NEKO
NEKO copied to clipboard
Add vqa playground notebook
Notebooks don't show up pretty in the Github diff viewer. You can see the pretty notebook here: https://github.com/eihli/NEKO/blob/vqa-eval-exploration/dev_notebooks/vqa_playground.ipynb
Explores the difference between training and evaluation runs. Trying to find why evaluation loss doesn't decrease.
Here's something that I found curious. The evaluation generates the correct target answer. But it includes some additional text after the correct text.
See the last screenshots.
Target answer: breakfast Predicted answer : breakfast Haw
Target answer: breakfast Predicted answer : breakfaststasy
etc...
The evaluation