Sure, it’s danish though, called Weekendavisen. Until earlier this year it didn’t have digital news, they do have a digital edition now, but I’ve never used it.
I recently ran into that issue - I have a nice fine-tuned yocto image and now a customer wants to "log in via ssh and install a custom raspbian package". I was thinking of the docker solution as well.
"We have small and fine-tuned ghetto, but customers don't want that. What to do?"
I have similar problem: we have consumer device on top of imx6. I used iMX6 SoloLite evaluation board kit (1GB of memory) with Fedora for arm as development prototype. For me, installing a package is not a problem at all: "dnf install package". I planed to use Raspbian or Fedberry on actual device, but our management blindly switched to Yocto, without comparing of alternatives. To install package on Yocto, we need to create our own recipe in our own layer,then rebuild Yocto, fix bugs and repeat, and then deliver update. We spent man-year to achieve same result as we already had with Fedora. We seriously lag behind schedule because two developers (and two consultants) are working on developing of our own custom OS instead of working on our application. Moreover, instead of single system for development/testing and production (Fedora or Debian on developer workstation, in Docker for CI, on actual device) we now have mix of two systems. I created configuration script with about fifty options to be able to compile for host and for Yocto at same time (including workaround for number of bugs in Yocto). Enormous development time was sink into ground because of this schizophrenia.
I already have OS. In my case, I just use chroot script to access full Fedora distro with all development tools on Yocto. If I will be able to replace Yocto, I will replace it with major distro, like Debian, Ubuntu, CentOS, or Fedora, not by yet another toy distro with thousands of packages per maintainer.
From the paper:
"we have to use all the GPUs in DGX1 (8 V100 GPUs, each with 16GB memory) for training. We distribute the generator computation task to 4 GPUs and the discriminator computation task to the other 4 GPUs. Training takes ∼10 days for 2K resolution."
As I don't have a DGX1 here, training the 2K resolution net for 10 days on a p3.16xlarge instance (also has 8 V100 GPUs) would cost USD 5875 on AWS.
(USD24.48 per hour on-demand pricing * 24 hours/day * 10 days)
The DXG-1 costs $129,000 so AWS is cheaper unless you need to do it 22 times. And you can have multiple instances running at once and get all of your results in ten days, instead of waiting ten days again for each run.
Well plus electricity. A DGX-1 takes 4 kilowatts or so that 10 day training run will take just about a megawatt hour or about $100 at retail. So the cross over point is more like 23 runs :)
DGX 1V has $7500 worth of CPU alone to feed the gpus. Throw in 8 TB of nvme ssd for training data and you're looking at something more like 1/5th the cost.
V100 has ~50% higher memory bandwidth than 2080Ti, so you probably wouldn't get 80% of the performance. Also, only two 2080Ti can be connected via nvlink.
Their GitHub README says 24GB, and 12/16 GB requires cropping and performance not guaranteed. I’ve only seen P100 with 16 each, and its the big Quadros that have 24
One of the example translates a full human pose to a video of a dancer. If the network would be trained on the facial pose(?) / features only, would that recreate something like the facial reenactment in http://niessnerlab.org/projects/thies2016face.html (source code for face2face is not public)?
Yes, but it just cuts out the face and pastes it on a different person/background. It does not do full reenactment where you keep the entire target video environment.
