• Stars
    star
    510
  • Rank 84,932 (Top 2 %)
  • Language
    Python
  • Created about 6 years ago
  • Updated over 5 years ago

Reviews

There are no reviews yet. Be the first to send feedback to the community and the maintainers!

Repository Details

MADE (Masked Autoencoder Density Estimation) implementation in PyTorch

pytorch-made

This code is an implementation of "Masked AutoEncoder for Density Estimation" by Germain et al., 2015. The core idea is that you can turn an auto-encoder into an autoregressive density model just by appropriately masking the connections in the MLP, ordering the input dimensions in some way and making sure that all outputs only depend on inputs earlier in the list. Like other autoregressive models (char-rnn, pixel cnns, etc), evaluating the likelihood is very cheap (a single forward pass), but sampling is linear in the number of dimensions.

figure 1

The authors of the paper also published code here, but it's a bit wordy, sprawling and in Theano. Hence my own shot at it with only ~150 lines of code and PyTorch <3.

examples

First we download the binarized mnist dataset. Then we can reproduce the first point on the plot of Figure 2 by training a 1-layer MLP of 500 units with only a single mask, and using a single fixed (but random) ordering as so:

python run.py --data-path binarized_mnist.npz -q 500

which converges at binary cross entropy loss of 94.5, as shown in the paper. We can then simultaneously train a larger model ensemble (with weight sharing in the one MLP) and average over all of the models at test time. For instance, we can use 10 orderings (-n 10) and also average over the 10 at inference time (-s 10):

python run.py --data-path binarized_mnist.npz -q 500 -n 10 -s 10

which gives a much better test loss of 79.3, but at the cost of multiple forward passes. I was not able to reproduce single-forward-pass gains that the paper alludes to when training with multiple masks, might be doing something wrong.

usage

The core class is MADE, found in made.py. It inherits from PyTorch nn.Module so you can "slot it into" larger architectures quite easily. To instantiate MADE on 1D inputs of MNIST digits for example (which have 28*28 pixels), using one hidden layer of 500 neurons, and using a single but random ordering we would do:

model = MADE(28*28, [500], 28*28, num_masks=1, natural_ordering=False)

The reason we plug the size of the output (3rd argument) into MADE is that one might want to use relatively complicated output distributions, for example a gaussian distribution would normally be parameterized by a mean and a standard deviation for each dimension, or you could bin the output range into buckets and output logprobs for a softmax, or mixture parameters, etc. In the simplest example in this code we use binary predictions, where are only parameterized by one number, hence the number of the input dimensions happens to equal the number of outputs.

License

MIT

More Repositories

1

nanoGPT

The simplest, fastest repository for training/finetuning medium-sized GPTs.
Python
22,607
star
2

minGPT

A minimal PyTorch re-implementation of the OpenAI GPT (Generative Pretrained Transformer) training
Python
15,735
star
3

char-rnn

Multi-layer Recurrent Neural Networks (LSTM, GRU, RNN) for character-level language models in Torch
Lua
11,228
star
4

convnetjs

Deep Learning in Javascript. Train Convolutional Neural Networks (or ordinary ones) in your browser.
JavaScript
10,642
star
5

nn-zero-to-hero

Neural Networks: Zero to Hero
Jupyter Notebook
8,476
star
6

micrograd

A tiny scalar-valued autograd engine and a neural net library on top of it with PyTorch-like API
Jupyter Notebook
5,613
star
7

neuraltalk2

Efficient Image Captioning code in Torch, runs on GPU
Jupyter Notebook
5,426
star
8

neuraltalk

NeuralTalk is a Python+numpy project for learning Multimodal Recurrent Neural Networks that describe images with sentences.
Python
5,352
star
9

arxiv-sanity-preserver

Web interface for browsing, search and filtering recent arxiv submissions
Python
4,943
star
10

ng-video-lecture

Python
2,074
star
11

reinforcejs

Reinforcement Learning Agents in Javascript (Dynamic Programming, Temporal Difference, Deep Q-Learning, Stochastic/Deterministic Policy Gradients)
HTML
1,273
star
12

makemore

An autoregressive character-level language model for making more things
Python
1,217
star
13

cryptos

Pure Python from-scratch zero-dependency implementation of Bitcoin for educational purposes
Jupyter Notebook
1,142
star
14

randomfun

Notebooks and various random fun
Jupyter Notebook
996
star
15

ulogme

Automatically collect and visualize usage statistics in Ubuntu/OSX environments.
Python
941
star
16

recurrentjs

Deep Recurrent Neural Networks and LSTMs in Javascript. More generally also arbitrary expression graphs with automatic differentiation.
HTML
918
star
17

arxiv-sanity-lite

arxiv-sanity lite: tag arxiv papers of interest get recommendations of similar papers in a nice UI using SVMs over tfidf feature vectors based on paper abstracts.
Python
864
star
18

tsnejs

Implementation of t-SNE visualization algorithm in Javascript.
JavaScript
815
star
19

pytorch-normalizing-flows

Normalizing flows in PyTorch. Current intended use is education not production.
Jupyter Notebook
790
star
20

paper-notes

Random notes on papers, likely a short-term repo.
660
star
21

svmjs

Support Vector Machine in Javascript (SMO algorithm, supports arbitrary kernels) + GUI demo
JavaScript
636
star
22

karpathy.github.io

my blog
CSS
472
star
23

lecun1989-repro

Reproducing Yann LeCun 1989 paper "Backpropagation Applied to Handwritten Zip Code Recognition", to my knowledge the earliest real-world application of a neural net trained with backpropagation.
Jupyter Notebook
425
star
24

deep-vector-quantization

VQVAEs, GumbelSoftmaxes and friends
Jupyter Notebook
422
star
25

covid-sanity

Aspires to help the influx of bioRxiv / medRxiv papers on COVID-19
Python
351
star
26

find-birds

Find people you should follow on Twitter based on who the people you follow follow
Python
305
star
27

forestjs

Random Forest implementation for JavaScript. Supports arbitrary weak learners. Includes interactive demo.
JavaScript
284
star
28

researchlei

An Academic Papers Management and Discovery System
Python
194
star
29

Random-Forest-Matlab

A Random Forest implementation for MATLAB. Supports arbitrary weak learners that you can define.
MATLAB
172
star
30

researchpooler

Automating research publications discovery and analysis. For example, ever wish your computer could automatically open papers that are most similar to a paper at an arbitrary url? How about finding all papers that report results on some dataset? Let's re-imagine literature review.
Python
167
star
31

nipspreview

Scripts that generate .html to more easily see NIPS papers
Python
147
star
32

ttmik

Talk to me in Korean Anki cards and related scripts
Python
103
star
33

tf-agent

tensorflow reinforcement learning agents for OpenAI gym environments
Python
99
star
34

gitstats

A lightweight/pretty visualizer for recent work on a git code base in multiple branches. Helps stay up to date with teams working on one git repo in many branches.
HTML
85
star
35

EigenLibSVM

A wrapper for LibSVM that lets you train SVM's directly on Eigen library matrices in C++
C++
74
star
36

MatlabWrapper

C++ convenience class to communicate with a Matlab instance. Send matrices back and forth, execute arbitrary Matlab commands, or drop into interactive Matlab session right in the middle of your C++ code.
C++
52
star
37

twoolpy

useful scripts to work with Twitter + Python. Requires the tweepy library.
Python
50
star
38

notpygamejs

Game making library for using Canvas element
JavaScript
41
star
39

scholaroctopus

A set of tools/pages that help explore academic literature
33
star
40

karpathy

root repo
19
star