Connect with us


Machine learning – Tensorflow and Pytorch



Spread Articles to Your Friends

TensorFlow is a free and open-source software library for dataflow and differentiable programming covering a range of jobs. It is a symbolic math library and is also used for machine learning applications such as neural networks. PyTorch is an open-source machine learning library based on the Torch library, which can be applicable for applications such as computer vision and natural language processing.

The two machine-learning frameworks Tensorflow and Pytorch compete closely for dominance in deep learning. Both have a strong base. In the market of machine-learning frameworks, the usage situation is getting tighter on Tensorflow and Pytorch. Other frameworks are virtually irrelevant. According to research by the KI magazine The Gradient, Facebook’s machine-learning framework Pytorch is becoming increasingly effective against Google’s Tensorflow – as far as its use in research is concerned. In order to make that determination, The Gradient has looked at past AI conferences and their submissions.

There was a strong increase in submissions based on Pytorch versus a declining number of Tensorflow papers. The absolute majority of submissions at all major conferences are now based on Pytorch, according to The Gradient. Thus, the sheet has turned completely in the course of only one year. The Pytorch dominance should not only be seen at special conferences, for example on image and language processing, but also in all other topics of machine-learning implementation. Overall, not only does the increase in Pytorch submissions lead to a majority over Tensorflow, but even the dedicated Tensorflow submissions have fallen in absolute terms compared to last year.

Beyond the use in the research dominates according to the findings of the gradient still Google’s Tensorflow in machine learning. When companies do Big Data training, Google’s frameworks are typically used. This has a lot to do with the flexibility of the implementation and, above all, the performance of the framework. Here, the different requirement profile between research and industry is particularly clear. While it is far more important to a researcher to be able to quickly change and recalculate his machine-learning model, the main focus in production would be on the quick completion of once defined processes. Pytorch is particularly popular among researchers because of its fast implementations. So it was easier to handle and easier to manipulate in operation. In addition, the learning effort is lower until the first start of a project. In research, machine learning is also typically tested on local computers or small networks. In industry, however, these factors played virtually no role. Here, other specific benefits would be much more important. For example, Tensorflow is optimized for Google Cloud and can offer significant performance benefits. This is exactly what scares researchers from Tensorflow because they did not want Google to keep the entire vertical process of machine learning in their own hands. In addition, the willingness of other competitors such as Microsoft, Amazon or Nvidia promotes the only remaining alternative, just to support Pytorch.

Regardless of the current outcome, the question of the dominance of Tensorflow or Pytorch remains meaningless in the medium term. Because the whole complex of machine learning moves forward at such a high speed that it can not be estimated whether current trends will stabilize or evaporate. What is clear is that Machine Learning will look drastically different in five years from today.

Continue Reading