Paddle: Baidu's open source deep learning framework | Hacker News Reader

Paddle: Baidu's open source deep learning framework

github.com

230 pointsthinxer19 comments

Looks well put together, well documented. Quickstart tutorial was easy to follow.

Why should I think about using this instead of (or in combination with?) the plethora of other similar offerings out there?

Well, if you are a "Internet celebrity hound", Andrew Ng (of Stanford, machine learning online course, and general HN fame ) is Chief Scientist at Baidu now.

Otherwise, I agree that it has an uphill battle with some other entrenched frameworks.

One thing which is missing is a "model zoo", a place where people share models (usually well-known) pre-trained, which is very useful for starting to use a framework.

Although this repo https://github.com/baidu/paddle_paddle_model_zoo, suggests they might be working on one.

There is a model zoo here: http://www.paddlepaddle.org/doc/demo/#model-zoo

Another deep learning framework, this time from Baidu.

Given TensorFlow's rising dominance with AI researchers and practitioners and the existence of other frameworks with large installed bases like Theano, Torch, and Caffe, I don't think this new framework has much chance of gaining wide adoption in the US or other markets in the West. In my opinion, TensorFlow's network effects are too large to overcome at this point.

However, Paddle could gain significant adoption in China, Baidu's home market.

EDIT: My opinion could be wrong. To find out, I've created an HN POLL so we can all see which deep learning frameworks the HN community would use to build new products and services today. Link to HN POLL: https://news.ycombinator.com/item?id=12391744

TensorFlow was open sourced 9 months ago. I think it's too early to discredit newer initiatives.

Agreed. There is a lot of hype around Big G's tech. A year or two ago, it seems some people were saying all microservices would be built using Kubernetes (because Big G's internal tech stack is lightyears ahead of others). Hasn't happened for sure. I think these are the early days ... the winner will be determined by how well the platform scales, how easy it is to maintain infrastructure and how easy it is to use. Just my personal opinion :)

Also see "The deep learning frameworks landscape, August 2016" from @fchollet: https://twitter.com/fchollet/status/765212287531495424

The vast majority of Tensorflow users are students from the Udacity course. So the "rising dominance" of Tensorflow is actually disputable. 95% of those users have little experience, let alone the opportunity of deploying to production.

As a technology, Tensorflow is not better than other frameworks in many ways. It has a nice website and a few tutorials, but it doesn't do well, for example, in large production environments.

The field of DL libraries is actually wide and changing quickly. TF, CNTK, DSSTNE and other libs all came out in the last 10 months or so.

So Paddle has a chance, especially in China, where Tensorflow is handicapped because people can't use the Google Cloud on which it is optimized.

vonnik: I agree Paddle has a chance in China. That's what I wrote above :-)

As to TensorFlow's dominance, maybe you're right, I could be wrong. That's why I created the poll, to find out which framework the HN community actually wants to use: https://news.ycombinator.com/item?id=12391744

> don't think this new framework has much chance of gaining wide adoption

I don't believe it needs to, I believe diversity in a relatively young field is an advantage as it helps to not get stuck in a local minimum, so to speak.

what exactly is the differentiation between the frameworks?

I think that TensorFlow gets so much news because it's Google's play and Google is a big player, less so because it's clear that it's the technical winner. Having only briefly looked into libraries, Caffe seems to be the winner in actual usage and in real results. It's also used in lots of benchmarking. TensorFlow is just too new to have much dominance.

Of course, it's quite likely that I'm very wrong too. But I think your poll is going to be largely answered by people who are not doing deep learning but might do it in the future, which is going to be based mostly on what hits the news, rather than on the merits of the packages themselves. When polling a general discussion community like HN, I think it would be important for people simultaneously vote on which of the packages they 1) have heard of, 2) have used, 3) would use for a new project today.

Caffe has the model zoo and is pretty damn good at CNNs where you don't need new algorithms. For research, theano may still beat tensorflow in many ways, but it remains to be seen.

Easy to use.

Are you asking what the point of open sourcing software is?

If you want to be cynical about it, consider that you already have it embedded in your infrastructure. Your real value is the stuff you've built on top of it. If it's open and widely used you'll have an easier time hiring experienced people; you might even acquire a company with something new and useful to you that's built on the stack you already use; and, you know, someone might even fix a bug or add some great new functionality that you can take advantage of.

Plus if your stack becomes dominant, it's the other guys' (your competitors) problem to change their stack instead.

Interesting project name. The authors seem to be telling us something about their philosophy of education :)

Reply on news.ycombinator.com