Phase II:
- A neural network with 62% accuracy on test set
- A simple flask API to consume the model and perform prediction
- Dockerization of the flask app
Phase III:
-
Moved 62% accurate Model and Flask Application to AWS Cloud , You can use this curl Request below here
curl --location --request POST 'https://dspd-service.lhth73asbhl1c.us-west-2.cs.amazonlightsail.com/predict' --form 'file=@"sad3.jpeg"' -
Trained a ResNet-50 model with 70% accuracy. The Notebook for that model is also attached
-
Performed Quantization using tflite which reduced the size by 10x.
-
Performed Pruning and reduced the model size by 50% but the accuracy got reduced from 70% to 54%.
Phase IV:
- Interpretability - Create a report on how various iterations of the model improved over time and in which areas it lacks including confusion matrix etc
- Polish flask application: Improve the business logic by integrating database and exposing end to end APIs as was mentioned in the Vision doc
- Test model generalization on different data sets specially including minority and under respresented groups.
- Iterate on model quality to strike a reasonable accuracy to size balance.
- Expose the final model through the flask application already on cloud
What we did in phase IV:
- Refined Business logic
- Integrated Databases
- Uploaded new end points to the server
- Uploaded more accurate model i.e 70%
- Created User Application in iOS.
Doesn't include the training and test data to reduce repo size, can be furnished upon request
Group members: Daniyal Raza 19839 Babar Shamsi 19840 Aisha Ghori Pathan 18297