What is this repo about?

The goal is to create NN1, NN2, and NN4 neural networks in order to predict if a handwritten digit is even or odd.

This implementation uses only numpy library for support. No other library used.

NN1 and NN2 are plain neural network structures.
NN4 uses batchnorm layer as well as dropout layer.

Implementation

Each models NN1 and NN2 contains layers and W. You can store weights in layers.W or W.

Code Structure

├── mlcvlab
│   ├── models # header file of add2 cuda kernel
|   |   ├── base.py
|   |   ├── nn1.py
|   |   ├── nn2.py
|   |   └── nn4.py
|   ├── nn
|   |   ├── activations.py
|   |   ├── basis.py
|   |   ├── batchnorm.py
|   |   ├── dropout.py
|   |   └── losses.py
|   └── optim
|       ├── adam.py
|       ├── async_sgd.py
|       ├── sgd.py
|       └── sync_sgd.py
|    
├── HW1_MNIST_NN1.ipynb
├── HW1_MNIST_NN1.py
├── HW1_MNIST_NN2.ipynb
├── HW1_MNIST_NN2.py
└── README.md

NN4 Backprop Images

Here is how I did the backprop for NN4 with batchnorm and dropout. Code for backprop is based on these formulas.

Run test cases

To run the test cases for activations.py file, run the following command and all tests should pass

$python test_activations.py

                TEST_SIGMOID_1 : True
                TEST_SIGMOID_2 : True
           TEST_SIGMOID_GRAD_1 : True
           TEST_SIGMOID_GRAD_2 : True
                TEST_SOFTMAX_1 : True
           TEST_SOFTMAX_GRAD_1 : True
                   TEST_TANH_1 : True
              TEST_TANH_GRAD_1 : True
                     TEST_RELU : True
                TEST_RELU_GRAD : True


TEST_ACTIVATIONS : True

To run the test cases for basis.py file, run the following command and all tests should pass

$python test_basis.py

                 TEST_LINEAR_1 : True
            TEST_LINEAR_GRAD_1 : True
                 TEST_RADIAL_1 : True
            TEST_RADIAL_GRAD_1 : True


TEST_BASIS : True

To run the test cases for losses.py file, run the following command and all tests should pass

$python test_losses.py

                 TEST_L2 : True
                TEST_L2_GRAD_1 : True


TEST_LOSSES : True

NOTE: Test cases are to be added for cross entrophy and cross entrophy grad.

NikolaAndro/MLCV_Package