Tensorboard SummaryWriter not writing images correctly

[import-antigravity (Robbie Dozier)]

I know TB is not made by pytorch but I think the issue probably has to do with the Pytorch SummaryWriter side of things.

I’ve been trying to train a CIFAR-10 DCGAN and pulling my hair out for a week because the images always come out like this:

Observe the extreme clipping/saturation.

Now, today I decided to write some training data to tensorboard as well and I got this:

Here is the training data observed with plt.imshow in a Jupyter notebook, which confirms that my data is not loaded incorrectly.

I’m completely stumped. Grayscale images (MNIST) work fine on TB. Does anyone have any insight? This is my code to write the images to tensorboard (I’m using pytorch lightning but that shouldn’t matter):

from torchvision.utils import make_grid
img = pl_module.sample_G(self._n * self._n)

train_batch = next(iter(pl_module.train_dataloader()))[:(self._n * self._n)]

# Concatenate images
grid = make_grid(img, self._n)
t_grid = make_grid(train_batch, self._n)

if pl_module.logger is not None:
    pl_module.logger.experiment.add_image('train', t_grid, pl_module.global_step, dataformats='CHW')
    pl_module.logger.experiment.add_image('generated', grid, pl_module.global_step, dataformats='CHW')

[import-antigravity (Robbie Dozier)]

Update: fixed it!

The problem is that the images were between -1 and 1, and torch.utils.tensorboard.summary.image() only accepts floating point images in [0, 1] or integer images in [0, 255]. I’m going to go ahead and open an issue on github because the code should be changed or the docs should be changed to make this more obvious.