16

I was looking at the Tensorflow MNIST example for beginners and found that in this part:

for i in range(1000):
  batch_xs, batch_ys = mnist.train.next_batch(100)
  sess.run(train_step, feed_dict={x: batch_xs, y_: batch_ys})

changing the batch size from 100 to be above 204 causes the model to fail to converge. It works up to 204, but at 205 and any higher number I tried, the accuracy would end up < 10%. Is this a bug, something about the algorithm, something else?

This is running their binary installation for OS X, seems to be version 0.5.0.

Community
  • 1
  • 1
pythonic metaphor
  • 10,296
  • 18
  • 68
  • 110
  • In case you're suffering from the same bug as I did, you may want tot check here: http://stackoverflow.com/questions/33712178/tensorflow-nan-bug – user1111929 Feb 10 '16 at 19:03

4 Answers4

28

You're using the very basic linear model in the beginners example?

Here's a trick to debug it - watch the cross-entropy as you increase the batch size (the first line is from the example, the second I just added):

cross_entropy = -tf.reduce_sum(y_*tf.log(y))
cross_entropy = tf.Print(cross_entropy, [cross_entropy], "CrossE")

At a batch size of 204, you'll see:

I tensorflow/core/kernels/logging_ops.cc:64] CrossE[92.37558]
I tensorflow/core/kernels/logging_ops.cc:64] CrossE[90.107414]

But at 205, you'll see a sequence like this, from the start:

I tensorflow/core/kernels/logging_ops.cc:64] CrossE[472.02966]
I tensorflow/core/kernels/logging_ops.cc:64] CrossE[475.11697]
I tensorflow/core/kernels/logging_ops.cc:64] CrossE[1418.6655]
I tensorflow/core/kernels/logging_ops.cc:64] CrossE[1546.3833]
I tensorflow/core/kernels/logging_ops.cc:64] CrossE[1684.2932]
I tensorflow/core/kernels/logging_ops.cc:64] CrossE[1420.02]
I tensorflow/core/kernels/logging_ops.cc:64] CrossE[1796.0872]
I tensorflow/core/kernels/logging_ops.cc:64] CrossE[nan]

Ack - NaN's showing up. Basically, the large batch size is creating such a huge gradient that your model is spiraling out of control -- the updates it's applying are too large, and overshooting the direction it should go by a huge margin.

In practice, there are a few ways to fix this. You could reduce the learning rate from .01 to, say, .005, which results in a final accuracy of 0.92. 

train_step = tf.train.GradientDescentOptimizer(0.005).minimize(cross_entropy)

Or you could use a more sophisticated optimization algorithm (Adam, Momentum, etc.) that tries to do more to figure out the direction of the gradient. Or you could use a more complex model that has more free parameters across which to disperse that big gradient.

dga
  • 21,757
  • 3
  • 44
  • 51
16

@dga gave a great answer, but I wanted to expand a little.

When I wrote the beginners tutorial, I implemented the cost function like so:

cross_entropy = -tf.reduce_sum(y_*tf.log(y))

I wrote it that way because that looks most similar to the mathematical definition of cross-entropy. But it might actually be better to do something like this:

cross_entropy = -tf.reduce_mean(y_*tf.log(y))

Why might it be nicer to use a mean instead of a sum? Well, if we sum, then doubling the batch size doubles the cost, and also doubles the magnitude of the gradient. Unless we adjust our learning rate (or use an algorithm that adjusts it for us, like @dga suggested) our training will explode! But if we use a mean, then our learning rate becomes kind of independent of our batch size, which is nice.

I'd encourage you to check out Adam (tf.train.AdamOptimizer()). It's often more tolerant to fiddling with things than SGD.

colah
  • 416
  • 3
  • 6
  • But doesn't this change the whole thing we are trying to optimize. With values `[1, 72, 5]` the difference is pretty significant. – Salvador Dali Nov 15 '15 at 11:44
  • 1
    `sum` gives the number of bits/nits/digits needed to encode all the data. Taking the mean over all examples, gives the average number of bits required per example. But the version above takes the mean over both dimensions. An untrained (unopinionated) MNIST model should, unsirprisingly, require 1 digit per digit, or `log(10)~=2.3` nits per digit. Switch it to log10 if you want it in digits. If you want to keep the physical meaning use `cost = -tf.reduce_mean(tf.reduce_sum(self.y_*tf.log(self.y),1))`. that is: `sum` across classes, `mean` over items. – mdaoust Nov 17 '15 at 01:20
15

Nan occurs when 0*log(0) occurs:

replace:

cross_entropy = -tf.reduce_sum(y_*tf.log(y))

with:

cross_entropy = -tf.reduce_sum(y_*tf.log(y + 1e-10))
Muaaz
  • 261
  • 1
  • 7
2

@dga nicely explained you the reason of such behavior (the cross_entropy becomes too huge) and thus the algorithm will not be able to converge. There are a couple of ways to fix this. He already suggested to decrease the learning rate.

Gradient descent is the most basic algorithm. Almost all other optimizers will be working properly:

train_step = tf.train.AdagradOptimizer(0.01).minimize(cross_entropy)
train_step = tf.train.AdamOptimizer().minimize(cross_entropy)
train_step = tf.train.FtrlOptimizer(0.01).minimize(cross_entropy)
train_step = tf.train.RMSPropOptimizer(0.01, 0.1).minimize(cross_entropy)

Another approach is to use tf.nn.softmax_cross_entropy_with_logits which handles numeric instabilities.

Salvador Dali
  • 214,103
  • 147
  • 703
  • 753