Closed shadowbourne closed 3 years ago
But please help as I think this tool is (or would be if i could get it to work) so useful!
I haven't tested tb-reducer
on Windows so maybe a path separator issue? Could you zip and upload your event files here?
Thanks for your files. So there were multiple problems here.
First, tb-reducer
rather short-sightedly did not yet support logs increasing in anything other than unit steps. It had to be
[
{"step": 0, "value": 0.8129},
{"step": 1, "value": 7234},
{"step": 2, "value": 0.5224},
...
]
This is now fixed in tensorboard-reducer==0.2.0
.
Second, it intentionally did not support runs with different step counts. In fact, there was a disclaimer in the readme about that (now removed). I should have looked at your screenshot more carefully. It clearly shows that seed0
reaches to about 50k steps, while seed1
only has ~18k steps. You can now explicitly tell tb-reducer
that you are aware of this and would like to reduce those two runs anyway (using however many steps are available in the shortest run) by passing the --lax-steps
flag.
Third, TensorBoard Reducer also intentionally did not support loading TensorBoard runs with non-identical sets of scalar tags. This constraint can also now be deactivated by passing --lax-tags
. However, this wasn't an issue in your case.
In summary, could you try pip install -U tensorboard-reducer
to make sure you're running 0.2.0 followed by tb-reducer -i 'seed?/*' -o out --lax-steps
and see if it works now?
Lovely work! new issues have emerged, but I shall close this issue and open another!
PS C:\Users\shadow\Desktop\tb> tb-reducer -i 'seed?/*' -o out Loaded 2 TensorBoard runs with 99 scalars and 383 steps each