memory_probe_ops

TensorFlow kernels for probing memory.

Usage:

1. Copy linux.memory_probe_ops.so or macos.memory_probe_ops.so locally and rename to memory_probe_ops.so
2. Make sure this file is in current directory, or add its location to $PATH
3. Run test_memory_probe.py
4. Check memory-probe-examples.ipynb for examples

On Linux you can test memory probe ops as follows:

# install memory_probe_ops

import urllib.request
response = urllib.request.urlopen("https://github.com/yaroslavvb/memory_probe_ops/raw/master/linux.memory_probe_ops.so")
open("memory_probe_ops.so", "wb").write(response.read())

import tensorflow as tf

memory_probe_ops = tf.load_op_library("./memory_probe_ops.so")
print("Memory usage: ")
sess = tf.Session()
print(sess.run(memory_probe_ops.bytes_in_use()))

Troubleshooting

• Getting tensorflow.python.framework.errors.NotFoundError: ./memory_probe_ops.so: invalid ELF header

Make sure you got the correct version of .so (Linux vs Mac)

Building

On MacOS

TF_INC=$(python -c 'import tensorflow as tf; print(tf.sysconfig.get_include())')
g++ -std=c++11 -undefined dynamic_lookup -shared memory_probe_ops.cc -o memory_probe_ops.so -fPIC -I $TF_INC -O2

On Linux

TF_INC=$(python -c 'import tensorflow as tf; print(tf.sysconfig.get_include())')
g++ -std=c++11 -shared memory_probe_ops.cc -o memory_probe_ops.so -fPIC -I $TF_INC -O2

Using Bazel

Put BUILD file with the following under tensorflow/core/user_ops/

tf_custom_op_library(
    name = "memory_probe_ops.so",
    srcs = ["memory_probe_ops.cc"],
)

Now run

bazel build --config=cuda --config=opt //tensorflow/core/user_ops:memory_probe_ops.so

The .so file is dropped under bazel-bin/tensorflow/core/user_ops/memory_probe_ops.so

Note

A similar op for max_bytes_in_use has been integrated into TensorFlow in ccf9a752