Improving mAP in Object Detection Task

[lcoandrade]

I'm working on an Object Detection Task. I'm using this Kaggle dataset, which I've converted to COCO.

I'm in the epoch number 13 so far, but the metrics I'm getting are quite low:

2023-09-21 02:06:20:rastervision.pytorch_learner.learner: INFO - metrics:
{'epoch': 13,
 'loss_classifier': 0.1262786090373993,
 'loss_box_reg': 0.1734391450881958,
 'loss_objectness': 0.09284689277410507,
 'loss_rpn_box_reg': 0.20699267089366913,
 'train_loss': 0.5995572805404663,
 'train_time': '0:05:27.749222',
 'mAP': 0.05325962360907719,
 'mAP50': 0.12719308604173504,
 'valid_time': '0:01:16.905776'}

In epoch 9 the metrics were better:

2023-09-21 01:23:34:rastervision.pytorch_learner.learner: INFO - metrics:
{'epoch': 9,
 'loss_classifier': 0.12986959517002106,
 'loss_box_reg': 0.1729908585548401,
 'loss_objectness': 0.10003477334976196,
 'loss_rpn_box_reg': 0.21598847210407257,
 'train_loss': 0.6188836693763733,
 'train_time': '0:05:46.995538',
 'mAP': 0.08408294456447805,
 'mAP50': 0.18128494012114718,
 'valid_time': '0:01:24.285483'}

What can I do to improve that? Consider that this is my configuration:

LR = 1e-4
IN_CHANNELS = 3
IMG_SIZE = (640,640)
BATCH_SIZE = 8
STRIDE = 640
DEVICE, NUM_DEVICES = 'cuda', 1
WORKERS = 2
BACKBONE = Backbone.resnet50

I processing using GPU P100 available in Kaggle.