100-days-of-code

My 100 days journey with coding to improve my Machine Learning, Deep Learning, Data Science skills. Posting current updates on twitter: https://twitter.com/AgnMikolajczyk

Day 1

I continued with wav2wav #autoencoder for audio style transfer. Loss decreased gradually during the training, so I've started writing audio generation script Nuty Also, I've added json configuration files with parameters of e.g. melspectrogram.

• Wav2wav autoencoder: https://arxiv.org/pdf/1904.08983.pdf

Day 2

Had a little fun with casual convolutions in wav-to-wav autoencoder in #pytorch. Reimplemented&tried to understand it better. Later, experimented a bit with different architectures.

• Casual convolutions: https://medium.com/the-artificial-impostor/notes-understanding-tensorflow-part-3-7f6633fcc7c7

Day 3

I came up with an idea for a fun project after working hours. I always wanted to train GANs but popular datasets seemed boring -> I started scraping data with Selenium! It's my first time trying it out and first successfully clicked button felt awesome

• Selenium: https://www.selenium.dev/
• Selenium for data scraping: https://medium.com/the-andela-way/introduction-to-web-scraping-using-selenium-7ec377a8cf72
• BeautifulSoup: https://www.crummy.com/software/BeautifulSoup/bs4/doc/

Day 4

Started generating random RPG-like pixel characters with LPC spritesheet: http://tinyurl.com/yb6yg7kw. I'm creating my own dataset for my experiment with Generative Adversarial Networks. Soon I'll start implementing it in #pytorch

• LPC spritesheet: http://tinyurl.com/yb6yg7kw.
• Pytorch: https://pytorch.org/

My notebooks: * Download random spritesheet with Selenium * Extract character from spritesheet

Day 5

Implemented DCGAN from pytorch tutorial. * Experimented with latent size (input for Generator) and feature map sizes * Added soft and noisy labels * Added Wasserstein loss which is a good solution for mode collapse * Wasserstein loss - https://developers.google.com/machine-learning/gan/loss * mode collapse - https://machinelearningmastery.com/practical-guide-to-gan-failure-modes/ * Added dropout in both generator and discriminator

• DCGAN Tutorial

Dataset

• Download dataset here

Results

• My final code with modified DCGAN

Day 6

Today practiced a bit with custom data loaders in pytorch: https://pytorch.org/tutorials/beginner/dataloadingtutorial.html Started writing generator code for my wav-to-wav autoencoder.

Day 7

First week finished! Today, I've experimented with a XLA for TPU with multiprocessing, the notebook shared on Kaggle for Flower classification challenge. https://kaggle.com/dhananjay3/fast-pytorch-xla-for-tpu-with-multiprocessing https://www.kaggle.com/c/flower-classification-with-tpus/notebooks?sortBy=relevance&group=everyone&search=pytorch&page=1&pageSize=20&competitionId=18278

Day 8

Continued having fun with #GenerativeAdversarialNetwork - this time Conditional GANs on retro pixel game dataset - http://github.com/AgaMiko/pixelcharactergenerator I've managed to modify code and run the training, can't wait to see the results

Day 9

Today finished Conditional DCGAN experiments! GAN generates a pixel character seen from selected angle.

• different learning rate for discriminator and generator
• soft labels
• added classification loss to the discriminator. Discriminator have to guess fake/real but also the character angle
• generator is conditioned with embedding from trainable look-up table that gives the info about the character view angle

• notebook with modified Conditional DCGAN