Skip to content

Instantly share code, notes, and snippets.

@followben

followben/1-image-classification.ipynb Secret

Last active October 26, 2018 05:55
Show Gist options
  • Star 1 You must be signed in to star a gist
  • Fork 0 You must be signed in to fork a gist
  • Save followben/fff5e238625e212f506d7d8fba87567f to your computer and use it in GitHub Desktop.
Save followben/fff5e238625e212f506d7d8fba87567f to your computer and use it in GitHub Desktop.
Download ZIP
Applying fast.ai Lesson 1 techniques to my own data, incl. less than stellar (!) results on images of galaxies
Display the source blob
Display the rendered blob
Raw
1-image-classification.ipynb
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
@AdityaSoni19031997
Copy link

AdityaSoni19031997 commented Oct 25, 2018 •
edited

Imagenet had galaxies as one of its classes? If not then we need to unfreeze the layers and retrain
Transfer learning is applied when we know that we have similar or related classes with the data in our hand

Check this https://datascience.stackexchange.com/a/28387/35644

@followben
Copy link
Author

I've updated the notebook, this time running 20 epochs instead of 4 (as suggested by Jeremy). Indeed, error rates for galaxies with the resnet34-based model are lower now, but still a long way from bears :)

@followben
Copy link
Author

3rd iteration, this time 10 epochs, then unfreeze, then 4 epochs. Still 74% accuracy is about as good as it gets.

@followben
Copy link
Author

4th iteration: 5 epochs, then unfreeze, then 4 epochs, resulting in 76% accuracy.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment