search

fizyr / keras-retinanet

Keras implementation of RetinaNet object detection.
Apache License 2.0
4.38k stars 1.96k forks source link

Low mAP scores during training on COCO: from 0.01 to 0.1 during 60 epochs #1467

Closed Dawars closed 4 years ago

Dawars commented 4 years ago

I've been training the model with default parameters for 60 epoch now on COCO (2017) and I'm not getting the expected performance

The command I used: python keras_retinanet/bin/train.py --tensorboard-dir ./logs_coco_vanilla --gpu 0 --steps 10000 --compute-val-loss --epochs 100 --batch-size 1 coco /workspace/datasets/coco

  1. The mAP scores start very low (0.01) and after 60 epochs only reach 0.1-0.2
  2. At epoch 30 the gap between the train/val loss starts to widen which I take to be the beginning of overfitting.

Screenshot from 2020-10-01 09-37-53 Screenshot from 2020-10-01 09-38-13

I'm training on a GTX 1080 Ti.

Is this normal? If now, what could be the cause?

pratibhashinde commented 3 years ago

I am also getting low mAP. first epoch's mAP is 0.0017. I am using ResNet50 backbone on my own dataset. I did not make any changes in the program. Is it normal or do I need to make some changes?