@guillaumekln
I am running ONMT-tf with tensorflow-gpu 1.8.0. I used
tensorboard --logdir="/OpenNMT09/toy-ende/run"
and I got
W0322 22:15:17.081071 Reloader tf_logging.py:121] Found more than one graph event per run, or there was a metagraph containing a graph_def, as well as one or more graph events. Overwriting the graph with the newest event.
TensorBoard 1.8.0 at http://ip-XXX-XX-X-XX:XXXX (Press CTRL+C to quit)
Is it the error of my installation (the most recent ONMT-tf build + tensorflow-gpu 1.8.0, or is it something else? ONMT-tf created the run folder automatically.
It’s a warning. Just open the URL in your browser.
Do you mean the /OpenNMT09/toy-ende/run must be accessible from the Web? Note I am running the training on an AWS EC2 Linux instance, I can’t open any browser there. BTW, I tried to move the folder to the Apache location and now can access the folder from my Windows machine by using something like https://xx.xx.xx.xxx/run, but I just see a folder structure. Is it possible to use Tensorboard when running a training on AWS EC2?
Right in the logs:
TensorBoard 1.8.0 at http://ip-XXX-XX-X-XX:XXXX (Press CTRL+C to quit)
That’s the URL you should open in your browser. You may need to open the port for the EC2 instance to make it accessible.
Maybe search for Tensorboard tutorials online.
Just FYI, probably the best “tutorial” is at Steps to train a Transformer model on the WMT English-German dataset page, Monitoring model training using TensorBoard section.
However, in my case, I had to also open the 6006 port as described on the AWS Train an MNIST Model and Visualize the Training with TensorBoard page.