training sets and inputs

Training your model
Forum rules
Read the FAQs and search the forum before posting a new topic.

Please mark any answers that fixed your problems so others can find the solutions.
Locked
User avatar
siliconunit
Posts: 8
Joined: Sun Feb 16, 2020 11:26 am
Has thanked: 1 time

training sets and inputs

Post by siliconunit » Sun Feb 16, 2020 11:55 am

Hi!,

Great project! getting decent results,(1080 8gb / ryzen 7), and my face B is homemade HD filming.
I 've got a couple of questions thought!:

1. Can I increase the number of A or B faces on the fly? ie stop the training, update the folders, that's it ? (I'm using Villain)

2.Once I train face A and face B , is there a way to recycle the face B training part ?



SU

User avatar
torzdf
Posts: 548
Joined: Fri Jul 12, 2019 12:53 am
Answers: 86
Has thanked: 16 times
Been thanked: 120 times

Re: training sets and inputs

Post by torzdf » Sun Feb 16, 2020 11:57 am

Hello and welcome!

Your first question is answered here:
app.php/faqpage#f3r4

As to the second, it is possible to just replace the "A" data, to recycle "B", but this usually results in identity leak, so it's best to start a new model with your required pair.
My word is final

User avatar
siliconunit
Posts: 8
Joined: Sun Feb 16, 2020 11:26 am
Has thanked: 1 time

Re: training sets and inputs

Post by siliconunit » Sun Feb 16, 2020 12:30 pm

Awesome! I missed that bit !:)
What about gray-scale images, are they a good thing or will get confusing?

Thanks!

User avatar
torzdf
Posts: 548
Joined: Fri Jul 12, 2019 12:53 am
Answers: 86
Has thanked: 16 times
Been thanked: 120 times

Re: training sets and inputs

Post by torzdf » Sun Feb 16, 2020 4:56 pm

Should be ok in moderation.
My word is final

User avatar
djandg
Posts: 22
Joined: Mon Dec 09, 2019 7:00 pm
Has thanked: 2 times
Been thanked: 2 times

Re: training sets and inputs

Post by djandg » Sun Feb 16, 2020 9:58 pm

FYI - I've tried all greyscale images as one of the faces for training, result was poor (color wise and sharpness wise) swapping on a color video at 250k iters, original model. B&W images were also of poor quality due to being extracted from a film from the 1930's.

User avatar
bryanlyon
Site Admin
Posts: 274
Joined: Fri Jul 12, 2019 12:49 am
Answers: 20
Location: San Francisco
Has thanked: 3 times
Been thanked: 78 times
Contact:

Re: training sets and inputs

Post by bryanlyon » Mon Feb 17, 2020 6:47 am

All greyscale/poor quality on one side is not-ideal and will almost always lead to failure. This is due to the inability for the model to learn sufficient information from the poor quality source data.

Locked