Tags are interpreted as independent runs

[pshvechikov]

This code

 writer = SummaryWriter(log_dir="runs/first_run") 
 writer.add_scalars('loss/pi', {"policy": pi_loss.data[0], 
                                "value": V_pi_loss.data[0], 
                                "entropy": pi_entropy.data[0]}, i_episode) 

yields in tensorboard page

I guess, this is not the way the Tensorboard is supposed to recognise tags.

Also, if I add, for example, 30 different summaries (including a bunch of histograms and scalars), CPU consumption increases significantly (up to the whole available cpu time). Tensorboard becomes almost irresponsible.
I guess such a behaviour may also be due to very many tags being interpreted by tensorboard as independent runs.

tensorboard-pytorch==0.8 was installed with pip
tensorboard is of 0.1.8 (more older versions of tensorboard are also susceptible to the same issue).

[pshvechikov]

In Tensorflow the result analogous to add_scalars() also needs a separate file_writer and thus also will result in additional lines in the runs section of tensorboard.

[lanpa]

Hi, I think this is a pretty hacky method. To plot many sequences at once, you should dump the data to json format with export_scalars_to_json() and plot with other software.

[pshvechikov]

Ok, thank you.