New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Unable to visualize Inception v3 graph in TensorBoard with TensorFlow 0.7.1 #1287
Comments
Maybe related to #716 |
Hi @dgolden1, Thanks for reporting and taking the time to include a clean repro. Would you mind trying this setup against the master branch? We've been doing some work in improving the pipeline for large graphs, so it might be that this is already fixed at head. |
I built from source on Ubuntu 14.04 and met the same issue. Any updates on this? |
dsmilkov is out for the rest of the week but I will be investigating this today |
From what I see locally, it seems like this was fixed in Tensorboard in commit 3212eb3. Basically, the graphdef contains huge embedded constant tensors, making the graphdef size too large for the client to handle when it is based from Tensorboard server to the client browser. That commit adds server-side filtering out of large embedded constants, making the client able to handle the served graph data. So, building Tensorboard from scratch on master should allow visualization of inception_v3. Also, the next tagged release should also include a rebuilt Tensorboard with the fix. @dgolden1 and @ffmpbgrnn, did you rebuild the Tensorboard frontend and backend explicitly? Perhaps rebuilding from the TF root doesn't rebuild the Tensorboard components? |
Thanks for the help, @jameswex. I tried building from source again, this time on Ubuntu 14.04 with the latest master (13ea3ca) using Anaconda Python 3.5.1. Results are the same: no graph is displayed. I built via:
I also tried explicitly building tensorboard and running it like:
with the same result. |
It turns out that just a bazel build will not fully rebuild the tensorboard front-end (just the back-end of tensorboard). If you want to manually rebuilt the tensorboard front-end, its currently a multi-step process. I believe in addition to the bazel build of tensorboard, you should also run "gulp vulcanize" in the tensorboard directory to rebuild the front-end HTML that communicates with the tensorboard back-end (see the tensorboard README.md for dependencies for running gulp commands). @danmane, can you confirm if there are additional steps beyond gulp vulcanize and bazel build? Thanks. |
In general, using gulp vulcanize then bazel build will get you the latest and greatest TensorBoard.. |
Some progress; I did gulp vulcanize and then the bazel build with the latest master (e4add49). As before, I'm on Ubuntu 14.04 on Amazon EC2 with Anaconda Python 3.5.1 Now, when attempting to visualize the graph (after dumping it via the same Python snippet in my original issue), I get this TensorBoard error:
The TensorBoard app shows this "Graph visualization failed" error Could this be a Python 2 vs. 3 issue? |
Was able to reproduce the issue using python3. The problem comes down to Fix is on the way. The commit should appear tomorrow. If you don't want to wait, a small fix that makes it work just for python3 is to replace line 66 in |
And also replace line 58 in |
@dsmilkov, making those changes worked! Thanks! After the change has been pushed to master, I'll test again, and close this issue if it works. |
I attempted to test the current master (d868f1e) but now I can't even open a session; I get this error:
Which sounds like another Python 2 vs 3 issue, unrelated to tensorboard. I'll try separately with the commit that included the |
That is a separate python 3 tensorflow (but not tensorboard) issue that I believe is a known issue |
I can't test 9c7be1c either because of yet another non-tensorboard Python 2 vs 3 issue. On 9c7be1c:
Maybe some Python 3 unit tests would be helpful? I'll try this again in a few days when there will hopefully be a version that works on Python 3. |
So we have python 3 tests, but they are not fully integrated requiring us to run them manually. A change yesterday broke TensorFlow on python 3 and fixes are on the way. |
Understandable, @dsmilkov, thanks for the explanation! |
I can confirm the graph can now be visualized properly in f952246. Thanks for working on this! |
i've got the similar issue in python2 |
updating tensorflow works |
Summary
Attempting to visualize the Inception v3 graph with TensorBoard results in an empty graph (after several minutes of loading).
Update: an earlier version of this issue indicated that the progress bar hung forever, but apparently, I just didn't wait long enough.
Environment info
Operating System: OS X 10.11.3, Chrome 48.0.2564.116, Anaconda 1.2.2
If installed from binary pip package, provide:
0.7.1
Steps to reproduce
/tmp/imagenet/classify_image_graph_def.pb
.tensorboard --logdir /tmp/inception_v3_log
Expected result: the graph
Actual result: Empty graph screen (after several minutes of loading with no movement of the progress bar)
A 91 MB file (same size as the graph protobuffer) called
events.out.tfevents.1456423256.[hostname]
is correctly saved to the log directory, so it seems that the graph is in there somewhere.What have you tried?
The text was updated successfully, but these errors were encountered: