0

Am training Mnist.csv for robustness adding Gaussian noise using python random library,but how will I decide on the mean and std of noise to be added to the dataset.Am using a standardized data with 785 (28281) column and training for fog,brightness and stride.

Navya
  • 9
  • 2
  • Be precise about your question, otherwise nobody will understand what you mean, if you are looking for a way to stabilize training a GAN then go to this [repo](https://github.com/MrForExample/Generative_Models_Collection) – Mr. For Example Dec 10 '20 at 04:52
  • Would like to more details :So my Mnist dataset has 785 columns and 1500 rows and I need to train my model using SVM for fog correction,am planning to do data augmentation with each data point as x = x+ N(0,sigma),my question is how do I decide on the sigma value. – Navya Dec 10 '20 at 13:03
  • This is hyperparameters turning problem, no clear theory to tell you what the optimal value is, usually we won't set std of gaussian noise too high when data augmentation, and we just try it out for a good value, but some people won't even bother to try, just set to a value where augmented images look not too different from the original images, This [article](https://medium.com/bethgelab/increasing-the-robustness-of-dnns-against-image-corruptions-by-playing-the-game-of-noise-4566b5c2c8d5) talk about this if you like to see more – Mr. For Example Dec 11 '20 at 01:33
  • Got it thank you Mr.For Example. – Navya Dec 13 '20 at 01:54
  • No problem, Glad to help – Mr. For Example Dec 13 '20 at 01:56

0 Answers0