This is the code for this video on Youtube by Siraj Raval as part of The Math of Intelligence course. A convolutional neural network implemented in pure numpy. It uses a MNIST-like dataset with about 30 alphanumeric symbols. The author trained a deep convolutional network using Keras and saved the weights using python's pickle utility. Only the the forward propagation code is rewritten in pure numpy (as opposed to Theano or Tensorflow as in Keras). Which lets us run the network as a demo via heroku. For backpropagation in numpy for a convnet see this

Live web app is here: Website

Dependencies are packaged in the flask folder, so this app does not have any external depencies. Run pip install -r requirements.txt to install them.

Install pip here.

to start the web app run python run.py . To start the notebook run jupyter notebook in terminal.

Install jupyter here.

Credits for this code go to greydanus. I've merely created a wrapper to get people started.