Skip to content

Latest commit

 

History

History
44 lines (43 loc) · 2.03 KB

README.md

File metadata and controls

44 lines (43 loc) · 2.03 KB

ClassificationCNN

  • This neural network is intended for the parking lot app
  • The current network that we will test is found from this website https://www.tensorflow.org/tutorials/images/deep_cnn
    • Origionally it classified images into the following categories: airplane, automobile, bird, cat, deer, dog, frog, horse, ship and truck
      • Remarks:
        • it does not have a category that will do pickup trucks ( if that becomes a problem we will work on a different network)
        • automobiles do not overlap with trucks
    • We have discarded all categories except automobile and truck
      • We will either add a category called not car or output that the spot is empty when the system is not very confident there is a car
    • It also has a peak performance of 86% accuracy but that might improve in our case
    • The model follows the architecture described by Alex Krizhevsky, with a few differences in the top few layers.
    • The layers are currently (in order):
      • convolutional layer
      • max pooling layer
      • normalization layer
      • convolutional layer
      • normalization layer
      • max pooling layer
      • fully connected layer with rectified linear activation
      • fully connected layer with rectified linear activation
      • softmax linear transformation to produce logits
    • We will train several networks to see what works best for our project
    • One network will be :
      • Convolutional layer
      • max pooling layer
      • normalization layer
      • fully connected layer with rectified linear activation
      • softmax linear transformation to produce logits
    • Another will be:
      • convolutional layer
      • fully connected layer with rectified linear activation
      • softmax linear transformation to produce logits
    • Another will be:
      • convolutional layer
      • max pooling layer
      • normalization layer
      • convolutional layer
      • fully connected layer with rectified linear activation
      • softmax linear transformation to produce logits
  • Things we will compare include:
  • accuracy
  • learning rate
  • total loss