How does deep learning work in real time? - api

I just have made api which recommends movies based on user input using Django. It’s designed as to execute deep learning function for every 10 inputs from each user. For this each 10 inputs, it takes around 10 seconds to give output. How does google or amazon can do real time data update without delay?

When it comes to execution there is no deep learning, there is just deep model. Let us assume that this is a regular feed forward model, then all you have to do is perform K (depth) matrix multiplications. For convolutional layers it is much more, but still matrix multiplications. All these operations are extremely simple to parallelize. In particular just running them on GPU will give you ~20x boost. Using tensorflow which can scatter computations across many cores/cpus/machines also works the same way. There are also numerous other optimizations possible, like training a small net to replicate behaviour of the big one, making use of sparsity of some matrices (if you are using relus - many neurons will produce zeros, thus - sparse matrices appear) etc.
That being said - network that takes 10s to process 10 inputs sounds like a horrible implementation, or the network is really huge, so before going for general optimization schemes - make sure your current code is fine. For example if you use tensorflow etc. it is important that many things before pushing your data take time - like loading libraries, starting sessions, session .run call itself etc.

Related

Parallelization strategies for deep learning

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.

How to make a model of 10000 Unique items using tensorflow? Will it scale?

I have a use case where I have around 100 images each of 10000 unique items. I have 10 items with me which are all from the 10000 set and I know which 10 items too but only at the time of testing on live data. I have to now match the 10 items with their names. What would be an efficient way to recognise these items? I have full control of training environment background and the testing environment background. If I make one model of all 10000 items, will it scale? Or should I make 10000 different models and run the 10 items on the 10 models I have pretrained.
Your question is regarding something called "one-vs-all classification" you can do a google search for that, the first hit is a video lecture by Andrew Ng that's almost certainly worth watching.
The question has been long studied and in a plethora of contexts. The answer to your question does very much depend on what model you use. But I'll assume that, if you're doing image classification, you are using convolutional neural networks, because, after all, they're state of the art for most such image classification tasks.
In the context of convolutional networks, there is something called "Multi task learning" that you should read up on. Boiled down to a single sentence, the concept is that the more you ask the network to learn the better it is at the individual tasks. So, in this case, you're almost certain to perform better training 1 model on 10,000 classes than 10,000 classes each performing a one-vs-all classification scheme.
Take for example the 1,000 class Imagenet dataset, and CIFAR-10's 10 class dataset. It has been demonstrated in numerous papers that first training against Imagenet's 1,000 class dataset, and then simply replacing the last layer with a 10 class output and re-training on CIFAR-10's dataset will produce a better result than just training on CIFAR-10's dataset alone. There are admittedly multiple reasons for this result, Imagenet is a larger dataset. But the richness of class labels, multi-task learning, in the Imagenet dataset is certainly among the reasons for this result.
So that was a long winded way of saying, use one model with 10,000 classes.
An aside:
If you want to get really, really interesting, and jump into the realm of research level thinking, you might consider a 1-hot vector of 10,000 classes rather sparse and start thinking about whether you could reduce the dimensionality of your output layer using an embedding. An embedding would be a dense vector, let's say size 100 as a good starting point. Now class labels turn into clusters of points in your 100 dimensional space. I bet your network will perform even better under these conditions.
If this little aside didn't make sense, it's completely safe to ignore it, your 10,000 class output is fine. But if it did peek your interest look up information on Word2Vec, and read this really nice post on how face recognition is achieved using embeddings: https://medium.com/#ageitgey/machine-learning-is-fun-part-4-modern-face-recognition-with-deep-learning-c3cffc121d78. You might also consider using an Auto Encoder to generate an embedding for the images (though I favor triplet embeddings as typically used in face recognition myself).

Is it possible to train Neural Network with low amount of instances?

I have faced some problem when I needed to solve Regression Task and use as minimum instances as possible. When I tried to use Xgboost I had to feed 4 instances to get the reasonable result. But Multilayer Perceptron tuned to overcoming Regression problems has to take 20 instances, tried to change amount of neurons&layers but the answer is still 20 .Is it possible to do something to make Neural Network solve Resgression tasks with from 2 to 4 instances? if yes - explain please what should I do to succeed in it? Maybe there is some correlation between how much instances are needed to train and get reasonable results from Perceptron and how features are valuable inside dataset?
Thanks in advance for any help
With small numbers of samples, there are likely better methods to apply, Xgaboost definitely comes to mind as a method that does quite well at avoiding overfitting.
Neural networks tend to work well with larger numbers of samples. They often over fit to small datasets and underperform other algorithms.
There is, however, an active area of research in semi-supervised techniques using neural networks with large datasets of unlabeled data and small datasets of labeled samples.
Here's a paper to start you down that path, search on 'semi supervised learning'.
http://vdel.me.cmu.edu/publications/2011cgev/paper.pdf
Another area of interest to reduce overfitting in smaller datasets is in multi-task learning.
http://ruder.io/multi-task/
Multi task learning requires the network to achieve multiple target goals for a given input. Adding more requirements tends to reduce the space of solutions that the network can converge on and often achieves better results because of it. To say that differently: when multiple objectives are defined, the parameters necessary to do well at one task are often beneficial for the other task and vice versa.
Lastly, another area of open research is GANs and how they might be used in semi-supervised learning. No papers pop to the forefront of my mind on the subject just now, so I'll leave this mention as a footnote.

Cells detection using deep learning techniques

I have to analyse some images of drops, taken using a microscope, which may contain some cell. What would be the best thing to do in order to do it?
Every acquisition of images returns around a thousand pictures: every picture contains a drop and I have to determine whether the drop has a cell inside or not. Every acquisition dataset presents with a very different contrast and brightness, and the shape of the cells is slightly different on every setup due to micro variations on the focus of the microscope.
I have tried to create a classification model following the guide "TensorFlow for poets", defining two classes: empty drops and drops containing a cell. Unfortunately the result wasn't successful.
I have also tried to label the cells and giving to an object detection algorithm using DIGITS 5, but it does not detect anything.
I was wondering if these algorithms are designed to recognise more complex object or if I have done something wrong during the setup. Any solution or hint would be helpful!
Thank you!
This is a collage of drops from different samples: the cells are a bit different from every acquisition, due to the different setup and ambient lights
This kind of problem should definitely be possible. I would suggest starting with a cifar 10 convolutional neural network tutorial and customizing it for your problem.
In future posts you should tell us how your training is progressing. Make sure you're outputting the following information every few steps (maybe every 10-100 steps):
Loss/cost function output, you should see your loss decreasing over time.
Classification accuracy on the current batch of your training data
Classification accuracy on a held out test set (if you've implemented test set evaluation, you might implement this second)
There are many, many, many things that can go wrong, from bad learning rates, to preprocessing steps that go awry. Neural networks are very hard to debug, they are very resilient to bugs, making it hard to even know if you have a bug in your code. For that reason make sure you're visualizing everything.
Another very important step to follow is to save the images exactly as you are passing them to tensorflow. You will have them in a matrix form, you can save that matrix form as an image. Do that immediately before you pass the data to tensorflow. Make sure you are giving the network what you expect it to receive. I can't tell you how many times I and others I know have passed garbage into the network unknowingly, assume the worst and prove yourself wrong!
Your next post should look something like this:
I'm training a convolutional neural network in tensorflow
My loss function (sigmoid cross entropy) is decreasing consistently (show us a picture!)
My input images look like this (show us a picture of what you ACTUALLY FEED to the network)
My learning rate and other parameters are A, B, and C
I preprocessed the data by doing M and N
The accuracy the network achieves on training data (and/or test data) is Y
In answering those questions you're likely to solve 10 problems along the way, and we'll help you find the 11th and, with some luck, last one. :)
Good luck!

Must each tensorflow batch contain a uniform distribution of the inputs for all expected classifications?

This is probably a newbie question but I'm trying to get my head around how training on small batches works.
Scenario -
For the mnist classification problem, let's say that we have a model with appropriate hyerparameters that allow training on 0-9 digits. If we feed it with a small batches of uniform distribution of inputs (that have more or less same numbers of all digits in each batch), it'll learn to classify as expected.
Now, imagine that instead of a uniform distribution, we trained the model on images containing only 1s so that the weights are adjusted until it works perfectly for 1s. And then we start training on images that contain only 2s. Note that only the inputs have changed, the model and everything else has stayed the same.
Question -
What does the training exclusively on 2s after the model was already trained exclusively on 1s do? Will it keep adjusting the weights till it has forgotten (so to say) all about 1s and is now classifying on 2s? Or will it still adjust the weights in a way that it remembers both 1s and 2s?
In other words, must each batch contain a uniform distribution of different classifications? Does retraining a trained model in Tensorflow overwrite previous trainings? If yes, if it is not possible to create small (< 256) batches that are sufficiently uniform, does it make sense to train on very large (>= 500-2000) batch sizes?
That is a good question without a clear answer. In general, the order and selection of training samples has a large impact on the performance of the trained net, in particular in respect to the generalization properties it shows.
The impact is so strong, actually, that selecting specific examples, and ordering them in a particular way to maximize performance of the net even constitutes a genuine research area called `curriculum learning'. See this research paper.
So back to your specific question: You should try different possibilities and evaluate each of them (which might actually be an interesting learning exercise anyways). I would expect uniformly distributed samples to generalize well over different categories; samples drawn from the original distribution to achieve the highest overall score (since, if you have 90% samples from one category A, getting 70% over all categories will perform worse than having 99% from category A and 0% everywhere else, in terms of total accuracy); other sample selection mechanisms will show different behavior.
An interesting reading about such questions is Bengio's 2012 paper Practical Recommendations for Gradient-Based Training of Deep
Architectures
There is a section about online learning where the distribution of training data is unknown. I quote from the original paper
It
means that online learners, when given a stream of
non-repetitive training data, really optimize (maybe
not in the optimal way, i.e., using a first-order gradient
technique) what we really care about: generalization
error.
The best practice though to figure out how your dataset behaves under different testing scenarios would be to try them both and get experimental results of how the distribution of the training data affects your generalization error.