Filtersize

Want to understand the training process better? Got tips for which model to use and when? This is the place for you


Forum rules

Read the FAQs and search the forum before posting a new topic.

This forum is for discussing tips and understanding the process involved with Training a Faceswap model.

If you have found a bug are having issues with the Training process not working, then you should post in the Training Support forum.

Please mark any answers that fixed your problems so others can find the solutions.

Locked
User avatar
Scrapemist
Posts: 16
Joined: Sat Nov 05, 2022 2:26 pm
Has thanked: 8 times

Filtersize

Post by Scrapemist »

I am comparing preset parameters between sae-df and saehd-df and I notice there is a difference in filter size.
Saahd has a dec_res_block and dec_output_kernel of both 1 and sea of 2 and 5.

Does this mean that smaller filters results in better output?
I'm assuming the assed -hd in saehd means higher definition.

User avatar
torzdf
Posts: 2651
Joined: Fri Jul 12, 2019 12:53 am
Answers: 159
Has thanked: 129 times
Been thanked: 622 times

Re: Filtersize

Post by torzdf »

No, decreased kernel size does not intrinsically mean better output, it just means it uses a larger kernel for convolutions.

Kernel size of 1 is a bit different too, in that it does not look at any neighbouring pixels, and tends to be used to reduce spatial data.

These are quite complicated concepts to get your around, and basically dictates (in fairly simple terms) how a layer gets sampled for passing on to the next layer. It's probably best to google around a bit about convolutional kernels to understand them better.

eg:
https://stats.stackexchange.com/questio ... -size-mean
and
https://stats.stackexchange.com/questio ... al-network

My word is final

Locked