Michael Firman mdfirman

mdfirman / 2dto3d.py

Last active June 30, 2017 11:12

Training a 2d -> 3d network

	# 2D -> 3D bottleneck architecture

	import tensorflow as tf
	layers = tf.contrib.slim.layers


	# Network inputs
	x_in = tf.placeholder(tf.float32, shape=(None, 64, 64, 3))
	y_in = tf.placeholder(tf.float32, shape=(None, 32, 32, 32, 1))
	phase = tf.placeholder(tf.bool) # for batch norm - see http://ruishu.io/2016/12/27/batchnorm/

mdfirman / 0_reuse_code.js

Created April 11, 2017 09:52

Here are some things you can do with Gists in GistBox.

	// Use Gists to store code you would like to remember later on
	console.log(window); // log the "window" object to the console

mdfirman / lasagne_minibatch_discrimination.py

Created October 20, 2016 14:04

A lasagne layer to perform minibatch discrimination

	class MinibatchDiscrimination(lasagne.layers.Layer):
	'''
	A lasagne layer to perform minibatch discrimination [1]. The idea is to improve
	generative advesarial networks (GANs) by appending a feature to each input example
	which takes information from all the other items in the minibatch.
	This means that the generator, when it comes to creating a new minibatch, is less
	likely to fixate on one point. Instead it should be incentivised to generate
	examples covering the full output space.

	Notation and formulation as described in [1], section 3.2

mdfirman / codestats.py

Last active April 8, 2016 17:17

Script to recursively count lines of code in files in folder

	#!/usr/bin/python
	# Script which prints statistics about number of lines from all the files
	# recursively below the current folder.
	# Currently just looks at .py and .ipynb files, but could be trivially modified
	# to look at other file types.
	# Caveat: Includes comments, blank lines etc in the count.

	# Bash solution:
	# #! /bin/sh
	# find . -name '*.py' \| xargs wc -l

mdfirman / io_mat_large.py

Last active October 20, 2016 14:07

Saving python dictionaries containing arbitrarily large numpy arrays --- a hacky fix for python 2.7 problem

	import numpy as np
	import scipy

	def savemat_large(savepath, dic, modify_in_place=True, **kwargs):
	"""
	Save a python dictionaries containing arbitrarily large numpy arrays.
	Essentially a wrapper for scipy.io.savemat, but first numpy arrays which
	are too large are split up.

	This is a hacky fix for the problem that individual objects bigger than

mdfirman / merging_from_master

Created July 21, 2015 16:05

	# https://help.github.com/articles/merging-an-upstream-repository-into-your-fork/

	git pull https://github.com/groakat/engaged_hackathon.git master