MNIST web

A website for testing neural networks on MNIST data set using 2 hidden layers.

Required Software

• Node.js
• PyTorch

Instructions

1. Run node --version on your cmd. If not found, go to here and download Node.js.

2. Check if you have PyTorch installed with this script:

   import torch
   print(torch.__version__)

3. Install required packages(while in the directory)

   npm install

4. Run the server by typing node server.js in the cmd.

5. Your server is now ready for requests.
   In the browser type http://localhost:8000/, fill the form and send.

6. Wait for network's results.

Syntax

General Settings

• neural_net - The neural net you want to use out of {"Basic", "Dropout", "Batch_norm", "Combine"}.

• epochs - The number of data iterations.

• learning_rate - The network's learning rate {small values like 0.01, 0.005, 0.001}.

• batch_size - Iterating the data using batch_size number of samples each iteration {normally 64}

• valid_split - Splits your train data to validation and training and evalutes the network after each epoch {ranges from 0 to 1}.

Structure Settings

• hidden1_size - The number of neurons in the first hidden layer {default is 100}.

• hidden2_size - The number of neurons in the second hidden layer {default is 50}.

Options

• write_test_pred - Writes predictions to file 'test.pred', number to represent a boolean(due to a bug in passing a boolean from node.js to python) {0, 1}.

• draw_loss_graph - Draws loss graph of training and validation, number to represent a boolean(due to a bug in passing a boolean from node.js to python) {0, 1}.