Create something I have recently learnt from the Fast.ai course. Implement it to help make life easier for myself and maybe others. Seemingly easy(was not) and quick(way not)
What it does
Takes in images of avocados. It is a resnet32 CNN thats trained on images picked by hand into the three classes. Once the model was trained on a set of just 55 images, it has returned an accuracy of 85%.
How I built it
Using the fast.ai library that is built on top of pytorch. It makes life easier since it has pre-trained model weights for a resnet32 CNN. Also using a Gcloud instance makes for fast processing.
Challenges I ran into
- Image data set. Avocados have various infographics on image google searches. But to download images of skins of avocados to represent the various classes soon proved that there have been many duplicates/ artificial photos.
Accomplishments that I'm proud of
Getting a decent accuracy on a small data set.
What I learned
How important good data is and how much more attention it needs.
What's next for Avocado ripeness classifier
The aim was sky high. We wanted to build a website and an api so that we could use our model in day to day applications.