Non-Max Supression Issue

[MohitDAngrish]

After training the detectron faster_rcnn_R_50_C4_3x for a certain usecase with the given below config :

cfg = get_cfg()
cfg.merge_from_file(model_zoo.get_config_file("COCO-Detection/faster_rcnn_R_50_C4_3x.yaml"))
cfg.DATASETS.TRAIN = ("my_dataset_train_new",)
cfg.DATASETS.TEST = ("my_dataset_test_new",)
cfg.DATALOADER.NUM_WORKERS = 4
cfg.MODEL.WEIGHTS = model_zoo.get_checkpoint_url("COCO-Detection/faster_rcnn_R_50_C4_3x.yaml")  # Let training initialize from model zoo
cfg.SOLVER.IMS_PER_BATCH = 4
cfg.SOLVER.MAX_ITER = 60000 #adjust up if val mAP is still rising, adjust down if overfit
cfg.SOLVER.WARMUP_ITERS = 1000
cfg.SOLVER.LR_SCHEDULER_NAME = 'WarmupCosineLR'
cfg.SOLVER.STEPS = []
cfg.OUTPUT_DIR = '/content/drive/MyDrive/ObjectDetectionDetectron/Models/faster_rcnn_R_50_C4_3xBestModelVal'
cfg.SOLVER.GAMMA = 0.05
cfg.MODEL.ROI_HEADS.BATCH_SIZE_PER_IMAGE = 64
cfg.MODEL.ROI_HEADS.NUM_CLASSES = 8
cfg.TEST.EVAL_PERIOD = 20
cfg.MODEL.ROI_HEADS.NMS_THRESH_TEST= 0.6 

Train loss is reducing upto 0.3 but the val loss is still at 0.9. Model seem to be overfitting. But that's not the case when i try doing predictions the output seemed reasonable but there was an issue of overlapping boxes as you can see below.

So here are my two questions:

1. If NMS and IOU is not working why the train loss is so low.
2. If NMS and IOU is working can somebody please point me towards the code where the non max suppression and iou is handled while training and evaluation of a model is done.