tfprof | Profiling Taskflow Programs through Visualization | Data Visualization library

I'll give it a try:

(1) From this example, looks like the first number is the "self" flops, the second number means the "total" flops under the naming scope. For example: for the 3 nodes respectively named random_uniform (if there is such a node), random_uniform/mul, random_uniform/sub, they respectively take 11.76k, 11.76k, and 1 flops, and in total 23.52k flops.

For another example: 23.83k = 23.52k + 300.

Does this make sense?

(2) The root node is a "virtual" top-level node added by the profiler, which doesn't have a "self" flops , or in other words, it has zero self flops.

(3) Not sure why it is 1. It would help if you can print the GraphDef and find out what this node really is, with print(sess.graph_def)

Hope this helps.

First of all, as of now, tfprof.model_analyzer.print_model_analysis is deprecated and tf.profiler.profile should be used instead according to the official documentation.

Given that we know the number of FLOP, we can get the FLOPS (FLOP per second) of a forward pass by measuring run time of a forward pass and divide FLOP/run_time

Let's take an easy example.

' ... no flops stats due to incomplete shapes' means that you have unknown [shape] information for some variables, e.g. when you process variable batch size ([None] shape) or a tf.while_loop for arbitrary amount of time, etc.

Accordingly to official tfprof docs (source):

• It must have known "shape" information for RegisterStatistics('flops') to calculate the statistics. It is suggested to pass in -run_meta_path if shape is only known during runtime. tfprof can fill in the missing shape with the runtime shape information from RunMetadata.

As for RunMetadata, there is a tf.RunMetadata() op in tensorflow, that should be what you need. Usually you pass it to a sess.run() op.

A solution for this is to pass RunMetadata during runtime, when all the shapes will be defined.

It looks like your virtualenv is configured correctly, but the code you are trying to use imports tf.contrib.tfprof, which is not currently (as of TensorFlow 1.2) supported on Windows.

However, since tf.contrib.tfprof is only used to provide profiling information, you should be able to run the code by manually removing the following line from train.py:

• Line 27: from tensorflow.contrib.tfprof import model_analyzer

...and passing the flag --show_graph_stats=false when running the script.

tfprof does static analysis of your graph and calculate the float operations for each graph node.

I assume you are using dynamic_rnn or something similar that has tf.while_loop. In that case, a graph node appear in graph once but is actually run multiple times at run time.

In this case, tfprof has no way to statically figure out how many steps (timestamp in your word) will be run. Hence, it only counts the float operations once.

A work around for now is probably multiply timesteps by yourself.

Both of the two answers are right.

In Ubuntu(python2.7 and GPU) I take a short cut and updated the tensorflow from 10 to 12 by: pip install \ -i https://pypi.tuna.tsinghua.edu.cn/simple/ \ https://mirrors.tuna.tsinghua.edu.cn/tensorflow/linux/gpu/tensorflow_gpu-0.12.1-cp27-none-linux_x86_64.whl In China I use the mirror in Tsinghua University for much faster installation.
It goes well directly.

But this seems don't apply to Mac(at least for python3.5 and cpu only) even if your version of the tensorflow is 12 or above that file has not been generated(even other two files are already listed there(//anaconda/lib/python3.5/site-packages/tensorflow/tools/tfprof/). So I just copied that file(it must be the same as that in my Ubuntu, right?) privided by Yaroslav Bulatov to the folder. And it works.