[SOLVED] MachineLearning Homework 6-Anime Face Generation

Anime Face Generation

TODO

● Train your own anime face generator using Generative Adversarial Networks.

Crypko

https://crypko.ai/#

Data Collections

●

Thanks to Arvin Liu for collecting this dataset.

Data Format

• ●  The download link is in the sample code.
• ●  Unzip cripko_data.zip, the data format is as below:
• ●  faces/
  ○ 0.jpg

  ○ 1.jpg ○…

• ●  Total 71314 .jpg files in a folder.
• ●  DO NOT use any extra data and

  pretrained models.

JudgeBoi – Submission Format

● You should generate 1000 images, and name each image <number>.jpg ○ e.g. 1.jpg, 2.jpg, … , 1000.jpg

● Use tar to compress your images, and name the file with .tgz as extension.

○ e.g. images.tgz

• ●  The untarred files should not contain the folder.
• ●  The compressing code is provided in the sample code.
• ●  To create such a compressed file by yourself, follow the 2 steps below:
  • ○  cd <the folder containing your generated images>
  • ○  tar -zcvf ../images.tgz *.jpg
• ●  The folder containing your generated images should only contain 1000 images.

JudgeBoi

• ●  5 submission quota per day, reset at midnight.
• ●  Users not in whitelist will have no quota.
• ●  Only *.tgz file is allowed, file size should be smaller than 2MB.
• ●  The countdown timer on the homepage is for reference only.
• ●  We do limit the number of connections and request rate for each ip. If you

  cannot access the website temporarily, please wait patiently.

• ●  Please do not attempt to attack JudgeBoi, thank you.
• ●  Every Wednesday and Saturday from 0:00 to 3:00 is our system

  maintenance time. If the website cannot be used during this time, please wait patiently for the completion of the maintenance.

JudgeBoi

• ●  Since the evaluation metric in this homework requires GPU computation, the JudgeBoi server cannot serve too many submissions at the same time.
• ●  Under normal circumstances, JudgeBoi will complete the evaluation within one minute. If pending conditions are encountered, it may be longer. Please wait patiently after you submit.
• ●  However, if you have waited more than two minutes for the progress bar to finish, please refresh the page and try to upload again.
• ●  Please DO NOT upload at the last minute; no one knows if you can upload successfully.

NTU Cool

● Upload your code to NTU Cool.
○ TAs will check your code if necessary.

● If you beat the boss baseline in JudgeBoi, you may submit a report named report.pdf to explain the method you use to obtain the extra 0.5 pt.

NTU Cool – Submission Format

● Zip your code and name the compressed file <student_id>_hw6.zip ○ e.g. b06901234_hw6.zip

• ●  Do not submit your model checkpoints and the dataset!!!
• ●  Do not submit your generated images (images.tgz)!!!
• ●  We can only see your last submission before the deadline.

Evaluation Metrics

● FID (Frechet Inception Distance) score

• ○  We use the FID score as one of the evaluation metrics.
• ○  The FID score assesses the similarity between two datasets of images, which is the lower

  the better in this task.

Evaluation Metrics

● AFD (anime face detection) rate

• ○  To detect whether an anime face is in a given image.
• ○  The detection rate is the higher the beter.

Grading (10pt + 0.5pt)

• ●  Code 4 pt
• ●  Simple Baseline 2 pt

○ FID ≤ 30000 AND AFD ≥ 0.00 ● Medium Baseline 2 pt

○ FID ≤ 11800 AND AFD ≥ 0.43 ● Strong Baseline 1 pt

○ FID ≤ 9300 AND AFD ≥ 0.53 ● Boss Baseline 1 pt

○ FID ≤ 8200 AND AFD ≥ 0.68

● Bonus 0.5 pt
○ Submit a PDF report to explain your method (< 100 words in English) if you beat the Boss

Baseline.

Regulation

• ●  Do NOT submit the training data to JudgeBoi directly.
• ●  Data Augmentation (shifting, flipping, …,etc.) on training data is allowed,

  while it is forbidden to submit these processed images to JudgeBoi.

Regulation (cont’d)

• ●  You should NOT plagiarize, if you use any other resource, you should cite it in the reference. (*)
• ●  You should NOT modify the generated images manually.
• ●  Do NOT share codes or generated images with any living creatures.
• ●  Do NOT use any approaches to submit your results more than 5 times

  per day.

• ●  Do NOT search or use additional data or pre-trained models.
• ●  Your final grade x 0.9 if you violate any of the above rules.
• ●  Prof. Lee & TAs preserve the rights to change the rules & grades.

  (*) Academic Ethics Guidelines for Researchers by the Ministry of Science and Technology

DCGAN (Sample code)

● Weight initialization ● Generator

○ ConvTranspose + BatchNorm + ReLU ● Discriminator

○ Conv + BatchNorm + LeakyReLU

DCGAN

WGAN-GP

●

Wasserstein GAN (WGAN)

• ○  Remove the last sigmoid layer from the discriminator.
• ○  Do not take the logarithm when calculating the loss.
• ○  Clip the weights of the discriminator to a constant.
• ○  Use RMSProp or SGD as the optimizer.

  Gradient penalty (GP)

• ○  Use gradient penalty instead of weight clipping.
• ○  Use Adam instead of RMSProp as the optimizer.

●

WGAN WGAN-GP

Spectral Normalization GAN (SNGAN)

● Discriminator
○ Perform spectral normalization on the weights of each layer.

SNGAN

Baseline Guide

● Simple
○ Random submission

● Medium

• ○  DCGAN + WGAN
• ○  2~6 hr

  ● Strong

• ○  DCGAN + SNGAN
• ○  2~6 hr

  ● Boss
  ○ AutoGAN, BigGAN, Progressive GAN, Self-Attention GAN, StyleGAN, StyleGAN2 ○ 6~16 hr

Links

● Colab