baidu salers logo prediction

---

Basic info

• This project is based on gluon or a more famous parent: mxnet.

Project code structure

• dataloader: data preprocess
  • dataPreProcess.py
    • 80% of dataset used for train, 20% of dataset used for valid.
  • augmentation.py
    • add padding, rotate
    • ColorJitterAug: 0.3
    • random gray, prob: 0.5
    • calc dataset mean and std（utils/calc_mean_std.py），used for dataset normalize
• logs
  • training logs
• model
  • define model
• utils
  • calc_mean_std.py
    • calc dataset mean and std，used for dataset normalize
• weights
  • save model weights
• train.py
  • train model code
  • change pretrianed_model_name for different pretrained models, pretrained models
• predict.py
  • predict code for test and valid
• predict_wrong_ana.py
  • predict valid dataset label, find what's wrong with model
  • used for find best concat params
• concateResult.py
  • embedding model predict result

Future Work

• valid data can join train, use cv for valid
• mixup：BEYOND EMPIRICAL RISK MINIMIZATION
• pseudo label
• more augmentation
• more init methods