Training the AI about starnet HOT 4 CLOSED

This is weird result. What image size do you use for input? 256 by 256? What learning rates do you use? I would try to increase those and train for less, you should not need 50 or 70 epochs to see improvements.

From what I see I think that you have too many examples of huge stars in your datasets and neural net starts to try and remove all the bright areas that it sees.

from starnet.

Remember that your training dataset should be representative of what you will be feeding in after. Just cropping a bunch of huge stars and nothing else is a bad Idea.

The fact is that you see stars with spikes only rarely is a problem, but I don't think that this is the way to deal with this.

from starnet.

Thanks, I appreciate your time to reply! This is a good introduction to neural nets for me, I'm super new to working it out. I now realise that the learning rates are important, and are dependant on the input data. I was mainly feeding it crops of big "problem stars" and I can see that's what caused the issue. The graphs went to hell when I did that.

I was a little surprised that there aren't more people wanting to cooperate with some data with me on this, nor have I seen much about people training it on any of the forums. It's an awesome toy :)

I'm training new weights now overnight to see how it goes. I'm using a gtx1080, and I'm finding the ram throughput is my bottleneck, the gpu is nearly idle most of the time and it's 15-20 mins per epoch. The lower resolution images made this go much faster.

How many images is a reasonable training set? I'm used to seeing 10s of thousands in other studies of input samples!

from starnet.

I actually don't think this is the perfect project for introduction, GANs are a bit tricky to train. But if you like it, then why not)

Most people just want to use the code and don't care what's under the hood. Plus, it is a bit tricky and time consuming to get training data...

For 1080 I would try to increase image size to 512x512 and maybe play with batch size, this could help keeping GPU more busy.

For the number of images there is only one answer - the more the better. I don't know myself how many will top performance for this architecture. One thing you should remember is that the way I treat data here is a bit non-conventional. Usually one input image is one training example which is resized to match input size. Here I use random crops from image without resizing. So if input is 256x256 pixels, then, for example, one image with 2560x2560 resolution is effectively 100 training images. So you relatively easy get 1k or more samples. Not quite Google scale, but decent.

from starnet.