Tensorboardx: how to merge multiple scalar into one figure

[blueardour]

Hi, thanks for the excellent library.

I have a question. When add a write.add_scalar, a new figure is generated.
I wonder how to merge multiple scalar into one figure? It is similar with the write.add_scalars function, however I have to call the write.add_scalar in different positions of the code.

[lanpa]

Sorry I can understand your question. Doesn't demo.py work for your case?

[blueardour]

Yes, the demo.py work perfect for me. Sorry for my unclear description.

I have a requirement like the following:
`

  writer = SummaryWriter()

  for iter in emuerate(xx):
     write.add_scalar('loss', loss, iter)
      if iter % 1000 ==0:
         write.add_scalar('top1', top1, iter)
         write.add_scalar('top5', top5, iter)

  write.close()

`
After I show the log in tensorboard, the loss, top1, top5 are in different figures.
I wonder if I could merge these plot into one figure. Could it possible provide some 'write.merge()' function. The write.add_scalars interface seems to enable multiple variable grouped into one figure, however it requires all variable in the same sentence.

[lanpa]

What arguments would be passed to write.merge()? Wouldn't it similar to add_scalars?

[blueardour]

Yes, I expect pass the variable name with the same x-axis(for example, all the 'loss', 'top1', 'top5' scalars employ the 'iter' as their x-axis) into the write.merge(). So that these scalars could be drawn in the same figure.
I'm sorry that I don't know whether it is hard to implement this function.

The following is some figure I draw with matplotlib.

By the kind of figure, the val loss in some trival is a little higher than the training loss, indicating a over fitting.

[lanpa]

Oh, I think I got your point. Does the following code work for you?

for n_iter in range(10000):
    writer.add_scalars('data/scalar_group', {'loss': n_iter*np.arctan(n_iter)}, n_iter)
    if n_iter%1000==0:
        writer.add_scalars('data/scalar_group', {'top1': n_iter*np.sin(n_iter)}, n_iter)
        writer.add_scalars('data/scalar_group', {'top5': n_iter*np.cos(n_iter)}, n_iter)

Looks like you use matplotlib, you may also interested in discussing #152 with us.

[blueardour]

Great. Thanks for the illustration. It solved my problem. I would also look at the #152 issue.