how much memory is needed for a model (rough estimate to select the cluster)?
Is there any other tensorflow core tutorial how to setup such big jobs? (other than Census)
When training a big model in distributed between-graph replication, does all the model need to fit on the worker, or the worker only does ops and then transmits the results to the PS. Does that mean that the workers can have low memory just for singular ops?
PS: With smaller models the network trained successfully. I am trying to deepen the network for better ROC.
Questions coming up from on-going troubleshooting:
When using the replica_device_setter with the parameter cluster, i noticed that the master has very little memory and CPU usage and checking the log placement there are very little ops on the master. I checked the TF_CONFIG that is loaded and it says the following for the cluster field:
u'cluster': {u'ps': [u'ps-4da746af4e-0:2222'], u'worker': [u'worker-4da746af4e-0:2222'], u'master': [u'master-4da746af4e-0:2222']}
On the other hand, in the tf.train.Clusterspec documentation, it only shows workers. Does that mean that the master is not considered as worker? What happens in such case?
Error is it Memory or something else? EOF Error?
Is it possible to share a queue between two graphs in TensorFlow? I'd like to do a kind of bootstrapping to select "hard negative" examples during training.
To speed up the process, I want separate threads for hard negative example selection, and for the training process. The hard negative selection is based on the evaluation of the current model, and it will load its graph from a checkpoint file. The training graph is run on another thread and writes the checkpoint file. The two graphs should share the same queue: the training graph will consume examples and the hard negative selection will produce them.
Currently there's no support for sharing state between different graphs in the open-source version of TensorFlow: each graph runs in a separate session, and each session uses an isolated set of devices.
However, it seems like it would be possible to achieve your goal using a queue in single graph. Simply construct a queue (using e.g. tf.FIFOQueue) and use tf.import_graph_def() to import the graph from the checkpoint file into the current graph. Using the return_elements argument to tf.import_graph_def() you can specify the name of the tensor that will contain the negative examples, and then add a q.enqueue_many() operation to add them to your queue. You would then fork a thread to run the enqueue_many operation in a loop. In your training graph, you can use q.dequeue_many() to get a batch of negative examples, and use them as the input to your training process.