5.4 Reduction in warmup echos in SGD Optimiser
Case 3. Changing the value of warm-up echos to 2.5
  1. lr0 -> 0.1
  2. momentum ->0.92
  3. lrf ->0.2
  4. weight_decay: 0.0005
  5. warmup_epochs -> 2.5
  6. optimizer : SGD
Reverting back to the best momentum that we found and reduce the warmup epochs to 2.5 to see if performance can be improved or helps in achieving better results. The experiment was conducted with 30 epochs on the sub-sampled tiles of xView Dataset. The result achieved was at a score of PR AUC 0.109 at mAP@0.5 | F1 max 0.09 at 0.095 confidences and were able to improve upon the performance on a lower scale. Figure 8 depicts the score of F1 max 0.09 at 0.095 confidence rate which is comparatively good.