![]() This dataset will be stored in the trainA folder. Each image is 200x200 pixels and contains one face. It contains Simpsons faces extracted from a few seasons. Thankfully, I stumbled across a dataset on kaggle that had everything I needed. And after I looked at the images I noticed that some of them didn’t include any faces at all. But it took so long to download all the images. Unfortunately to do that it looks like you need a developer key from google console. Initially my thinking was to scrape images from google images. To create this dataset I needed to find close up shots of Simpsons characters and close up shots of regular people. ![]() !git clone import os os.chdir('pytorch-CycleGAN-and-pix2pix/') !pip install -r requirements.txt DatasetsĬreating the dataset is harder than you would initially think. Installįirst we need to install CycleGAN. So this isn’t likely to turn out well.īut I’m going to attempt this anyway. It’s worth noting that the paper explicitly mentions that large geometric changes are usually unsuccessful. Cyclegan would translate that into a Simpsons Character. The idea is that you upload a photo of your face. I was inspired by sites like turnedyellow and makemeyellow. I thought this could be potentially applied to The Simpsons. ![]() Such as converting horses to zebras (and back again) and converting photos of the winter to photos of the summer. It’s been applied in some really interesting cases. Cyclegan is a framework that is capable of unpaired image to image translation.
0 Comments
Leave a Reply. |