Feature : Feedback Loop for Ramp

[kshitijrajsharma]

Issue:

• Contributes to DISCUSSION : Apply Feedback Loop on the Predictions  #106
• Resolve Add Backend Model and CURD API for feedback #110

What does this PR DO ?

PR Implements the Awaiting feedback loop method for RAMP Basemodel !
PR Focuses Mainly on Backend changes to implement feedback method

• Introduces Feedback Model
• CURD API for Feedback
• Connects feedback data to training Methodology
• Basic Frontend to showcase backend flow ( Mapper view and Validator View )

Methodology
When new feedback is submitted backend picks up the loaded model , zoom level of prediction and base image when mapper is providing feedback , That Published model acts as base model where new feedback will be fit into . Initially Feedback will go back to manager/moderators of model's page . Moderators can see the feedbacks for their published model . They have option to validate good feedbacks while discarding the bad feedbacks ! Feedback loops runs with 1 batch size and 1 epoch by default to avoid overfitting . Maintains different version of model, it will be a new training instance which means manager can always go back to previous version of model if things doesn't work out after applying feedback !

Limitation :

• Feedbacks needs to be completed for the grid single feedback or couple of feedback can't be supplied to model , for eg : if model makes two bad prediction in area while submitting feedback full input for whole grid should be supplied Like this , if one feature is accepted /modified or deleted for feedback mapper needs to make sure all the yellow box is completed for feedback

How to test ?

• Test Feedback CURD API

Run predictions from published model , Submit feedbacks from Mapper view , Validate them from Manager View and Apply those to published model and publish new version of model with feedback ! **Mapper View :**

**Manager View :**

Currently Accepted Feedback :

• Missing Feature : Mapper can create new feature if missing
• Modify Feature : Mapper can correct if predictions are not correct / misaligned
• Delete Feature : Mapper can delete invalid outliers

TODO :

• Stabilize This Method
• Add good methods and direction to frontend