The following is a excerpt that was made, as part of my final capstone project.
Introduction
The hardware and software section will be primarily exploring the two key parts in the development of neural networks. Currently the two competing software libraries for the development of neural networks are PyTorch and Tensor Flow. And the two competing hardware platforms to train models is between AMD and Nvidia [6]. In this section I will explore the benefits and disadvantages of each.
Deep Learning Software & Hardware Selection
When looking into developing our model I identified the 2 key choices, software selection and hardware selection. I identified framework selection as a key choice since, it would act as the key building block in constructing the model, and effect how fast I could train them. Where as hardware selection was important since it would be the primary limiting factor in how fast I could train the model, and how complex I could make the model.
Software Selection
Due to the exponential expansion of machine learning (ML) research and computing power seen over the last decade. There has also been an explosion of new types of software infrastructure to harness it. This software has come from both academic and commercial sources. The need for this infrastructure arises from the fact that there needs to be a bridge betIen theory and application. When I looked at what Ire the most popular frameworks, I found it was a mix of strictly academic and commercial driven software. The four main frameworks Ire Caffe, Theano, Caffe2 + PyTorch, and Tensor Flow (TF).
When I went about choosing a framework, I considered three different factors, community, language, and performance. Community was one the biggest factors, since I had no real production experience in doing any sort of large scale ML modeling and deployment. The only framework that fulfilled this need was Google’s Tensor Flow. It had been released in 2015 and had been made available to the open source community. Leading to many academic researchers to contribute and influence its development. Which has resulted in many other companies using it in their production deep learning pipelines. The combination of both software developers and scientists using it has led to a lot of community driven development. This has lead to making it easier to use and deploy. A side effect of this large amount of adoption is the generation of detailed documentation. Written by the community, large amount of personal, and company blogs, detailing how they used TF to accomplish their goals. The only real competitor at the time of writing it this is Facebook’s Caffe 2 + PyTorch Libraries which was just open sourced early this year.
The other factor was the language interface it would use. I wanted an easy to use interface, with which to build out the model. When I looked at what was available, I found that all of the popular frameworks were written in C++ and CUDA, but had a easy to use Python based interface. The only framework out of the four mentioned above, that only had C++ based interface was Caffe.
The most important part of framework selection was the performance aspect. Most if not all ML research and production use cases happen on Nvidia GPU hardware. This is due to Nvidia’s development of their CUDA programming framework for use with their GPUs. It makes parallel programming for their GPUs incredibly easy. This parallelization is what lets the complex matrix operations be computed with incredible speed. There were only two frameworks out of the four I mentioned, that used the latest version of CUDA in its code base. Which were TF and Caffe 2 + PyTorch, however Caffe 2 + PyTorch was not as robust as Tensor Flow in supporting the different versions of CUDA.
In the end I choose to go with TF since it had a better community and CUDA support. I did not choose to go with its nearest competitor, since it was not as well documented, and its community was just starting to grow. Whereas TF has been thoroughly documented and has had large deployments outside of Google (such as at places like LinkedIn, Intel, IBM, and UBER). Another major selling point for TF is the fact that, it is free, continually getting new releases, and has become an industry standard tool.
| Deep Learning Software Frame Works | ||||
| Name | Caffe | Theano | Caffe 2 + PyTorch | Tensor Flow |
| Computational Graph Representation | No | Yes | Yes | Yes |
| Release Date | 2013 | 2009 | 2017 + 2016 | 2015 |
| Implementation language | C++ | Python & C | C++ | C++, JS, Swift |
| Wrapper languages | N/A | Python | Python, C++ | C, C++, Java, GO, Rust, Haskell, C#, Python |
| Mobile Enabled | NO | NO | YES | YES |
| Corporate Backing | UC Berkeley | University of Montreal | ||
| CUDA enabled | NO | YES | YES | YES |
| Multi GPU Support | NO | NO | YES | YES |
| Exportable Model | YES | NO | YES & NO | YES |
| Library of pretrained models | YES | NO | YES | YES |
| Unique Features | Don’t need to code to define Network | First to use CUDA and Computational Graph in Memory | Uses the original developers of Caffe and Theano frameworks VISDOM – Error function Visualization Tool PoIrs Facebook ML | Tensor Board – Network Visualization and Optimization Tool Developed by Google Deep Brain PoIrs Google ML |
| Under Active Development | No | No | Yes | Yes |
| NOTE | The reason as to why PyTorch and Caffe 2 are always mentioned together is because they are meant to be used together. PyTorch is much more focused on research and flexibility. Where as Caffe 2 is more focused on production deployment and inference speed. Facebook’s researchers use PyTorch to prototype models, then translate the model into Caffe 2, using their model transfer tool known as ONIX. | |||
Table 1 A summary of all information of note that I collected during my research