Tensorflow 2 is on the way

Feature Requests
#1

Cool…

For those of you who don’t know why this is excellent news if you have a fairly recent GPU it won’t work with the existing releases of DAE owing to CUDA 9 being the only offering.

Essentially Tensorflow 2 means CUDA 11 (I hope) which has the knock-on effect of enabling those of us with newer NVIDIA cards to finally use our GPUs with DAE.

As to timescales I was just told that…

"A Tensorflow 2 update will be available soon."

Also, I got confirmation that the GAN can create “over a billion images” - I knew it would be a large number - just not what the number was.

I just completed 100,000 GAN images - I got one duplicate (I just generated one more and so now have no dupes)

1 Like
Holy Cuda 10.0, Batman
#2

Generating one more so that you have no duplicates in the set sort of defeats the point of that metric, right?

Thanks for the latest. Cheers!

K

#3

Nope - the point of generating digital signatures was to identify duplicate occurance which I knew was likely. DAE tell me there are over a billion possibilities which means the actual number is most likely either 2147483648 or 4294967296.

Personally I want to know if a duplication has occurred specifically so I can generate extra and be sure that I have 100,000 different images as I want to have 100k unique images as rewards for my game (it started out as 10k uniques…).

Images 80498 and 99082 (out of 100,000) generated the same signature which indicated a strong likelyhood of duplication (but not a certainty). I therefore had to visually compare them and, yes, they were identical.

The likelyhood of duplication increases as you have more images that you don’t want to match.

This is best illustrated by the “Birthday Paradox” where you take a group of people and see if any share a birthday. It always seems really wierd that if you have 23 people then there is a > 50% chance that two have the same birthday (it’s true)

Applying the same calculations to AI images actually reveals that the 50% probability occurs at either 54,563 or 77,164 (and in my case it happened at 80,498)

If I generated another 100,000 images some MAY be duplicates but this is not guaranteed to happen as there are vastly many more new images still available. All that can be said with a fair amount of confidence is that one of the 100k new images will be a duplicate (so throw it away and make another)

2 Likes
#4

Yeah I misunderstood what you were trying to do nbd

#5

Any updates coming your way about Tensorflow 2 being implemented? Thanks

1 Like
#6

I haven’t heard anything yet but I know that DAE definitely read the forums so merely asking the question here is a good idea :slight_smile:

I’ve just asked for a few small changes to the CLI as well which may, or may not, get implemented (chances are good I’ll get this particular one - not saying why). The requested change is simple but everything takes time.

Also, the recent Log4j security issue will probably delay things a little (not that it should affect DAE really) - I was informed of the update before the Log4j problem was found.

Personally I’ll just patiently wait for the TF2 version as the benefits should be worth the wait - it SHOULD work on my new laptop’s GPU (RTX 3060) which will be cool as current version don’t support newer GPUs

#7

I always get the impression that DAE must have a very restricted programming/development team. For example, they are the only company I know of that hasn’t already released a Windows 11 version of their software.

1 Like
#8

At least they finally released a Java 11 version :slight_smile:

1.2.7 - i.e. latest was first to use it

1 Like
#9

I am also quietly and patiently waiting for TF2.
i purchased in early 2020 and this software is fantastic… However I have a 3080 desktop and would love to use this software to its potential with 4k video etc. Recently been very impressed with art breeder… The morph animation function is something that could be an amazing addition to DAE… TF2 support is the big thing I’m awaiting for… Can’t wait, I hope this gets released very soon!

#10

Peardox - thanks for all your notes here and on your blog related to installing DAE, and using the GPU version.

As a new prospective user I am blown away by the potential of this product, but with no support for CUDA 11 I am finding it hard to even test it for my intended use. Hopefully this release is coming very soon.

1 Like
#11

Well, I just saw marketing materials for 30% off for Easter, so does that mean it’s still being developed?

#12

of course, a lot stuff is in the making.

#13

Oh, that’s great news! Like what? I didn’t know a dev read any of this

#14

Please make sure Aarch64 is included this time - essential for Mac M1 (side benefit is my Pi and Nano can also use it)

1 Like
#15

Im not a dev, but im reading all messages here :slight_smile: a lot of improvements, new models and new features.

#16

Well, Management reasd this stuff so the messaage will get thru :slight_smile:

1 Like
#17

Hello Peardox,
Is there any update to the TENSORFLOW2 ? so we can use this great software as intended, on a modern GPU. It is taking quite along time. I hope there is an update very soon… I only open up the software to check if there is a TF2 update… :frowning:

#18

They told me about TF2 in Decemeber

No news since

1 Like
#19

The world is at hell and sales of DAE are probably slow so depending on the Dev team implementing things like Automatically rewrite TF 1.x and compat.v1 API symbols  |  TensorFlow Core takes time and lots of trial and error. I know people are still alive Facebook is getting updates :grin: so no mass exodus I myself have picked up more interest in finding ways to create such art at a faster process.

Now I’m stuck on a site where you tell it what you want and the AI generates it for you. The beauty of it the small photos are zero charge and for DAE anything at 512x512 is a excellent sample for the other 3 renderings.

Shawn

#20

Well, six months and no new version

So… I decided to write my own and it’s coming along nicely - supports all NVidia GPUs released in the last eight years and a few from before then.

While it’s not a replacement for DAE it does work well. Styling is about the same speed but creating new styles is far faster - my 3060 can bang a medium quality one out in about half an hour (really good quality takes longer of course).

I’ll be releasing it for those of you who want to play with it in the next month (planned 1st August but alpha will be available as well)

Note that I’ve only been working on this for about a month so it’s currently nowhere near as flexible as DAE (it’s NOT a replacement - it’s an experiment)

I’m also looking at making it work with AMD GPUs (theoretically this is already the case for Linux - Windows is WIP for the project that makes it possible)