svmihar / sertifGAN

stylegan2 interpolation on various e-certificate shared online [publicly]

• satuin gambar, dengan cara cari seluruh format gambarnya pake glob, terus pindahin ke satu folder aja
• search duplicated dataset with phash
• buat model untuk cek ini sertif yang beneran apa nggak -> simple cnn would do i guess?
• training pipeline di kedro

dataset

• semua, full size, >90 accuracy on sertif_clf
  • sertifikat.zip (restricteD)
• resized, non transparent, converted to 512*512
  • sertifikat_resized.zip (restricted)

see you in a month. lol.

batch #1

training is actually done, first batch evaluations are:

• log loss didn't converge
  • images are too "unique"
  • incosistent padding / whitespaces that surrounds the image
    • 
  • karena di paksa resize 512 * 512, some vertical aspect ratio certificate got squished. guess will try to find them and kill delete them

batch #2

fixes to do:

• sertifikat classifier
• gak usah di dedupe, 'agak mirip' is fine -> set the threshold to be >=2
• re-label the sertifikat classifier images, for a better criteria
• watch dog kalo model udah selesai training -> upload ke drive
• watch dog biar modelnya yang disimpen cuman 2 aja