Development

Which Deep Learning Framework Should You Choose?

Dr Justin Chan
Dr Justin Chan
April 17, 2019·7 min read
If you are into machine learning and Deep Learning, you’ll know how important Deep Learning Frameworks can be. They make the whole task of coding your models so much easier simply by abstracting out the details that you could do without. Problem is, there are so many frameworks out there it can be a challenge picking the right one for your project. Giants like Facebook, Google, Microsoft, and Amazon have all come up with their own deep learning frameworks, each of which is better than the others in some ways. Today, we will be counting down the top 5 most popular deep learning frameworks to use in 2019.
  1. Caffe2:
Caffe is a fairly light-weight and high-level framework. It is ideal if you want to create deep learning models that work with image data. However, it is not very helpful in working with Recurrent Neural Networks and language processing models. Here are some of its features:
  1.  Lightweight: Since it is not computationally too expensive, it is easy to deploy models created in Caffe to mobile devices or other computationally constrained platforms.
  2.  High-level API: Even if your knowledge in machine learning or the math behind it is limited, you can easily build deep learning models using Caffe. This helps beginners get a good feel of deep learning before actually jumping into more complex coding.
  3.  Caffe Model Zoo: This framework gives access to a plethora of pre-trained networks, models and weights, ready for you to apply to your deep-learning problems.
Deep Learning
  1. Deeplearning4j:
Deeplearning4j is Java’s reply to Python frameworks. In fact, since it has been implemented in Java, it is actually more efficient than Python-based frameworks. Some of its Deeplearning4j’s features have been highlighted below:
  1. Supports Tensors: DeepLearning4j uses the tensor library called ND4J. This allows java deep-learning programmers to work tensors.
  2. Hardware support: The framework supports both CPUs and GPUs.
  3.  Organized structure: Loading data and training algorithms are treated as separate processes, so the process of deep learning is more organized.
  4. Models supported:  The framework allows the creation of models like ANNs, CNNs and even RNNs and LSTMs.
  1. Keras:
The Keras framework is the ideal starting framework if you are just beginning to learn deep learning and are comfortable with Python coding. It is a high-level API, which was developed to facilitate fast experimentation. Here are some of its key features:
  1. Compatibility with lower-level frameworks: Keras can be wrapped over other deep learning frameworks like Tensorflow, CNTK, and Theano. In this way, new users can use the higher-level interface of Keras with the lower-level features of these other frameworks working under the hood.
  2. Models supported: Keras supports models like CNNs and RNNs, which a lot of other frameworks don’t. With a high-level interface, it allows beginners to get acquainted with these complex network models without feeling too intimidated by the complexity.
  3. Hardware support: Keras can run seamlessly on both the CPU and GPUs.
  1.     Pytorch:
Although fairly new, Pytorch is gaining quite a lot of popularity due to its ease of use and efficiency. Created by Facebook Artificial Intelligence Research team (FAIR), Pytorch is already competing for neck-to-neck with Tensorflow. The framework provides a fairly easy to use API and coding with it feels more Pythonic, so you feel quite at home. Deep Learning Some of the features that make Pytorch stand out include:
  • Dynamic computation graphs: Pytorch allows you to create graphs dynamically as you go along coding.
  • CUDA support: Support for CUDA ensures that the code can be run on GPU. This speeds up running of your code and performance of your system.
  1.     Tensorflow 2.0:
Tensorflow is the brainchild of the makers of Google. It is hailed today as the best framework by most programmers. And there are multiple reasons for that:
  1. Organized coding: Its system of using datasets, layers and feature columns, though not exactly intuitive feeling at first, can help you organize your code better.
  2. Massive developer community: Tensorflow has rich, detailed documentation available online. You can find enough learning content on both its website as well as from other third-party sources all over the web.
  3.  Visualization using Tensorboard: One of Tensorflow’s highlight features is the Tensorboard. This is a great interface that lets u monitor the model training process through different visualizations. With Tensorflow 2.0 you can even see how well your model is performing and identify bottlenecks.
  4. Tensorflow serving: This lets you serve your models on a specialized gRPC server. So you can deploy your models in a production environment, with distributed training.
  5. TfLite: Lets you conveniently transfer your model to mobile devices, so you can easily make quick inferences on your cell phones or tablets.
In spite of the high popularity of Tensorflow, a very large portion of the developers have started to take a liking to Pytorch. What ensued is an Internet-wide debate between the supporters of the two. Which One is better? Pytorch or Tensorflow? Let’s do a comparative study to see who wins.
  1. Ease of learning and coding: Pytorch feels more ‘Pythonic’ and has an object-oriented approach. It is speedy and efficient and to code using Pytorch you don’t need to learn any extra concepts. Tensorflow, however, requires you to understand concepts of Variable scoping, placeholders and sessions and you need to do more boilerplate coding. This is not very likable by programmers.
  2. Documentation: There is a lot of documentation available online for both frameworks. However, since Pytorch is fairly newer, the volume may be slightly less than that of Tensorflow, but that is expected to level up quite quickly. With Tensorflow 2.0, a lot of the coding style has changed, so the Tensorflow community has to work towards building a good documentation base with the new style too.
  3. Graph construction: Pytorch supports dynamic computation graphs and Imperative programming instead of symbolic programming. This allows for more flexible coding. Although Tensorflow was 1.x was lacking in this area owing to its define-and-run strategy, Tensorflow 2.0 made up for that with Eager execution. Thanks to eager execution, graph creation using Tensorflow has also become flexible.
  4. Debugging:  Pytorch uses the pdb (python standard debugger) for debugging. You can also debug using Pycharm, so that makes code written Pytorch easier to debug. and you don’t need to learn it separately. Tensorflow has its separate debugger (tfdbg), so that’s one extra thing to learn. However, Tensorflow 2.0 is now integrated with Keras, so you can use tf.keras for easier and more flexible debugging.
  5. Distributed Training: You can use Tensorflow to easily distribute your training load over multiple devices like GPUs. This is something that Pytorch still does not support.
  6. Serialization: Sometimes you may need to deploy your models to devices in use-cases where Python is not an option. With Tensorflow, you can save the entire graph as a protocol buffer along with parameters and operations. This then facilitates the transfer of the graph to other languages like C++ or Java. This is one feature that is not supported by Pytorch.
  7. Deployment: Both are easy to wrap for small scale server deployments, for example, if you want to use your model on a flask web server. For mobile and embedded system development, however, Tensorflow works better than other frameworks including Pytorch. Moreover, you can use Tensorflow serving for easy server-side deployments, something that Pytorch lacks.
Our Verdict Pytorch can be ideal in research applications, or small scale projects, where rapid prototyping is more important. However, for large-scale deployments, Tensorflow is more feasible. It is also useful when you need to deploy your model to other platforms or use it in mobile or embedded systems. Tensorflow’s define-and-run strategy of using static graphs had put it at a disadvantage from Pytorch’s dynamic run-as-you-go strategy, but the alpha version of Tensorflow 2.0 is certainly changing the game.
🎬DDI MediaNewDDI Smart Block

Turn this article into a video

Instantly repurpose any DDI article into a professionally produced short-form video.

Try DDI Media →
Dr Justin Chan
Dr Justin Chan

Dr Chan founded DataDrivenInvestor.com (DDI) and is the CEO for JCube Capital Partners. Specialized in strategy development, alternative data analytics and behavioral finance, Dr Chan also has extensive experience in investment management and financial services industries. Prior to forming JCube and DDI, Dr Chan served in the capacity of strategy development in multiple hedge funds, fintech companies, and also served as a senior quantitative strategist at GMO. A published author at professional journals in finance, Dr. Chan holds a Ph.D. degree in finance from UCLA.

More in Development

Development

Personal Lessons from Startups - Part 4 of 5

Sonal Panda·6 min
Development

The Archaic Life of Fishermen Confronts the Progressives of the Windmill Ecosystem

Susan Mackenzie Andersen·16 min
Development

Development Process: From Product Concept to Manufacturing

Shachar Oz·16 min