• Final Presentation : https://docs.google.com/presentation/d/1TAuehsTdoA2fRd5p711V0sqtGuEJubtq2tC9DbNnp-A/edit?usp=sharing
• This Code is tested on MNIST Dataset.

Aryan Karn - 20185106 MNNIT-Allahabad ECE

Doubly Convolutional Neural Networks (NIPS 2016) by Shuangfei Zhai, Yu Cheng, Weining Lu and Zhongfei (Mark) Zhang

• Final Project contains main DCNN code written in python
• Forword Pass contains Matlab code of forword pass for understanding and proof of concept
• Correlation contains python code to extract weights from caffee model, matlab code to find averaged max transalation correlation of a layer, ploted results
• Presentation Raw files, images, graphs for Presentation

Code is written in python and would require following libraries:

• numpy
• theano (tensor)
• lasagne

Our Code is tested on Windows10 intel(i7)64 and MaC without CUDA but should run on any OS satisfying above pre-requisites.

Default Parameters:

num_epochs = 100
learning_rate = 1e-2 
metafilter_shape = [(2, 1, 6, 6), (4, 2, 6, 6)]
image_shape = (1, 28, 28)
kernel_size = 5
kernel_pool_size = 2
learning_decay = 1e-5
dropout_rate = 0.5
batch_size = 200

These parameters can be found and changed just below main function declaration.

You can run the code using following command inside final project.

sudo python main_final.py

Note: sudo access is required is to write results in a file

• Doubly Convolutional Neural Networks

• Getting Started with CNN Lasagne

• Lasagne Docs

• Theano Docs

• Final Presentation