Tensorboard Embedding wrong images/labels about tensorboardx HOT 11 CLOSED

Yes, and please have a test on the function before PR.

from tensorboardx.

Ok, I think I understand what is happening. according to the tensorboard embedding doc

The sprite should have the same number of rows and columns

There is no such control in the code right now, and this explains why TB is link wrong images and labels (although it is a poor implementation from them to expect a squared image IMHO) . In fact if you:

• Get the images number n_i
• Find the minimum squared sprite dimensions which can hold the images using (int(ceil(n_i**0.5))))**2
• Concat black images at the end of the valid images
• Call make_grid using the square root size (which now does exist as integer)

Everything should works. You can see the different results for a batch of 10 images below

The first one produces wrong association, while the second one works!

from tensorboardx.

great!
But I think you should fix:
https://github.com/lanpa/tensorboard-pytorch/blob/master/tensorboard/embedding.py#L15
and make nrow correct.
Would you like to send a PR?

from tensorboardx.

I' m working on that function indeed. Just one question:

• What is the use of xx?

from tensorboardx.

https://github.com/lanpa/tensorboard-pytorch/blob/master/tensorboard/embedding.py#L17

Just a dirty fix for torchvision. lol

from tensorboardx.

I make a PR as soon as I learn how it works :D

from tensorboardx.

changing nrow = int(math.floor(math.sqrt(label_img.size(0))))
to nrow = int(math.ceil(math.sqrt(label_img.size(0)))) would be OK. Am I right?

from tensorboardx.

no, beacuse you also have to concat black images to the original ones.
Should i fork the repo and then make a pr?

from tensorboardx.

I'm getting an unexpected behaviour...the label are linked correctly, but the images of the first column (except for the first element, i.e m[0,0] ) are shown as black. Any idea?

7 and 4 are black.

I've checked the dimension of the output of make_grid to find that somehow a pixel get append in both size, maybe the problem is there (which explain some of your fixes btw)

from tensorboardx.

Fixed and PR made :)

from tensorboardx.

Should work now. Thanks!

from tensorboardx.