meme classifier

My first hands-on deep learning classifier using data from imagenette and google images and building a small API of it.

Sequence of steps followed

• Data collection from imagenette and google images
• Load and Train on raw data with resnet34 architecture
• Data cleaning by removing similar images and from_toplosses
• Unfreeze model and train on cleaned images dataset
• Data Interpretation by confusion_matrix and from_toplosses
• Comparison with resnet50 architecture
• Prediction on test set and exporting the trained model
• Building a small API by taking above exported model

Files

-meme_classifier.ipynb : notebook used for performing above steps, using fastai

-export.pkl : Exported model after training (84 MB file)

-meme_api.py : A small Starlette API which accepts file upload as well as image URL and runs them against pre-calculated model to give prediction. Local Usage:

uvicorn meme_api:app

The model is deployed on Render at https://meme-classifier.onrender.com.

Template of the same can can be found here.

Examples

Input Image (upload): data/Unknown.jpg

Prediction: http://127.0.0.1:8000/upload

Input Image (URL): https://upload.wikimedia.org/wikipedia/commons/thumb/0/04/Greenland_467_%2835130903436%29.jpg/640px-Greenland_467_%2835130903436%29.jpg

Prediction: http://127.0.0.1:8000/classify-url?url=https%3A%2F%2Fupload.wikimedia.org%2Fwikipedia%2Fcommons%2Fthumb%2F0%2F04%2FGreenland_467_%252835130903436%2529.jpg%2F640px-Greenland_467_%252835130903436%2529.jpg

References

https://course.fast.ai

https://github.com/fastai/fastai