Torch.utils.tensorboard adding images with step

1

Hi, I’m using this util of visualizing:

from torch.utils.tensorboard import SummaryWriter

for image visualization and I’ve noticed that as more images I right each step (which is in my case epoch), the step count and points are changed, i.e:

When about 15 epoch were finished, although I write each epoch fixed amount of same images,
only steps 1,2,3,4,8, … (some more, doesn’t matter).
Now after more then 30 epochs I’ve noticed that steps 4,8 disappeared and 3 and 11 are showed.

Before:
image

Now:
image

image

I’ve only documented the the disappearing of 4, the two last images are consequent at the step bar.

How can I see all the steps ?
Does Tensorboard swallows them and that is a bug and bad behavior or there is something in the settings of the external api (in the browser) or some argument in the code needed to be passed for not having this behavior ?
Thank you in advance.

It happened only for images, for loss and accuracy and other point recordings I still see all the steps.

2

Oh, I think that parameter is responsibale for it, correct me if I am wrong:

max_queue

image

3

This is because by default Tensorboard samples some logged results, not all.

If you want to increment the number of sampled images just set the “samples_per_plugin” argument when loading the Tensorboard:

tensorboard --logdir <your_directory> --samples_per_plugin “images=< number_of_images>”

Example

tensorboard --logdir runs --samples_per_plugin “images=10000”

Note: Maybe the documentation say to set “0” to load all images. However, it will not according to this reported bug –samples_per_plugin “images=0” is documented to show all images, but shows no images at al