Deep Learning Framework Power Scores 2018

I just published the “Deep Learning Framework Power Scores 2018: Which deep learning frameworks are tops in usage, interest, and popularity?” in Towards Data Science.

I’ve been studying deep learning frameworks and was trying to decide which one to learn next. I couldn’t find a good analysis of which ones were most used and most popular, so I gathered a bunch of data and made my own scoring system.

The FastAI framework is new and not very in demand, but I included it because I think it’s likely to become much more popular.

What do you think? I’d appreciate any feedback, especially on the scoring metrics.

Due to popular demand, I added more frameworks. And due to François Chollet’s suggestions I improved several of the metrics.
Would love to hear your thoughts!

That is a great graph. What is your own oppinion what it will look like after 5 years. Do you think there will be pytorch at the place of Tensorflow and second place? I started learning Tensorflow instead of because there was not any good documentations. I think if Jeremy or someone else could make a documentation the library could get more attention.

The release of fastai v1 will be on Oct 2nd, and it will include documentation :slight_smile:


I will add that date to my calendar.

Excited for v1! That’s ahead of schedule. Awesome!

I think PyTorch and FastAI will grow in usage and popularity. It sounds like Keras become more tightly linked with TensorFlow. That should help keep TensorFlow popular.

I’m currently using Tensorflow just because there is more examples and tutorials but I think could start using and PyTorch after the v1 release. What you Jeffhale think about the growth speed? At least what I have heard the PyTorch is growing very fast and I think growing speed is one of the most important indicators when we try to predict which library is popular in five years from now. Is Tensorflow growing as rapidly and is there some groups which use tf and other groups which use pt?

Today is the day!

1 Like

Hi on the past I have experimented Tensorflow but as the time pased each time I visited the documentation many things was changing . Also the developer way of doing things on Tensorflow (create the graph then execute in a session) I don’t liked, PyTorch have a different learning curve than Tensorflow(very short) that in few lessons you can… I have the audacious to say … in ONE LESSON … you can jump in for doing things. I find out to be much more natural at least for me. I hope you enjoy it too. There are dozend of examples on the youtube channels:

Take a look on this:

Thanks for reply! It sounds like that Google can mess up Tensorflow like they did with AngularJS which is not anymore very popular. I have learned to use TF but I cant say that I’m really good with it. I think only reason why I would still use Tensorflow is tensorboard but I have read that some people have made it to work with pytorch and I think they will add this kind of thing some day to the library.

Hi, you totally right about Angular thing. I saw people who loves Javascript deciding that wouldn’t learn angular 2+ if they need to learn Typescript to do it.

I also saw that library that interfaces Tensorboard with pytorch. Appears to be very nice, but I still don’t needed to use it. I really don’t know how to use it or why use it. Because in Pytorch world you build what you think in terms of computational graph you don’t need a Ploting graph to know how is your model. To me appear much more like a cosmetic to impress some one instead of I really need it. Like I said, I don’t see any needs to use that so … to me still useless.