Skip to content

Instantly share code, notes, and snippets.

@usholanb
usholanb / rnn.py
Created August 3, 2016 20:09
Python RNN with numpy only
#############################################################################
# Abstract: This code trains a model to predict next character based on
# the previous ones.
#
# More details:
# 1. Each character is represented as a vector of size 256. It is all
# zeros except the index where that characters stands in ASCII table.
# 2. Code divides the text on chunks of input_size.
# 3. The teaching labels of all the characters in a chunk are the next
# characters, so we just shift the chuck to the right and assign the
@usholanb
usholanb / nn.py
Created August 3, 2016 20:08
Python NN with numpy only (MNIST dataset)
#####################################################################################
# This code trains a model that predicts which number from 0 to 9 is drawn on the
# MNIST pictures of 20x20 pixels. There are 60000 training examples and 10000
# testing examples.
# The methods used: Mini-batching, Weight Decay, Momentum, Dropout,
# Xavier's Initialization
# Number of hidden layers: 3 ( Adjustable )
# Sizes of mini-batches and hidden layer are easily adjustable
# Non-Linear functions: Sigmoid, Funny Tanh
# Linear: ReLU