SummaryWriter.add_scalar() number format and Transforms.TensorboardX

1st Question : How can I control the number format of scalars in TensorboardX board in a graph?

Let’s

# proper imports writer = SummaryWriter() x = 5.123456789 epoch = 3

If I write, writer.add_scalar( 'Raw integer scalar', x, epoch) everything works fine, but the output format in TensorboardX it varies.

If I write, writer.add_scalar( 'Formatted raw integer scalar', '{:0.2%}'.format(x), epoch) I get this kind of error

there is no workspace variable

from line 404, file tensorboardX/writer.py

if self._check_caffe2_blob(scalar_value): scalar_value = workspace.FetchBlob(scalar_value)

and there is a check for a Caffe variable although I am using PyTorch.

2nd Question : How can I use TensorboardX from Transforms module, what are the dependencies and their minimum versions?

Thanks, Vassilis

[mgoldey]

This error is indeed triggered whenever a string is passed in as the second value of add_scalar. That’s because add_scalar expects scalars. That’s a perfectly natural assumption for tensorboardX to make. It would be better to throw a different error in this situation, but both my fundamental problem and the OP’s problem are not tensorboardX’s problem. I would politely suggest throwing a TypeError rather than leaving users thinking that they have a package installation problem when they see the NameError.

example code to trigger a NameError that probably ought to throw a TypeError

from tensorboardX import SummaryWriter
writer = SummaryWriter()
writer.add_scalar("some_variable", "a_string", 0)

[mgoldey]

The specific invocation in transformers could be passing something that’s a string type, so this looks like the logic for whether it’s in caffe2 is falsely triggering. I think I can suggest a one line fix for this, but there might be a more elegant way. Specific logic in transformers triggering this issue for me is just a for loop over a dictionary’s items. https://github.com/huggingface/transformers/blob/02e5f79662d72cccdca81a47e3001a5f6d36e5b1/src/transformers/trainer.py#L556-L557