RBM procedure using tensorflow

rbm_MNIST_test.py

import tensorflow as tf
import numpy as np
import input_data
import Image
from util import tile_raster_images


def sample_prob(probs):
    return tf.nn.relu(
        tf.sign(
            probs - tf.random_uniform(tf.shape(probs))))

alpha = 1.0
batchsize = 100

mnist = input_data.read_data_sets("MNIST_data/", one_hot=True)
trX, trY, teX, teY = mnist.train.images, mnist.train.labels, mnist.test.images,\
    mnist.test.labels

X = tf.placeholder("float", [None, 784])
Y = tf.placeholder("float", [None, 10])

rbm_w = tf.placeholder("float", [784, 500])
rbm_vb = tf.placeholder("float", [784])
rbm_hb = tf.placeholder("float", [500])
h0 = sample_prob(tf.nn.sigmoid(tf.matmul(X, rbm_w) + rbm_hb))
v1 = sample_prob(tf.nn.sigmoid(
    tf.matmul(h0, tf.transpose(rbm_w)) + rbm_vb))
h1 = tf.nn.sigmoid(tf.matmul(v1, rbm_w) + rbm_hb)
w_positive_grad = tf.matmul(tf.transpose(X), h0)
w_negative_grad = tf.matmul(tf.transpose(v1), h1)
update_w = rbm_w + alpha * \
    (w_positive_grad - w_negative_grad) / tf.to_float(tf.shape(X)[0])
update_vb = rbm_vb + alpha * tf.reduce_mean(X - v1, 0)
update_hb = rbm_hb + alpha * tf.reduce_mean(h0 - h1, 0)

h_sample = sample_prob(tf.nn.sigmoid(tf.matmul(X, rbm_w) + rbm_hb))
v_sample = sample_prob(tf.nn.sigmoid(
    tf.matmul(h_sample, tf.transpose(rbm_w)) + rbm_vb))
err = X - v_sample
err_sum = tf.reduce_mean(err * err)

sess = tf.Session()
init = tf.initialize_all_variables()
sess.run(init)

n_w = np.zeros([784, 500], np.float32)
n_vb = np.zeros([784], np.float32)
n_hb = np.zeros([500], np.float32)
o_w = np.zeros([784, 500], np.float32)
o_vb = np.zeros([784], np.float32)
o_hb = np.zeros([500], np.float32)
print sess.run(
    err_sum, feed_dict={X: trX, rbm_w: o_w, rbm_vb: o_vb, rbm_hb: o_hb})

for start, end in zip(
        range(0, len(trX), batchsize), range(batchsize, len(trX), batchsize)):
    batch = trX[start:end]
    n_w = sess.run(update_w, feed_dict={
                   X: batch, rbm_w: o_w, rbm_vb: o_vb, rbm_hb: o_hb})
    n_vb = sess.run(update_vb, feed_dict={
                    X: batch, rbm_w: o_w, rbm_vb: o_vb, rbm_hb: o_hb})
    n_hb = sess.run(update_hb, feed_dict={
                    X: batch, rbm_w: o_w, rbm_vb: o_vb, rbm_hb: o_hb})
    o_w = n_w
    o_vb = n_vb
    o_hb = n_hb
    if start % 10000 == 0:
        print sess.run(
            err_sum, feed_dict={X: trX, rbm_w: n_w, rbm_vb: n_vb, rbm_hb: n_hb})
        image = Image.fromarray(
            tile_raster_images(
                X=n_w.T,
                img_shape=(28, 28),
                tile_shape=(25, 20),
                tile_spacing=(1, 1)
            )
        )
        image.save("rbm_%d.png" % (start / 10000))

The print result of reconstruction error under my attempt is

0.481194
0.286333
0.152281
0.149891
0.149818
0.147523
0.147973

First, here's the appropriate weight visualization from DeepLearnToolbox with same configuration.

And here are the six image files. Apparently, there's something wrong with the code, but I don't know how to fix.

So, now I've implemented correctly. The code issue of previous revision is the wrong use of assign and the missing of divide by batchsize for rbm_w. Now I've confirmed that the final result is right. See the weights visualization below.

Current revision is far away from using rbm through tensorflow in practice.

1. There are redundancy in the computation of data flow graph. (calculating rbm_w, rbm_vb, rbm_hb)
2. There are too much context switching between GPU calculation and memory operation (the feed of o_w, and assign of o_w).

Thank you, a nice and clean example!

Google suggests disabling automatic calculation of gradients. I'm not sure, but I believe the gradients are calculated automatically through out the calculation steps to be used later in the backpropagation (and not lazily only if backpropagation is really done). As the gradients are calculated manually here, the automatic calculation should be switched off.

https://www.tensorflow.org/versions/r0.7/api_docs/python/train.html#stop_gradient

Thank you for share. This is a great example for doing customized update on tensorflow.

Wow that is great example THANK YOU. Do you know how to remove the redundancy in graph?

Im trying to implement pretraining of autoencoders with rbm:
https://github.com/Cospel/rbm-ae-tf

Share my code here, everything was done in one session.run()
https://github.com/hanhongsun/tensorflow_script/blob/master/rbm.py

Line 29 is wrong: h1 = tf.nn.sigmoid(tf.matmul(v1, rbm_w) + rbm_hb) should be h1=sample_prob(tf.nn.sigmoid(tf.matmul(v1, rbm_w) + rbm_hb))

你的sample_prob没有按照概率 采样吧。
your function named "sample_prob", is not a sample by fixed-probability.