search

rykov8 / ssd_keras

Port of Single Shot MultiBox Detector to Keras
MIT License
1.1k stars 553 forks source link

Re-implemented SSD has high loss and low acc during training #158

Open Yurodidon opened 4 years ago

Yurodidon commented 4 years ago

I implemented the SSD follow this project, but when I was training on PASCAL VOC2007 dataset, the acc became smaller and smaller, at the same time the loss is around 4.0.

The codes are available on https://github.com/CeLaMbDa/keras_SSD

Ryo-Kawanami commented 4 years ago

I'm also same situation... I have trained during 16 epoch, but loss, acc, val_loss, val_acc are almost same.

Ryo-Kawanami commented 4 years ago

Could anybody who knows this situation help?

Yurodidon commented 4 years ago

Hi Ryo-Kawanam, I think I found the reason of this situation! Check these:

  1. Do not use metrics in training, I don't know why, but still, when I delete those metrics, the network works much better.
  2. You should compile model after freezing layers, this was pointed out in the official document, please refer to https://keras.io/getting-started/faq/#how-can-i-freeze-keras-layers
  3. Use a fit learning rate, the can't be to small or too big, if it is too small it might lead the weights can't change.