Inverted colors in Tensorboard SummaryWriter add_image() function

user_8449237 0
Viewed 774

There is an image stored in image_tensor (image_tensor of size (3,256,512), storing values in the interval 0,255) which I would like to display in Tensorboard (TensorboardX for PyTorch, more specifically) via the add_image() function for SummaryWriter. When I add the image to the Tensorboard via writer.add_image("color_image",image_tensor,self.step), the colors are inverted.

When I write the image to a file via scipy.misc.imsave("/write/to/path/image.png",np.transpose(image_tensor.data.cpu().numpy(),(1,2,0))), the image is perfectly fine.

Only thing I change for the second line is changing CxHxW to HxWxC, but I don't think that this is the root of this color inversion issue. What might be the problem?

torch pytorch tensorboard python-3.x python
1 Answers

I had a similar problem, where I was also unable to track the problem. The solution that worked for me but is unfortunately a little bit cumbersome is:

Take your image and plug it into a matplotlib figure then use add_figure.

For example:

fig, ax =  plt.subplots(2,3)
# add your subplots with some images eg.
ax[0,0].imshow(image_1)
# etc.
writer_semisuper.add_figure("testfig", fig, 0)

This shows exactly the same plot that you created, but with lower resolution. So if your plot works in jupyter or saved on disk it should also work in tensorboard.

user_114811660
Related