├── app // For future implementation in
│ ├── Models
│ │ ├── __init__.py
│ │ ├── FeedForward.py // FeedForward model in PyTorch
│ │ ├── Perceptron.py // Perceptron model in pytorch
│ │ └── CNN.py // CNN model in PyTorch
│ ├── train
│ │ ├── train_CNN.ipynb // Training notebook for CNN
│ │ └── train_feedforward.ipynb // Training notebook for FeedForward network
│ ├── savedModels
│ │ ├── MNIST_CNN.pth // Trained CNN weights
│ │ ├── MNIST_FeedForward.pth // Trained FeedForward network weights
│ │ └── MNIST_Perceptron.pth // Trained perceptron weights
│ ├── utils
│ │ ├── __init__.py
│ │ └── camera.py
│ ├── CNNClassifierApp.py // GUI app code which loads CNN automatically
│ └── FeedforwardClassifierApp.py // GUI app code which loads FeedForward network automatically
├── LICENSE // License
├── assets
│ ├── out_vid1.mkv
│ ├── out_vid2.mkv
│ ├── output1.png
│ ├── output2.png
│ └── wrongop_1MNIST.png
└── README.md // This file
In this project I experimented a bit with different neural network architectures and saw how good each of the architectures were performing on MNIST dataset. Along with that I created a basic GUI application to take input from user and display the predicted number.
Different Architectures:-
I tried with a perceptron first but as you know the results were not very good. The accuracy was low and even with the data from MNIST it was not giving accurate results.
Next I tried a feed forward neural network with one hidden layer. The results were better than before i.e I got an accuracy of 80% within the first few epochs. When I tested it on the MNIST test set I got around 82% accuracy which is pretty good. There is one problem. The FeedForward neural networks do not take the shape into consideration. They just see the pixel intensities at locations and produces an output. Now assume a senario where 3 was written in such a way that it is covering all the activated parts for 8, then the model can be fooled easily.
Then I implemented a CNN architecture and still got a ~83% accuracy within a few epochs. Now when I tried it on custom that I made I got good results. The reason for this is that CNN takes into consideration the shapes i.e it doesnot just take the intensities into consideration, it also extracts features and gives results based on them. The convolutional layers create feature maps that record a region of image which is ultimately broken into rectangles and sent out for nonlinear processing. The only problem with CNN is that they need a lot of data to get trained but here as MNIST dataset has a lot of images we have no issues.
To get a local copy up and running follow these simple steps.
I strongly recommend the use of miniconda to create virtual environments
Install the following python modules:-
Thats it 😄
1st activate your environment. Then follow the given steps.
git clone https://github.com/SravanChittupalli/DigitClassifier-App-using-PyTorch.git
cd DigitClassifier-App-using-PyTorch/app
python CNNClassifierApp.py
Now you can find a GUI app on your screen. Play with it and let me know if there are any problems by creating an issue
Distributed under the MIT License. See LICENSE for more information.