Skip to content

Instantly share code, notes, and snippets.

@myme5261314
Last active April 10, 2020 06:59
  • Star 28 You must be signed in to star a gist
  • Fork 7 You must be signed in to fork a gist
Star You must be signed in to star a gist
Save myme5261314/005ceac0483fc5a581cc to your computer and use it in GitHub Desktop.
RBM procedure using tensorflow
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))
@myme5261314
Copy link
Author

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.

DeepLearnToolbox

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

rbm_0
rbm_5

@myme5261314
Copy link
Author

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.

rbm_5

@myme5261314
Copy link
Author

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).

@keskival
Copy link

keskival commented Apr 3, 2016

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

@hanhongsun
Copy link

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

@Cospel
Copy link

Cospel commented Apr 13, 2016

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

@hanhongsun
Copy link

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

@madzhen
Copy link

madzhen commented Sep 11, 2016

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))

@jxyyjm
Copy link

jxyyjm commented May 4, 2017

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.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment