Behavior of global_step=None

[DrJimFan]

This is probably a stupid question: I don't understand the behavior when global_step=None. I'd expect the plot's x-axis to correspond to wallclock time, but in reality it squeezes the curve to a single dot.
Both this repo and the official TB repo say global_step is optional. Under what situation do I actually set it to None?

[lanpa]

I always pass n_iter explicitly. It's more intuitive.

However you can try this for wall clock time:

import time
for n_iter in range(100):
    s1 = torch.rand(1) # value to keep
    writer.add_scalar('data/scalar1', s1[0]) 
    time.sleep(0.1)

Note that it's important to have a sleep.

[DrJimFan]

Yes, I know you can emulate wall clock time that way. However, I do want to understand what does global_step=None mean, and why is it even optional at all.

[lanpa]

Setting global_step=None means no step information is saved in protobuf (while time is still logged). So if users don't want to maintain a global step variable in the code, they can simply ignore this argument.

[DrJimFan]

Gotcha. I guess it's always good to make it explicit then. Thanks!
Happy Black Friday. ;)