Test for learning convergence for different tasks using overfit_batches

[dtronmans]

Purpose

• Tests for learning convergence using the overfit_batches argument, after 200 epochs with a batch size of 4 the loss should be under a certain threshold
• The threshold is determined by the loss formulation: higher for detection, instance segmentation and keypoint detection because the loss is composite
• Actually found out through these tests that we were wrongfully multiplying by batch_size in efficient_keypoint_bbox_loss. It was correctly learning and metrics were improving over time, but still this is an unnecessary multiplier

Specification

Dependencies & Potential Impact

Deployment Plan

Testing & Validation