A simple script for parsing marian training logs and displaying them as scalars in tensorboard. For visual monitoring and comparison.

• virtualenv -p python3 p3
• source p3/bin/activate
• pip install tensorboardX tensorboard

Assume you have a directory containing one or more Marian NMT trainings to compare. They have structure like this:

.
├── src-tgt-demo-1
│   └── model
│       └── train.log    # tb_log_parser.py requires train.log created by marian on this path. Other files such as checkpoints and running scripts can be present nearby.
├── src-tgt-demo-2
│   └── model
│       └── train.log    # and similarly here
├── ...                  # there can be more runs
│
├── tb_log_parser.py      
└── tb-monitored-jobs    # tb_log_parser.py requires this in the same directory 

If you have active trainings in src-tgt-demo-1 and src-tgt-demo-2, register them for monitoring by including the names into tb-monitored-jobs file, so it looks like this:

$ cat tb-monitored-jobs 
src-tgt-demo-1
src-tgt-demo-2

Then you can run the log parser:

$ ./tb_log_parser.py 
update loop src-tgt-demo-1
 current modification time: 1572357911.0 last: 0
last line id: 26122
update loop src-tgt-demo-2
 current modification time: 1572358555.0 last: 0
update loop src-tgt-demo-1
update loop src-tgt-demo-2
update loop src-tgt-demo-1
update loop src-tgt-demo-2

It will go through tb-monitored-jobs, and for each monitored job, it parses data from the lines of train.log, which newly appeared after last update loop, saves the values in tensorboard-readable format into src-tgt-demo-1/tb directory, together with pickled variables for restoring current parsing state. The logs are revisited every 5 seconds. You can update tb-monitored-jobs meanwhile to add or remove monitored jobs.

Let it running in the terminal window, open new terminal, source p3/bin/activate, tensorboard --logdir=. --port=6006. Wait until it writes it's ready, then open a webbrowser, go to localhost:6006 and analyze.

Install, follow the steps in Run section above and observe some training and validation scalars in tensorboard in your browser. Two sample train.log files are included in this repo.

• Licence: MIT, see Licence.md

• There are many points and features to explain. Feel free to ask.

• Concat: Dominik Macháček: machacek [at] ufal.mff.cuni.cz