Image logging to tensorboard

[tmquan]

Hi @williamFalcon ,

Thanks for your nice work.
I am just wondering is it possible to log the image tensor to tensorboard to train such a U-net?

Bests,

[williamFalcon]

yup!

in the lightningModule you can call self.experiment. that’s a summarywriter for tensorboard. for the api of that refer to pytorch summarywriter.

https://pytorch.org/docs/stable/tensorboard

it’s all explained here: https://williamfalcon.github.io/pytorch-lightning/Trainer/Logging/#tensorboard-support

realized self.experiment from docs. adding now!

[williamFalcon]

@tmquan can you update to the latest test-tube version and try again?

[celsofranssa]

Is there an approach to only log a subset of images instead of all over the validation_step or test_step call?

[epignatelli]

@williamFalcon I don't think it's a feasible solution.

It means returning the outputs of the forward pass from training_step.

However, as soon as you override training_epoch_end, it will stack things up in the GPU, until it fill its/their memory.

[ananyahjha93]

@ceceu have you tried

torchvision.utils.make_grid()

and then log images?

[ananyahjha93]

@epignatelli you can create a variable such as

self.imgs = # something

and then plot these. This will work with DDP and the regular single GPU case but will not work with DP the default version does not support state maintenance.

[williamFalcon]

example:

https://pytorch-lightning-bolts.readthedocs.io/en/latest/vision_callbacks.html#confused-logit

[celsofranssa]

Thanks @ananyahjha93 and @williamFalcon, I guess the callback approach is the best practice.

[qmpzzpmq]

Hi, right now is there any method to call tensorboard's add_image add_audio

[tmquan]

@qmpzzpmq
There should be

tensorboard = self.logger.experiment
grid = torchvision.utils.make_grid(...)
tensorboard.add_image(f'{stage}_samples', grid.clamp(0., 1.), self.current_epoch*self.batch_size + batch_idx)