/kaggle-planet

Code for the Kaggle competition "Planet: Understanding the Amazon from Space"

Primary LanguagePythonMIT LicenseMIT

Kaggle Planet competition

  1. This project is for the competition and is implemented in PyTorch.

  2. I used resnet-152 and got my best score 0.92563 on Kaggle. I have tried some model fusion tricks but it didn't work better.

Training tricks

  1. Firstly randomly crop the image. When the val loss doesn't decrease any more, cancel the crop setting for possible improvement of the model.

  2. The weight decay is used periodically.

  3. Use SGD + Momentum + Nesterov.

  4. When the validation loss decreases to very low, validate the model more frequently to find the best one.

  5. Initially use small batch size for fast convergence. Later on increase the batch size for more steady training.