search

Kaixhin / FGLab

Future Gadget Laboratory
https://kaixhin.github.io/FGLab/
MIT License
223 stars 34 forks source link

Tensorboard Integration #21

Open xanderdunn opened 8 years ago

xanderdunn commented 8 years ago

FGLab has some built-in graph UI: experiment

TensorFlow's TensorBoard has a full-featured, interactive web GUI for visualizing training output: 687474703a2f2f72757373656c6c73737465776172742e636f6d2f732f74656e736f72626f782f74656e736f72626f6172645f6c6f73732e706e67

Would it be possible to store the TensorBoard summary files in FBLab's database and then link from the FBLab web UI to open the TensorBoard UI? Ideally, FBLab would even be able to read stored values (train loss, val loss, etc.) directly from the TensorBoard summary files.

This would be an awesome integration for TensorFlow users.

Kaixhin commented 8 years ago

If there are summary/log files being stored in a directory then FGLab will certainly store them in MongoDB's GridFS - but as individual files rather than the entire directory, so you would have to download all files one by one/utilise the API to get them via a script.

It's possible to add custom fields with text, and I could implement extra processing to recognise HTTP links and make them actual links, but this would require the aforementioned script to have been run, and TensorBoard to have been started.

For FGLab to try and produce its own graphs, you would have to have something parse the files written by the SummaryWriter in a different directory and place the JSON graphs into the experiment directory to be uploaded to FGLab.

Given TensorBoard's wealth of features and tight integration with TensorFlow, how exactly are you hoping to use FGLab with TensorFlow?

xanderdunn commented 8 years ago

I'm hoping to use FBLab to store and present experiment results.

Tensorboard is a nice UI for viewing the results of a specific experiment, but it offers no solution to storing or viewing the results of a myriad of parameters. Tensorboard doesn't do this: experiments

This goal could be achieved without any kind of FGLab<->Tensorboard integration. I think it's purely a matter of convenience, as opposed to needing to find the TensorBoard summary files manually on disk and get to it externally from the FGLab view.

Kaixhin commented 8 years ago

Option 1, which is to present TensorBoard graphs natively in FGLab, would require parsing the files somewhere - probably on the server. It should be possible, but I don't have the resources to work on that.

Option 2 is to devise a script that makes it easy to download the files for TensorBoard to launch, and document it in the examples. This is something I can tackle once I have more time. Let me know if you want to try either option yourself and I can answer questions about the FGLab code.

Kaixhin commented 8 years ago

Just a quite note to say that this is very low priority, i.e. I'm very unlikely to get around to this myself, but if you want to attempt it I'll do my best to answer questions.