lingxuez / rbm_ae

My python implementation of RBM and denoising Autoencoder for pre-training

RBM, Autoencoder, Denoising Autoencoder and pretraining

This is for Assignment 2 of CMU 10-807 "Topics in Deep Learning" from Lingxue Zhu.

Data

The MNIST dataset is stored under directory data/, downloaded from the course website.

RBM

To train an RBM with a given dimension of hidden layer dim_hidden and number of steps for contrastive divergence k_CD, run

python trainRBM.py <dim_hidden> <k_CD>

The output will be written to RBMresults/, including the learned weights and bias terms. As an example, the current directory contains the output from the following command

python trainRBM.py 100 1

To generate new images using a Gibbs chain starting from random configuration of the visible variables, run

python sampleRBM.py

The default setting generates 100 new images using the Gibbs sampler with 1000 steps.

(Denoising) Autoencoder

To train a (denoising) autoencoder with a given dimension of hidden layer dim_hidden and probability of adding noise noise_probability, run

python trainAutoencoder.py <dim_hidden> <noise_probability>

When noise_probability is set to 0, no noise is introduced and a regular autoencoder is learned. The output will be written to Autoresults/, including the learned weights and bias terms. As an example, the current directory contains the output from the following command

python trainAutoencoder.py 100 0.25

Pretraining a Neural Network

To fine-tune an one-layer neural network to classify the 10 handwritten digits with unsupervised pretraining, run

python trainNN.py <dim_hidden> <method>

where method defines the pretraining method, being one of

1. RBM, using the results from RBM, the default is with k_CD=1.
2. Auto, using the results from autoencoder without adding noise.
3. DenoAuto, using the results from denoising autoencoder, the default is to use noise_probability=0.25.
4. Random, randomly initialize the weights without pretraining.

Note that the pre-trained weights must be under directory RBMresults/ or Autoresults/ when using the corresponding method(s). The current directory contains the pre-trained weights for 100 dimension of hidden layer, and is ready to run the following commands:

python trainNN.py 100 RBM

python trainNN.py 100 Auto

python trainNN.py 100 DenoAuto

python trainNN.py 100 Random

See the directory NNout/ for sample output.