In this experiment, we decrease the weight decay of our model to 0.000499. The result obtianed was AUC 0.111 at mAP@0.5 | F1 max 0.09 at 0.113 confidence conducted with 30 epochs on the sub-sampled tiles of xView Dataset. The experiment was successfully able to improve upon the performance. Figure 12 and Figure 13 shows the F1-Confidence and Precision- Recall curve respectively which depicts the score of AUC 0.111 at mAP@0.5 which is comparatively better result as compared to the results achieved. With the best weight decay of 0.000499anddecrease the warmup bias learning rate to 0.099 in order to verify if decreasing the warmup bias learning rate helps in getting better results. The result obtained was PR AUC 0.109 at mAP@0.5 | F1 max 0.08 at 0.099 confidence. The experiment was not able to improve upon the performance.
Fig 12. F1-Confidence curve for weight decay Fig 13. Precision – Recall curve for weight decay