Deep learning from scratch

Implementing deep learning algorithms in python with numpy.

Dependencies

• Numpy
• Matplotlib (for Loss plots only)

Logistic Regression

Implementation -> logistic-regression.ipynb

Architecture

• ReLU in hidden layers and Sigmoid in the output layer.
• Number of layers and number of units in each layer can be set using layers_dims hyper-parameter.
• Uses sigmoid cross entropy for loss computation.
• Vectorized implementation.

Softmax Classification

Implementation -> softmax-classification.ipynb

Architecture

• ReLU in hidden layers and Softmax in the output layer.
• Number of layers and number of units in each layer can be set using layers_dims hyper-parameter.
• Uses softmax cross entropy for loss computation.
• Vectorized implementation

Vanilla char level RNN

Implementation -> vanilla-char-rnn.ipynb

Architecture

• tanh as activation for hidden state, Softmax at output.
• Adagrad optimization.
• Uses softmax cross entropy for loss computation.