2023-Q1-AI 19. CycleGAN

19.1. Materials / Video 🔴 5 Jun 2023, 19:00

Video: https://youtube.com/live/EUy3JqHVQTU?feature=share

Jamboard: https://jamboard.google.com/d/1Ihbq4z-TQWoHH_7spW9Q9fJt0z2tVfU-VFAPtQ2MMrI/edit?usp=sharing

Paper: https://arxiv.org/pdf/1703.10593.pdf

19.2. Implementēt CycleGAN

Implement CycleGAN based on instructions in 19.1

Template: http://share.yellowrobot.xyz/quick/2023-5-31-F72F1C44-8785-4680-9956-E7E34B00CC5E.zip

Iesniegt ekrānšāviņus ar labākajiem rezultātiem un programmas pirmkodu.

19.3. Mājasdarbs - Pievienot UNet Pix2Pix ģeneratoru

Implementēt Ģeneratora arhitektūru Pix2Pix, izmantojot pētījumu: https://arxiv.org/abs/1611.07004

Implementēt Summer2Winter Yosemite datu kopu https://www.kaggle.com/datasets/balraj98/summer2winter-yosemite

Iesniegt ekrānšāviņus ar labākajiem rezultātiem un programmas pirmkodu.

Video / Jamboard

Jamboard iedotas tiesības: vecins.valters@gmail.com

Video RTMP key: ea11-mrgb-4jg2-4ajc-d4hr

Kods iepušots GIT

Iepriekšējā gada video

Video: https://youtu.be/9iEaQMn_5fc

Jamboard: https://jamboard.google.com/d/1t4I-Ow2nPsjABaiD2Qd8z_auo3bRkGZTFX5f3nmXSqI/edit?usp=sharing

Būtu labi pārtaisīt uz horse2zebra dataset (Aigai bija labi apmācīts, Reinis arī pārziana)

Hoses to zebras source code:

https://www.kaggle.com/code/balraj98/cyclegan-translating-horses-zebras-pytorch

Good material: https://towardsdatascience.com/cycle-gan-with-pytorch-ebe5db947a99

https://blog.jaysinha.me/train-your-first-cyclegan-for-image-to-image-translation/

nn.ReflectionPad2d(3),

https://www.machinecurve.com/index.php/2020/02/10/using-constant-padding-reflection-padding-and-replication-padding-with-keras/

Attīstība:

1. DTN-GAN

2. ⚠️ Pix2Pix, PatchGAN => UNet architecture piespiež saglabāt aptuvenas features (Parasts GAN ar UNet prior)

   • ⚠️ “PatchGAN” classifier, which only penalizes struc- ture at the scale of image patches

   • Markov random field -> independence separated by patch diameter

   • https://arxiv.org/pdf/1611.07004.pdf

3. LSGAN - https://arxiv.org/pdf/1611.04076v3.pdf

   1. Alternative to WGAN

4. CycleGAN

   1. https://arxiv.org/pdf/1703.10593.pdf

5. RecycleGAN - Temporal Information in Loss

6. BicycleGAN - VAE tipa GAN - nav saistīts ar RecycleGAN un CycleGAN

7. Better CycleGAN

   1. Same discriminator 3 classes! https://ssnl.github.io/better_cycles/report.pdf

LSGAN

Pix2Pix - PatchGAN Discriminator (pa labi full Image GAN vs PatchGAN)

Pix2Pix infilling task

Pix2Pix GAN -> UNet

RecycleGAN paper: CycleGAN mode collapse - pa vidu obama - mainās viens pikselis, kurš iekodē informāciju.

we get better outputs with our approach combining the spatial and temporal constraints.

BicycleGAN: Z injection