MNSIT Digits using PyTorch

The following repository contains files that have been tested to try to achieve an accuracy of 99.4% in more than one epoch, while keeping the model under 8k parameters

Code Iterations:

• S7- Model 1.ipynb :

  Target:

  • Get the Setup Right
  • Get a pretty decent accuracy, don't worry about number of params right now
  • Make sure the data loaders are working
  • Setup the basic training loop
  • Create a Baseline Model
  • 15 Epoch

  Result:

  • Parameters : 25,648
  • Best Training Accuracy : 99.61
  • Best Test Accuracy : 99.07

  Analysis:

  • The model is a bit large compared to our target model of having under 8k parameters
  • The Model is Overfitting

  S7- Model 2.ipynb :

  • Target:

    • Reduce the number of parameters
    • Add regularization techniques like Batch Normalization and Dropout to prevent overfitting

    Result:

    • Parameters : 12,100
    • Best Training Accuracy : 99.15
    • Best Test Accuracy : 99.11

    Analysis:

    • The model is not overfitting, but we need to reduce the number of params, since our target is to make the model under 8k Params
    • The Model seems to be generalizing well

• ##S7- Model 3.ipynb :

• ##S7- Model 4.ipynb :

• ##S7- Model 5.ipynb :