Training UGATIT in FP16
Hello, this is one of my full-time self learning projects. This was done in fastai2⁹, my GPU setups for this will be two 2070 Supers. The goal of this was to use fp16 training on U-GAT-IT. I did not do any of this in a controlled and meticulous experiment fashion. While I do agree that this is very important, I found that I had to forgo this in order to get something working. For the first two months of this I did not even have a working model for example. The target audience for this post is someone with either a single MOOC or introductory deep learning class, if you are in this category and do not understand something, feel free to leave a comment and I will try to answer your question, and update the article to be clearer. If you do not have any knowledge on deep learning this article might be a bit difficult, feel free to ask questions and I will do the best to answer them, and I really recommend learning fastai at https://course.fast.ai/
- Training U-GAT-IT with fp16
- Results
- Background
- FP16
- Batch Size of 3
- 2-category Categorical Cross Entropy VS Binary Cross Entropy with Logits
- Tanh Loss
- Normalization
- Tooling
- Contact Me
- References
from nbdev import export2html
from pathlib import Path
from PIL import Image
Hello, this is one of my full-time self learning projects. This was done in fastai2⁹, my GPU setups for this will be two 2070 Supers. The goal of this was to use fp16 training on U-GAT-IT. I did not do any of this in a controlled and meticulous experiment fashion. While I do agree that this is very important, I found that I had to forgo this in order to get something working. For the first two months of this I did not even have a working model for example. The target audience for this post is someone with either a single MOOC or introductory deep learning class, if you are in this category and do not understand something, feel free to leave a comment and I will try to answer your question, and update the article to be clearer. If you do not have any knowledge on deep learning this article might be a bit difficult, feel free to ask questions and I will do the best to answer them, and I really recommend learning fastai at https://course.fast.ai/
Outputs
Image.open('imgs/big_img13.png')