michaelgreen

8 anni fadiscuss

I would say if you really wanted to make this easier for streamers to use you could use helm to package up a kubernetes configuration (a bit of work) and then put it into something like google cloud launcher so you could just point and click to set this up.

michaelgreen

8 anni fadiscuss

Is there a name for the pattern recognition section? I'd be interested in reading more about it and how you use it as a measurement/how the results correlate with the decision process. Does it tie in with the behavioral aspect at all or are they completely independent?

michaelgreen

8 anni fadiscuss

Yeah they added it in the most recent version of iOS (11.3) but progressive web apps have a very loose definition the most basic of which is, as you correctly stated, using a service worker. However service workers are just the ground floor of a much bigger building of features that should be able to be leveraged cross browsers.

michaelgreen

8 anni fadiscuss

We built a PWA with vanilla javascript that does a lot of augmented reality work with the camera. The main issue is the lack of iOS support. Apple recently updated iOS to include support for the service worker (literally like this week) but it still lacks a huge number of features that chrome on Android has and basically every desktop browser has.

You do get access to a lot of the core sensors on the device but the inconsistencies on iOS just really make life a lot harder. Hopefully this will change over the next year but I'm not sure how quickly it will happen. So basically yes, you should build PWAs but until apple gets completely on board you can't really leverage them to their full extent like you can with chrome on android.

michaelgreen

8 anni fadiscuss

No problem (:

michaelgreen

8 anni fadiscuss

I've tried it with an external monitor which works great but I haven't tried it with any games. From looking at the eGPU forums it seems like people aren't running into that many problems with it.

michaelgreen

8 anni fadiscuss

So there I didn't really follow any blog post or anything, there's a lot of gists about setting it up but they become relatively out of date fairly quickly and are usually specific to their setup.

I would checkout: https://egpu.io/forums/mac-setup/wip-nvidia-egpu-support-for...

A lot of people have put in a lot of work to make it as easy as possible to setup. I would just make sure you setup things one at a time and don't immediately jump to trying to get TF or pytorch to work after installing the drivers/ following any sort of guide. Verify your Cuda installation first by building the sample programs and running them. (see http://docs.nvidia.com/cuda/cuda-installation-guide-mac-os-x...)

Other than that maybe checkout: https://gist.github.com/jganzabal/8e59e3b0f59642dd0b5f2e4de0... https://gist.github.com/smitshilu/53cf9ff0fd6cdb64cca69a7e28...

The main thing is just getting the GPU drivers setup, after that installing tensorflow requires some modifications to the source (relatively trivial find and replace), and I don't think you have to do anything special for the latest version of pytorch.

Other tips:

1. Make sure you're using a thunderbolt 3 cable with whatever egpu you have, other cables will not work even if they're usb c. (USB 3.1 != Thunderbolt 3 != USB C) Read about the differences.

2. I would recommend the AKITO Node enclosure as it seems like most people use it and the community is really small already so if you aren't it might be more difficult to debug issues, but I wouldn't say you couldn't use something else.

3. You'll want to have a docker container that has the CPU version of tensorflow/whatever so you can use that when you don't have the GPU readily available as tensorflow won't work if you installed it with GPU support and your GPU isn't there.

4. If you're trying to use Nvidia-Docker I'm not sure if anyone has gotten that to work on mac because device binding isn't supported on mac for docker. You might be able to get this to work by modifying the docker VM but I'm not sure.

michaelgreen

8 anni fadiscuss

In practice they work just fine, it's just very annoying to initially set them up and you have to do some patching to get the latest version of tensorflow working with it. I use one every day and I haven't run into any problems while running models with both pytorch and tensorflow.

michaelgreen

8 anni fadiscuss

So I've been using an eGPU for about 3 months now and it is amazing. This isn't officially supported so you end up having to do a lot of work arounds to get things like tensorflow/pytorch working.

It doesn't let you hot plug your eGPU into your computer (you have to restart for it to work) but for officially supported eGPUs you're now able to do that.

I've trained various models using a Titan XP and it's so awesome to be able to maintain the portability of your laptop and still get all of that power. A large benefit is also not having to move training data around between servers and other machines if you have it on an external drive or just on your laptop.

After you get everything up and running there isn't that much maintenance or anything you have to do regularly to keep it working. The speedup is incredible and it was definitely worth it for me personally, however it's not a walk in the park to get it set up initially.

michaelgreen

8 anni fadiscuss

I've been using an eGPU for a few months now and It's fantastic. It's a bit annoying to deal with all of the workarounds for libraries like pytorch/tensorflow so you can use the latest version but other than that it's great.

michaelgreen

8 anni fadiscuss

Wow thanks, reading it now (:

michaelgreen

8 anni fadiscuss

Okay I see what you're saying and I completely agree.

michaelgreen

8 anni fadiscuss

Jeff Dean talks about AutoML using RL and in the paper " Neural Architecture Search with Reinforcement Learning" it also talks about this.

Also it seems different from more traditional hyperparameter optimization because it makes novel cells. So the structure of the network isn't limited to our existing library of layers/cells.

https://arxiv.org/abs/1611.01578 https://youtu.be/HcStlHGpjN8?t=2073

michaelgreen

8 anni fadiscuss

This is huge, and it's only an alpha. I begun reading about AutoML/Neural Architecture searches around ~year ago and something I've been thinking about is:

Why doesn't this just move the optimization problem? Aren't you now just optimizing your DeepRL network rather than the network you're trying to optimize?

michaelgreen

8 anni fadiscuss

The first thing that comes to mind is peer reviews. Although it's not inherently quantitative data and doesn't remove any of the bias/politics/etc. it can be a good indicator of who others feel are helping them.

However as the writer of the article pointed out often times you can't see when other people are improving your ability to do work whether that's through writing more documentation, improving some internal infrastructure, or other often invisible things that can drastically improve your life as an engineer.

I think an effective approach would be through time tracking (i.e: writeup on how you spent your time for the day) in conjunction with some analysis on how the time you spend on certain things correlates with what other engineers are doing. This of course would rely on everyone being detailed in their writeups and the analysis still might not catch certain things because it's difficult to write down exactly how you spent your time. Perhaps you could only write down the three most difficult/annoying problems you ran into and just track that.

I remember reading about some company that is trying to make this easier by tracking computer use and creating graphs of how people interact within their organization and how that changes over time but that doesn't seem like it would work everywhere.

I don't know if there is a way to truly quantify teamwork or organizational impact. I think the important thing is to encourage your organization to communicate more: talk about what you're doing, why you're doing it, and how it helps others within your organization. If your organization is fairly big write about it and share your writeup internally. Train people within your organization to speak up if they feel they aren't being appreciated for the impact they made.

Ultimately I think it's an extremely hard problem because it stems from the question of what having a positive impact means. The answer to that can branch off into millions of different small tasks that could be totally unique to a subset of people within your org.

----

The best you can do is:

* Measure what you can.

* Know that's not the entire picture and it never will be.

After that try to create a culture that:

1. Speaks up about their own impact

2. Shares the impact of others' work, especially of those who might not speak out much

3. Is openly appreciative of and encourages 1 and 2.

comments