https://github.com/kuleshov/deep-learning-models
Implementations of popular deep learning models in Theano+Lasagne
https://github.com/kuleshov/deep-learning-models
Last synced: about 1 year ago
JSON representation
Implementations of popular deep learning models in Theano+Lasagne
- Host: GitHub
- URL: https://github.com/kuleshov/deep-learning-models
- Owner: kuleshov
- License: mit
- Created: 2016-11-28T22:04:46.000Z (over 9 years ago)
- Default Branch: master
- Last Pushed: 2017-07-12T03:13:07.000Z (almost 9 years ago)
- Last Synced: 2025-04-06T08:04:33.621Z (about 1 year ago)
- Language: Python
- Homepage:
- Size: 81.1 KB
- Stars: 24
- Watchers: 3
- Forks: 20
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
- awesome-deeplearning-resources - Implementations of many popular deep learning models in Theano+Lasagne
README
Deep Learning Model Zoo
=======================
This repository contains implementations of various deep learning algorithms in Theano/Lasagne.
## Running a model
To run a model, you may use the `run.py` launch script.
```
python run.py train \
--dataset \
--model \
--alg \
--n_batch \
--lr \
-e \
-l
```
Alternatively, you may use the `Makefile` included in the root dir; typing `make train` will start training. There are also several additional parameters that can be configured inside the `Makefile`.
The model will periodically save its weights and report training/validation losses in the logfile.
## Algorithms
The following algorithms are available.
### Supervised learning models
* `softmax`: simple softmax classifier
* `mlp`: multilayer perceptron
* `cnn`: convolutional neural network; solves `mnist` and achieves reasonably good accuracy on `cifar10`
* `resnet`: small residual network; achieves an accuracy in the 80's on `cifar10`
### Semi-supervised models
* `ssdadgm`: semi-supervised deep generative models (in progress)
### Unsupervised models
* `vae`: variational autoencoder
* `convvae`: convolutional variational autoencoder
* `sbn`: vae with discrete latent variables, trained with neural variational inference (reduces to sigmoid belief network)
* `adgm`: auxiliary deep generative model (unsupervised version)
* `convadgm`: convolutional auxiliary deep generative model (unsupervised version)
* `dadgm`: discrete-variable auxiliary deep generative model (unsupervised version, also trained with NVIL)
* `dcgan`: small deep convolutional generative adversarial network (tested on mnist)
## Datasets
The following datasets are currently available:
* `cifar10`: color images divided into 10 classes (32x32x3)
* `mnist`: standard handwritten digits dataset (28x28)
* `digits`: sklearn digits dataset (8x8); can be used for quick debugging on a CPU
## Optimization methods
Currently, we may train the models using:
* `sgd`: standard stochastic gradient descent
* `adam`: the Adam optimizer
## Feedback
Send feedback to [Volodymyr Kuleshov](http://www.stanford.edu/~kuleshov). Some models contain snippets from other users' repositories; let me know if I forgot to cite anyone.