This project is aimmed at reproducing the results in Learning with Bad Training Data via Iterative Trimmed Loss Minimization.

• Pytorch
• Numpy
• scikit-learn

• Linear Regression: Tests on sample sizes & Test on alphas has been finished.

  • Results of test on sample sizes fit the original data pretty well (Figure provided).
  • Results of test on alphas are not so correct (Figure provided).
  • Test on "large noise" has not been tested as the definition of "large" noise need further investigation.

• WRN: There are some issues with training.

  • With Dataloader as input, the accuracy is around 50%.
  • With torch.Tensor as input (numpy data converte to torch.Tensor), the accuracy is only 16%.

Therefore, I did not produce any of the results from WRN.

Note:

• The sample sizes for Linear Regression do not exceed 10000 because of the possibility the process being killed on my computer.
• WRN is trained with 200 epoches.
• WRN source is from link

Detailed test code can be found in train_*.py