Error using Visdom as exp. tracking system for image-segmentation

[idkwid2022]

Describe the bug

AttributeError: 'VisdomLogger' object has no attribute 'writer'

Expected result

No error.

Reproduction

Please see line 137, src/templates/template-vision-segmentation/vis.py.
logger.writer.add_image is for tensorboard, no?

logger.writer.add_image

Steps to reproduce

Choose template for vision-segmentation
Choose Visdom for exp. tracking system.

[vfdev-5]

@idkwid2022 thanks for raising this issue. You are right this line :

code-generator/src/templates/template-vision-segmentation/vis.py

Lines 137 to 142 in f137219

	logger.writer.add_image(
	tag=tag,
	img_tensor=grid_pred_gt,
	global_step=global_step,
	dataformats="HWC",
	)

is working only with TensorBoard. Would you like to suggest a fix ?

Quick fix could be just to make sure that this code :

code-generator/src/templates/template-vision-segmentation/main.py

Lines 114 to 141 in f137219

	# Log validation predictions as images
	# We define a custom event filter to log less frequently the images (to reduce storage size)
	# - we plot images with masks of the middle validation batch
	# - once every 3 validations and
	# - at the end of the training
	def custom_event_filter(_, val_iteration):
	c1 = val_iteration == len(dataloader_eval) // 2
	c2 = trainer.state.epoch % 3 == 0
	c2 |= trainer.state.epoch == config.max_epochs
	return c1 and c2
	# Image denormalization function to plot predictions with images
	mean = (0.485, 0.456, 0.406)
	std = (0.229, 0.224, 0.225)
	img_denormalize = partial(denormalize, mean=mean, std=std)
	exp_logger.attach(
	evaluator,
	log_handler=predictions_gt_images_handler(
	img_denormalize_fn=img_denormalize,
	n_images=15,
	another_engine=trainer,
	prefix_tag="validation",
	),
	event_name=Events.ITERATION_COMPLETED(
	event_filter=custom_event_filter
	),
	)

is executed only if exp_logger is TensorBoard.

Another fix could be to find some common method we could call with all supported exp tracking systems...

[idkwid2022]

@vfdev-5 I don't think I have qualification for any suggestion, I'm a total newbie in this.

If I may, I think it should be like in image classification template since the logger options is the same for every template (template-vision-classification/utils.py, line 213, function setup_exp_logging) other option is to disable the loggers other than tensorboard, but that's a bummer isn't it?

Anyhow, I kinda did make it run on my notebook, but I forgot which lines I changed. The main change is on vis.py, line 137:

logger.vis.image(grid_pred_gt.transpose(2,0,1), win=tag, opts=dict(title=tag))

But that's just to make it run only on visdom. So maybe like this:

#::: if (it.logger === 'tensorboard') { :::#
        logger.writer.add_image(
            tag=tag,
            img_tensor=grid_pred_gt,
            global_step=global_step,
            dataformats="HWC",
        )
#::: } else if (it.logger === 'visdom') { :::#
        logger.vis.image(grid_pred_gt.transpose(2,0,1), win=tag, opts=dict(title=tag))
#::: } :::#

Like on image classification template? I don't know about the other logging system. I'm sorry I can't be much of a help.

[vfdev-5]

@idkwid2022 no problems about your qualifications.
I agree with your suggestion for tb and visdom only if/else branches. For others we can update later if there's a requirement.
I have to check classification template but I do not remember if we were logging any images there...

[idkwid2022]

I meant about how the classification template deals with different loggers. I have not use the template, so I don't know if it logs images.