Tensorflow-implementation-of-Restricted-Boltzmann-Machine-including-GBRBM-BBRBM-

The application of RBM through Matlab have several drawbacks. One serious problem is that the computation time is very time-consuming. It is because that the matlab is difficult to run in cloud server.So I wrap the code of rbm based on matlab to python based on tensorflow. We can use jupyter notebook to run the code for training or inference quickly.

test RBM applications:

import numpy as np

import tensorflow as tf

import os,sys

from rbm_py3 import rbm

import time

from matplotlib import pyplot as plt

from math import *

from tensorflow.examples.tutorials.mnist import input_data

sess = tf.InteractiveSession()

mnist = input_data.read_data_sets('MNIST_data',one_hot=True)

data = mnist.train.next_batch(50000) data_x = data[0] data_y = data[1]

my_rbm = rbm(784,100, learning_rate=0.01, momentum=0.8, rbm_type='gbrbm', relu_hidden = True, init_method = 'uniform' )

my_rbm.plot=True my_rbm.pretrain(data_x,batch_size=100,n_epoches=10)

my_rbm.save_model_weights()

restore_path = os.getcwd() re_rbm = rbm(784,100,learning_rate=0.01,momentum=0.8,rbm_type='gbrbm',relu_hidden = True) re_rbm.restore_weights(restore_path,'rbm')

x_up = my_rbm.rbmup(data_x)

test DBN applications:

import numpy as np

import tensorflow as tf

import os,sys

from rbm_py3 import rbm

import time

from matplotlib import pyplot as plt

from math import *

from dbncreateopts import dbncreateopts

from DBN import dbn

from tensorflow.examples.tutorials.mnist import input_data

sess = tf.InteractiveSession()

mnist = input_data.read_data_sets('MNIST_data',one_hot=True)

data = mnist.train.next_batch(50000) data_x = data[0] data_y = data[1]

opts = dbncreateopts('gbrbm',50,100)

dbn_test = dbn(n_in = 784,hidden_layers_sizes=[100,100],opts=opts) dbn_test.train(data_x)

rbm_0 = dbn_test.rbm_list[0] rbm_1 = dbn_test.rbm_list[1] fig = plt.figure() ax0 = fig.add_subplot(2,1,1);ax0.plot(rbm_0.train_error) ax1 = fig.add_subplot(2,1,2);ax1.plot(rbm_1.train_error)

references:

https://github.com/meownoid/tensorfow-rbm https://github.com/lyy1994/generative-models/blob/master/RBM/RBM.ipynb http://deeplearning.net/tutorial/rbm.html