每10k-12k步骤使用ssd_mobilenet_v2_quantized_300x300_coco的对象检测模型的丢失增加

Question

I'm re-training ssd_mobilenet_v2_quantized_300x300_coco object detection model on a custom dataset. The dataset consists of approx 2.6k images and 19 classes. After the training step reaches 10k-12k the loss graph starts increasing. This happens even if I change my model to ssd_mobilenet_v2_coco and at the same step range. I couldn't find anything that is related to this behaviour in the config file. Also this disappers when using faster_rcnn models. When the issue arises the mAP becomes almost constant. Also tha accuracy doesn't go beyond 50%. Can anyone explain this behaviour ?

Sample Dataset:

Loss Graph

a) ssd_mobilenet_v2_quantized_300x300_coco

b) ssd_mobilenet_v2_coco

Config File: a) ssd_mobilenet_v2_quantized_300x300_coco

https://pastebin.com/BBwqEruK

b) ssd_mobilenet_v2_coco

https://pastebin.com/F8K0BhQV

Answer 1

What about your training loss? Notice that total_loss is the validation loss here.

If your training loss is decreasing while the validation loss is increasing, this is clearly a sign of overfitting, you may use regularization loss during training by adding the following in the config file, in part train_config

add_regularization_loss: true

just as batch_size: 24