search

Hexmagic / LaneDetection

百度车道线比赛
5 stars 1 forks source link

loss #2

Closed kaishijeng closed 3 years ago

kaishijeng commented 3 years ago

From train.py, it seems loss function is sum of dice loss and focus loss. Is this true? Do you have tried training with dice loss or focus loss only?

Thanks,

Hexmagic commented 3 years ago

single focal loss not worked well

kaishijeng commented 3 years ago

Thanks for an answer. The trained model (myself or yours) doesn't perform very well on my own videos. My trained model resolution is at stage 2. Not able to train with stage 3 resolution due to GPU memory. Any idea to improve this?

Thanks,

Hexmagic commented 3 years ago
  1. use free colab get more gpu memory
  2. simplfy model architecture or use other model ,I suggest you have a look at https://github.com/qubvel/segmentation_models.pytorch
  3. change image preprocessing
  4. 16 bit training ,reference to https://github.com/PyTorchLightning/pytorch-lightning,[this worked on some gpu ]