Fastai for PyTorch: Fast and accurate neural nets using modern best practices | Hacker News Reader

Fastai for PyTorch: Fast and accurate neural nets using modern best practices

fast.ai

279 pointsstablemap55 comments

Jeremy from fast.ai here - I'm sitting here at the PyTorch developer conference, listening to all the really great new stuff being announced. :) Happy to answer any questions about the fastai library release here. It's a rewrite from scratch compared to v0.7.

Docs here: http://docs.fast.ai . GitHub repo here: https://github.com/fastai/fastai . It's also available now on Google Cloud Platform, including example notebooks and datasets (Viacheslav Kovalevskyi from Google has posted a walk-thru here: https://blog.kovalevskyi.com/google-compute-engine-now-has-i... ). AWS support coming soon.

Thanks for all your hard work Jeremy. In Cleveland, we are running a study group with about 30 people on the 2018 version of the course. Because of logistics, we aren't planning on rescheduling to match the new course, but its still nice knowing that the library is being invested into.

Hello and thank you for the great work! Hopefully, the new version will work on MacOS with Anaconda, because 0.7 failed there due to the unsatisfied dependency on PIL. Apparently, the default Anaconda channels doesn't include the python 3.x-compiliant PIL package for Darwin. That blocked my progress on ML course.

You can always get a server with a GPU instance in the cloud somewhere. That might cost a few $ but then again, so does a GPU.

MacOS is really hard to work with for deep learning. Try https://salamander.ai/ . $0.19/hour for a server with NVIDIA GPU, fastai v1 pre-installed with notebooks and datasets. Or try Google Cloud preemptible instances - similar cost IIRC, but a little more setup to do.

Oh, it's $0.36/hour for salamander. the equivalent on gcp (n1-16-standard with attached k80) is $0.29/hour. though as you said, Salamander is easier to get started with (just 1 click after signing up!)

ps: i'm the creator of salamander

Are you sure you didn’t want pillow, which provides the PIL import and is the follow-on from the original PIL?

I was just listening to the the podcast when I saw this post. It's great to hear you can make a lot of progress on ML without a PHD. :)

What podcast?

I assume they are referring to the TWiMLAI podcast that was posted this morning https://twimlai.com/twiml-talk-186-the-fastai-v1-deep-learni...

When someone has a very high IQ, like Jeremy Howard obviously does, PhD becomes just a superfluous credential.

Great - can't wait to try out the new version! I am running fast.ai in a virtual environment. How does one upgrade to the latest version for both Fast.ai and PyTorch. Are the new versions backwards comaptible with the earlier version - would it break any code? If not, would you recommend I clone the current install and upgrade that to be on the safe side. Thanks!

None of fastai is backwards compatible with v0.7 - although the API has a similar structure, so it shouldn't be hard to rewrite. PyTorch v1.0 also has some big differences to the version we used in the course.

I suggest you create a new conda environment and then follow the instructions in the readme to install both.

Blog announcement here: https://cloud.google.com/blog/products/ai-machine-learning/i...

For a little latitude, I believe there are PyTorch VMs available on the other cloud providers as well.

Hey Jeremy, just wanted to say thanks very much for making the course and the library, both have been invaluable for us in getting started with deep learning (and achieving useful results) quickly. Both parts of the 2018 course have been a really great tour through the different areas of deep learning, and provided good jumping off points for digging deeper on the parts we got interested in.

I have a gaming system (Ryzen 1950X w/ GTX 1080Ti). Is there an easy way to create a bootable external media with most of the stuff preconfigured already?

I remember it being a pain to get all the drivers working and there was a lot of version incompatibility in the Python ecosystem.

I'd rather not mess with the bootloader and just have an external drive I can select to boot to in the BIOS. For some reason Windows wiped out GRUB and I can't boot my old Ubuntu distro anymore.

The story on python is much better nowadays. You can install cuda and drivers from nvidia, and then install anaconda, then just run:

    conda install -c pytorch -c fastai fastai pytorch-nightly cuda92

If you don't use conda, things can get more complicated, depending on what libs you have installed.

You have a couple of options to consider: 1. Install a modern Linux distribution on the bare metal, get https://github.com/NVIDIA/nvidia-docker and run the GPU-accelerated container https://github.com/Paperspace/fastai-docker 2. Try running the Ubuntu Hyper-V image provided by Canonical/Microsoft

P.S. I'm considering building a DL box based on Ryzen 2600 w/ GTX 1070Ti, so could you please share your experience with Ryzen and Linux.

My DL box has a Ryzen 1800x and a GTX 1080ti running Ubuntu 18.04. The Ryzen has given me very few issues, Nvidia gave me some setup issues, but now that it's all configured its rock solid.

A Hyper-V image would be great if it can take advantage of the GPU and not cause driver issues. Didn't know they had one for DL. Or is it just pure Ubuntu you mean?

Shameless plug, but another option for running your experiments on a GPU is Spell (https://spell.run). Very simple to use and will allow you to use as many GPUs as you want at once and only pay for the compute time. You can also run GPU-enabled kernels for Jupiter.

Fast.ai instructions here: https://learn.spell.run/fast_ai

Why do you hate variable names longer than 4 characters so much? :b

Hi Jeremy, not a question just a short thank you note. I think the value you produce with fast.ai is absolutely fantastic, and I really want to communicate my gratefulness for all of the excellent content and courses and blog posts. Especially having access to them for free. Personally, it has helped me to learn so much and to actually implement these techniques in a workplace setting. Again, thanks.

Regards

Hey Jeremy, will you update your courses to reflect the new API changes? I just started on your deep learning course, using pip to install the fastai library.

The live version of the course, updated for fastai v1, is beginning later this month, with publicly available videos to follow in early 2019.

Hi Jeremy!

What is the main difference with Keras+TF, I've only worked with them (still not mastering them), why should I consider fastai?

Thank you!

Possibly addressed in a passage from this post: http://www.fast.ai/2018/10/02/fastai-ai/

> It’s important to understand that these improved results over Keras in no way suggest that Keras isn’t an excellent piece of software. Quite the contrary! If you tried to complete this task with almost any other library, you would need to write far more code, and would be unlikely to see better speed or accuracy than Keras. That’s why we’re showing Keras in this comparison - because we’re admirers of it, and it’s the strongest benchmark we know of!

Hey Jeremy, are you going to be recommending GCP/AWS hosted fastai as the best way to get started now over Paperspace etc?

This is great news! Here has been my experience using Fast.ai : I had been training a deep learning network using Keras with Tensorflow, for diagnosing medical images - and it took me several months of hard work - tweaking parameters, training and testing to get acceptable levels of accuracy for our models. And then last month, I switched to Fast.ai (their pre-release version) and I was blown away - my models trained faster, and I matched and finally exceeded accuracy levels acheived with my earlier models. And I accomplished what had taken several months in Keras, in just a few days! And the biggest reasons for it were in my view, fast.ai's learning rate finder, the differential learning rates, and Test Time augmentation - all which are advanced features built into fast.ai. And the other great thing is that fast.ai uses the best defaults automatically, and it trains much, much faster than Keras / TF for some reason.

So I can't wait to try the new release out. I think Fast.ai has set a new bar for deep learning frameworks in terms of speed and ease of use. Thank you for all your great work!

Do you know if you can do object detection on images right out of the box? Or is it just classification?

I am using it for classification right now - but pretty sure you should be able to do object detection quite easily as well.

It supports segmentation right now, but object detection won't be supported "out of the box" until 1.1 (in a month or two).

You can certainly train a YOLO or SSD pytorch model with fastai, however.

Hello Jeremy,

One question, there is any best practice to transform video into [n] frames to then use model.predict(n) to make a "live classification / object detection"?

Kind regards from Dublin!

There is a learning rate finder for Keras:

https://github.com/surmenok/keras_lr_finder

differential learning rates are now called discriminative learning rates.

I haven't used fast.ai yet, but extensively use pytorch for anything to do with deeplearning. (it is so much better than tensorflow) It's great to see a keras like higher level abstraction library for deep learning.

A few questions:

1. what is the benefit of using fast.ai to someone well acquainted with pytorch (for academic use)

2. how well does fast.ai interface with pytorch itself ? Can parts of program be in fast ai and other parts be in pytorch ?

3. Am I correct in assuming that despite being very fast, fast.ai is still slower (even if marginally so) than pytorch itself ?

1. Less boilerplate, so you can focus on your algorithm; best practices built in, so you might be surprised at how your speed and accuracy may improve

2. Very close integration with PyTorch. fastai is designed to extend PyTorch, not hide it. E.g. fastai uses standard PyTorch Datasets for data, but then provides a number of pre-defined Datasets for common tasks

3. fastai is not slower than PyTorch, since PyTorch is handling all the computation. It'll othen be faster than your handwritten PyTorch however, since we went to a lot of effort to use performant algorithms.

Thanks for replying.

From appearance alone, everything about fast.ai seems to be positive with no obvious negatives over native pytorch.

I often require to train baselines for my projects. That will be the perfect place to start using fast.ai. I am sold.

I have heard a few people lament the lack of keras in pytorch. Fast.ai also appears to take care of that.

Thanks for the great work.

1. You get a state-of-the art training loop that includes the latest research findings such as 1-cycle.

2. Yes. You can make everything project-specific if you want to. e.g. Image segmentation datasets are just slightly modified Dataset classes.

3. fastai extends PyTorch in a very Pythonic OO sense, so I think the only speed issues could come from that, and maybe maintaining a few extra dicts in memory. If it's about i/o, probably not, if it's about parallelization (for nlp), definitely not. In fact, I can't think of a way in which there's a significant speed penalty.

@jph00 where do I find lstm/gru/seq2seq layers for time-series sequence predictions (not text)? Also interested in autoencoder implementations. The fast.ai docs search does not really work for this. What do you think about other notable APIs built on top of pytorch such as Pyro and AllenNLP?

Hi Jeremy, I am a beginner. Going to start your ML course,then follow up with DL- 1 & 2. My question is when will these 3 be updated to the new version of fastai. I am aware of the fellowship for DL-1, but there's a 12 hour difference. Also, I want to start with ML,so, I want to learn with the new version. Thanks

On the side, I am also taking Andrew's coursera course for a theoretical grounding , Bengio and Goodfellow's Machine Learning Book ,and Hands-On Machine Learning with Tensorflow and Sci-kit( O'reilly book).

Cannot recommend fast.ai enough. How they and Jeremy especially communicate is absolutely outstanding, and I guarantee you will learn a lot by the results-driven approach that they have.

ZDNet just published an article with some more background about this BTW : https://www.zdnet.com/google-amp/article/fast-ais-new-softwa...

Can this be used / intended for production ? Is it fast/efficient enough on common hw ?

We are not really doing cutting edge or research stuff. Just developing a big dumb resnet and hoping to scale up to 10s of GPUs over 2019.

I was really happy to read a comment here about how this framework was used to reproduce a 6 month project in 2 weeks :)

It's very fast indeed: http://www.fast.ai/2018/08/10/fastai-diu-imagenet/

For inference, new features being announced today at the conference will help a lot too.

Will serving be as fast as TensorFlow? I assume PyTorch's dynamic network is inherently slower than TensorFlow's compute graph approach

I'm a big fan - I've been using their old version of their library for NLP on medical text, but the course and library certainly removed a lot of the "mystique" around deep learning for me.

This is awesome! I'd love to use Fastai, but my company's pretty committed to Tensorflow right now. Is there any work being done to port Fastai over to Tensorflow?

Is it meant to be a better alternative to Keras?

Keras doesn't support PyTorch, so it's not exactly an alternative. But it would be fair to call it an inspiration for fastai.

Is there a guide available for image segmentation? I've only really used TF in the past

There's a development notebook for segmentation on the Carvana dataset. https://github.com/fastai/fastai_docs/blob/master/dev_nb/006...

Most of the functions in there are pre defined in fastai so you can remove much of the code there in practice.

We'll be adding a proper segmentation example soon.

The name fastai sounds suspicious. There is "fast" and "AI" in it, so I fear that it makes the easy things easier and the hard things harder. It just sounds like too much fluff. We obviously don't have AI yet (but pretty convincing machine learning models) and we won't get it in the next years (not fast) - so if software, a company or an organization puts "ai" in its name (or chooses it as top-level domain) I suspect they just want to ride the hype train and am very skeptical ...

@Topic: What is the difference to Keras, PyTorch, etc.? They are already pretty high-level and the basic models for common tasks are available in all libraries at that point.

I mean, _just_ basing your impression off of the name is not going to get you very far. As a first impression, sure, but if a bunch of comments are saying they really like it, maybe it's worth doing a bit more exploration.

And AI is commonly-used as basically a synonym for Machine Learning in the ML community. Yes, the word has multiple definitions and connotations, and some people prefer to avoid using the term AI at all because of that. But that doesn't meant that someone using the term AI to refer to ML is wrong or up to something.

Furthermore, just because non-legitimate organizations want to ride the hype train, it doesn't mean legitimate ones shouldn't benefit from that effect as well. fast.ai is an organization actually writing software for doing machine learning, as opposed to a completely unrelated company (like an electric toothbrush company) claiming their latest product "uses AI" whatever that means.

The funniest part is that the slogan for fast.ai "Making neural nets uncool again". If that isn't explicitly going against the hype and saying this is _actual_ neural nets and not just fluff, then I don't know what is.

> I mean, _just_ basing your impression off of the name is not going to get you very far.

There are so many neural networks libraries now and I don't find the time to test all of them - dismissing libraries that use .ai in their website's TLD (or in their name) served me very well so far. Sure if it really makes neural networks that much simpler to train, I will take a look, but for now it does not seem to be more high-level than Keras or PyTorch.

It does not matter if you can describe your network in 50, 100 or 200 lines of code as the hard part is to make it learn, choose hyper parameters, change loss functions, etc. - this is possible in all frameworks.

And I don't think AI is used as a synonym for machine learning in the community, but is used to describe AGI.

Reply on news.ycombinator.com