Home

Awesome

Introduction

NeuThink is experimental deep learning library build on top of pytorch Neuthink is research project aimed to explore how the concept of differential programming can be implemented in context of Python language. One of definitions of differential programming is that it "enables programmers to write program sketches with slots that can be filled with behavior trained from program input-output data" ([Bošnjak et all, 2016] The goal of Neuthink is too see how we can express such sketches in a way that can simplify creating complex deep learning models for non-experts.

Motivational Example

Consider the simple classical task of classification of figures into fixed number of classes: Circles, Rectangles and Triangles (example dataset). Lets say, we want to build a program that takes the list of figure images and selects only Triangles. The sketch of such program may look like this (in pseudo-python):

list_of_figures = load_from_folder(folder_name)
filter(lambda x: x==“triangle“, map(get_shape, list_of_figures))

Here we specify that we need to load figures from some folder, apply some function get_shape to it and then filter only triangles. In this case the behavior that need to be learned from data get_shape() function. In ideal world we would be able to just specify dataset of known figures and tell to compiler to make this get_shape() somehow. In reality, however, there exist many details, that need to be specified, so with most popular deep learning frameworks one can not just do it - lot of other code need to written that takes care of data loading, model specification, model training and prediction for a new data. Neuthink aims to eliminate all that code, so examples like above could work. Here is real Neuthink-based code that is sufficient to solve the task:

import neuthink.metaimage as Image
#load data
images = Image.Load(«./train_figures»)
#specify sketch - resize image, change to grayscale and map from image to target class
images = images.Resize(scale=(32,32),source="image").Mode().dMap(target="target_class")
get_shape = images.compile()  #this actually trains the model, filling in missing behavior. this returns function get_shape

Test = Image.Load("./test") #load some test images
get_shape(Test) #the new function can be just applied to the data to get new predictions
Images.Export("get_shape") # we can export this new function for future use

# so all we will need to use it is just import
from get_shape import get_shape

Neuthink currently supports simple image classification models, text classification, sequence tagging and character level language models. Also, more complex models can be constructed by combining existing primitives.

Installation

IMPORTANT: THIS IS THE WORK IN PROGRESS AND RESEARCH PROJECT. NOT RECOMMENDED FOR PRODUCTION USE, POORLY DOCUMENTED AND CONTAINS KNOWN BUGS. USE FOR YOUR OWN RISK

No installation/setup script is avaliable at the the time, if you want to use the library, it need to be installed manually

  1. Install all dependencies