TensorFlow vs Theano vs Torch vs Keras: Deep Learning Library
Daniel Johnson
Artificial intelligence is growing in popularity since 2016 with, 20% of the big companies using AI in their businesses (McKinsey report, 2018). As per the same report AI can create substantial value across industries. In banking, for instance, the potential of AI is estimated at $300 billion, in retail the number skyrocket to $600 billion.
To unlock the potential value of AI, companies must choose the right deep learning framework. In this tutorial, you will learn about the different libraries available to carry out deep learning tasks. Some libraries have been around for years while new library like TensorFlow has come to light in recent years.
8 Best Deep learning Libraries /Framework
In this list, we will compare the top Deep learning frameworks. All of them are open source and popular in the data scientist community. We will also compare popular ML as a service providers
Torch
Torch is an old open source machine learning library. It was first released was 15 years ago. It is primary programming languages is LUA, but has an implementation in C. Comparing PyTorch vs TensorFlow, it supports a vast library for machine learning algorithms, including deep learning. It supports CUDA implementation for parallel computation.
Torch deep learning tool is used by most of the leading labs such as Facebook, Google, Twitter, Nvidia, and so on. Torch has a library in Python names Pytorch.
Infer.net
Infer.net is developed and maintained by Microsoft. Infer.net is a library with a primary focus on the Bayesian statistic. Infer.net is a visualization tool for Deep Learning designed to offer practitioners state-of-the-art algorithms for probabilistic modeling. The library contains analytical tools such as Bayesian analysis, hidden Markov chain, clustering.
Keras
Keras is a Python framework for deep learning. It is a convenient library to construct any deep learning algorithm. The advantage of Keras is that it uses the same Python code to run on CPU or GPU. Besides, the coding environment is pure and allows for training state-of-the-art algorithm for computer vision, text recognition among other.
Keras was developed by François Chollet, a researcher at Google. Keras is used in prominent organizations like CERN, Yelp, Square or Google, Netflix, and Uber.
Theano
Theano is deep learning library developed by the Université de Montréal in 2007. Comparing Theano vs TensorFlow, it offers fast computation and can be run on both CPU and GPU. Theano has been developed to train deep neural network algorithms.
Microsoft Cognitive Toolkit (CNTK)
Microsoft toolkit, previously known as CNTK, is a deep learning library developed by Microsoft. According to Microsoft, the library is among the fastest on the market. Microsoft toolkit is an open-source library, although Microsoft is using it extensively for its product like Skype, Cortana, Bing, and Xbox. The toolkit is available both in Python and C++.
MXNet
MXnet is a recent deep learning library. It is accessible with multiple programming languages including C++, Julia, Python and R. MXNet can be configured to work on both CPU and GPU. MXNet includes state-of-the-art deep learning architecture such as Convolutional Neural Network and Long Short-Term Memory. MXNet is build to work in harmony with dynamic cloud infrastructure. The main user of MXNet is Amazon
Caffe
Caffe is a library built by Yangqing Jia when he was a PhD student at Berkeley. Comparing Caffe vs TensorFlow, Caffe is written in C++ and can perform computation on both CPU and GPU. The primary uses of Caffe is Convolutional Neural Network. Although, In 2017, Facebook extended Caffe with more deep learning architecture, including Recurrent Neural Network. Caffe is used by academics and startups but also some large companies like Yahoo!.
TensorFlow
TensorFlow is Google open source project. TensorFlow is the most famous deep learning library these days. It was released to the public in late 2015
TensorFlow is developed in C++ and has convenient Python API, although C++ APIs are also available. Prominent companies like Airbus, Google, IBM and so on are using TensorFlow to produce deep learning algorithms.
TensorFlow Vs Theano Vs Torch Vs Keras Vs infer.net Vs CNTK Vs MXNet Vs Caffe: Key Differences
| Library | Platform | Written in | Cuda support | Parallel Execution | Has trained models | RNN | CNN |
|---|---|---|---|---|---|---|---|
| Torch | Linux, macOS, Windows | Lua | Yes | Yes | Yes | Yes | Yes |
| Infer.Net | Linux, macOS, Windows | Visual Studio | No | No | No | No | No |
| Keras | Linux, macOS, Windows | Python | Yes | Yes | Yes | Yes | Yes |
| Theano | Cross-platform | Python | Yes | Yes | Yes | Yes | Yes |
| TensorFlow | Linux, macOS, Windows, Android | C++, Python, CUDA | Yes | Yes | Yes | Yes | Yes |
| MICROSOFT COGNITIVE TOOLKIT | Linux, Windows, Mac with Docker | C++ | Yes | Yes | Yes | Yes | Yes |
| Caffe | Linux, macOS, Windows | C++ | Yes | Yes | Yes | Yes | Yes |
| MXNet | Linux, Windows, macOS, Android, iOS, Javascript | C++ | Yes | Yes | Yes | Yes | Yes |
Verdict
TensorFlow is the best library of all because it is built to be accessible for everyone. Tensorflow library incorporates different APIs to built at scale deep learning architecture like CNN or RNN. TensorFlow is based on graph computation, it allows the developer to visualize the construction of the neural network with Tensorboad. This tool is helpful to debug the program. Finally, Tensorflow is built to be deployed at scale. It runs on CPU and GPU.
Tensorflow attracts the largest popularity on GitHub compare to the other deep learning libraries.
Comparing Machine Learning as a Service
Following are 4 popular DL as a service providers
Google Cloud ML
Google provides for developer pre-trained model available in Cloud AutoML. This solution exists for a developer without a strong background in machine learning. Developers can use state-of-the-art Google’s pre-trained model on their data. It allows any developers to train and evaluate any model in just a few minutes.
Google currently provides a REST API for computer vision, speech recognition, translation, and NLP.
Using Google Cloud, you can train a machine learning framework build on TensorFlow, Scikit-learn, XGBoost or Keras. Google Cloud machine learning will train the models across its cloud.
The advantage to use Google cloud computing is the simplicity to deploy machine learning into production. There is no need to set up Docker container. Besides, the cloud takes care of the infrastructure. It knows how to allocate resources with CPUs, GPUs, and TPUs. It makes the training faster with paralleled computation.
AWS SageMaker
A major competitor to Google Cloud is Amazon cloud, AWS. Amazon has developed Amazon SageMaker to allow data scientists and developers to build, train and bring into production any machine learning models.
SageMaker is available in a Jupyter Notebook and includes the most used machine learning library, TensorFlow, MXNet, Scikit-learn amongst others. Programs written with SageMaker are automatically run in the Docker containers. Amazon handles the resource allocation to optimize the training and deployment.
Amazon provides API to the developers in order to add intelligence to their applications. In some occasion, there is no need to reinventing-the-wheel by building from scratch new models while there are powerful pre-trained models in the cloud. Amazon provides API services for computer vision, conversational chatbots and language services:
The three major available APIs are:
- Amazon Rekognition: provides image and video recognition to an app
- Amazon Comprehend: Perform text mining and neural language processing to, for instance, automatize the process of checking the legality of financial document
- Amazon Lex: Add chatbot to an app
Azure Machine Learning Studio
Probably one of the friendliest approaches to machine learning is Azure Machine Learning Studio. The significant advantage of this solution is that no prior programming knowledge is required.
Microsoft Azure Machine Learning Studio is a drag-and-drop collaborative tool to create, train, evaluate and deploy machine learning solution. The model can be efficiently deployed as web services and used in several apps like Excel.
Azure Machine learning interface is interactive, allowing the user to build a model just by dragging and dropping elements quickly.
When the model is ready, the developer can save it and push it to Azure Gallery or Azure Marketplace.
Azure Machine learning can be integrated into R or Python their custom built-in package.
IBM Watson ML
Watson studio can simplify the data projects with a streamlined process that allows extracting value and insights from the data to help the business to get smarter and faster. Watson studio delivers an easy-to-use collaborative data science and machine learning environment for building and training models, preparing and analyzing data, and sharing insights all in one place. Watson Studio is easy to use with a drag-and-drop code.
Watson studio supports some of the most popular frameworks like Tensorflow, Keras, Pytorch, Caffe and can deploy a deep learning algorithm on to the latest GPUs from Nvidia to help accelerate modeling.
Verdict
In our point of view, Google cloud solution is the one that is the most recommended. Google cloud solution provides lower prices the AWS by at least 30% for data storage and machine learning solution. Google is doing an excellent job to democratize AI. It has developed an open source language, TensorFlow, optimized data warehouse connection, provides tremendous tools from data visualization, data analysis to machine learning. Besides, Google Console is ergonomic and much more comprehensive than AWS or Windows.