Project: A simple exercise in neural networks

Description and Purpose:

Handwritten number recognition

Solution:

In this project, we are trying to build a simple neural network model so that handwritten numbers stored as images (28x28 pixels) in the mnist dataset can be recognized by the computer.

Code and Results:

• The main code is in src.

• In model_creation we built our model as a json dict file.

• Our model summary:

• Visualize Sample Training Data:

  We use below code to show some of our datas:

      for i in range(5):
      plt.subplot(1,5,i+1)
      plt.imshow(train_imag[i], cmap='gray')
      plt.title(train_labels[i])
      plt.axis('off')

  and the result is:

  

  Our plot losses code:

  losses = history.history['loss']
  val_losses = history.history['val_loss']
  
  plt.plot(losses)
  plt.plot(val_losses)
  plt.xlabel('Epochs')
  plt.ylabel('Loss')
  plt.legend(['loss', 'val_loss'])

  result:

  

• Evaluation on Test Data:

  Code:

  test_labels_p = my_model.predict(test_x)
  test_labels_p = np.argmax(test_labels_p, axis=1)
  
  n = 0
  f, axs = plt.subplots(1,10,figsize=(15,15))
  for i in range(len(test_labels)):
      if n >= 10:
    break
      if (test_labels_p[i] != test_labels[i]):
      axs[n].imshow(test_img[i], cmap='gray')
      axs[n].set_title(f'{test_labels[i]} -> {test_labels_p[i]}')
      axs[n].axis('off')
      n = n+1

  Result: