How to debug a memory leak in TensorFlow

Use Graph.finalize() to catch nodes being added to the graph

The most common mode of using TensorFlow involves first building a dataflow graph of TensorFlow operators (like tf.constant() and tf.matmul(), then running steps by calling the tf.Session.run() method in a loop (e.g. a training loop).

A common source of memory leaks is where the training loop contains calls that add nodes to the graph, and these run in every iteration, causing the graph to grow. These may be obvious (e.g. a call to a TensorFlow operator like tf.square()), implicit (e.g. a call to a TensorFlow library function that creates operators like tf.train.Saver()), or subtle (e.g. a call to an overloaded operator on a tf.Tensor and a NumPy array, which implicitly calls tf.convert_to_tensor() and adds a new tf.constant() to the graph).

The tf.Graph.finalize() method can help to catch leaks like this: it marks a graph as read-only, and raises an exception if anything is added to the graph. For example:

loss = ...
train_op = tf.train.GradientDescentOptimizer(0.01).minimize(loss)
init = tf.initialize_all_variables()

with tf.Session() as sess:
    sess.run(init)
    sess.graph.finalize()  # Graph is read-only after this statement.

    for _ in range(1000000):
        sess.run(train_op)
        loss_sq = tf.square(loss)  # Exception will be thrown here.
        sess.run(loss_sq)

In this case, the overloaded * operator attempts to add new nodes to the graph:

loss = ...
# ...
with tf.Session() as sess:
    # ...
    sess.graph.finalize()  # Graph is read-only after this statement.
    # ...
    dbl_loss = loss * 2.0  # Exception will be thrown here.

Use the tcmalloc allocator

To improve memory allocation performance, many TensorFlow users often use tcmalloc instead of the default malloc() implementation, as tcmalloc suffers less from fragmentation when allocating and deallocating large objects (such as many tensors). Some memory-intensive TensorFlow programs have been known to leak heap address space (while freeing all of the individual objects they use) with the default malloc(), but performed just fine after switching to tcmalloc. In addition, tcmalloc includes a heap profiler, which makes it possible to track down where any remaining leaks might have occurred.

The installation for tcmalloc will depend on your operating system, but the following works on Ubuntu 14.04 (trusty) (where script.py is the name of your TensorFlow Python program):

$ sudo apt-get install google-perftools4
$ LD_PRELOAD=/usr/lib/libtcmalloc.so.4 python script.py ...

As noted above, simply switching to tcmalloc can fix a lot of apparent leaks. However, if the memory usage is still growing, you can use the heap profiler as follows:

$ LD_PRELOAD=/usr/lib/libtcmalloc.so.4 HEAPPROFILE=/tmp/profile python script.py ...

After you run the above command, the program will periodically write profiles to the filesystem. The sequence of profiles will be named:

• /tmp/profile.0000.heap
• /tmp/profile.0001.heap
• /tmp/profile.0002.heap
• …

You can read the profiles using the google-pprof tool, which (for example, on Ubuntu 14.04) can be installed as part of the google-perftools package. For example, to look at the third snapshot collected above:

$ google-pprof --gv `which python` /tmp/profile.0002.heap

Running the above command will pop up a GraphViz window, showing the profile information as a directed graph.