BUG: Validation breaks cached image sizes

[feffy380]

The validation split helper function does not adjust the image sizes list, so images get stretched due to being assigned the wrong size.

sd-scripts/library/train_util.py

Line 1939 in 0778dd9

sizes = [None] * len(img_paths)

sd-scripts/library/train_util.py

Lines 1976 to 1998 in 0778dd9

	# We want to create a training and validation split. This should be improved in the future
	# to allow a clearer distinction between training and validation. This can be seen as a
	# short-term solution to limit what is necessary to implement validation datasets
	#
	# We split the dataset for the subset based on if we are doing a validation split
	# The self.is_training_dataset defines the type of dataset, training or validation
	# if self.is_training_dataset is True -> training dataset
	# if self.is_training_dataset is False -> validation dataset
	if self.validation_split > 0.0:
	# For regularization images we do not want to split this dataset.
	if subset.is_reg is True:
	# Skip any validation dataset for regularization images
	if self.is_training_dataset is False:
	img_paths = []
	# Otherwise the img_paths remain as original img_paths and no split
	# required for training images dataset of regularization images
	else:
	img_paths = split_train_val(
	img_paths,
	self.is_training_dataset,
	self.validation_split,
	self.validation_seed
	)

@rockerBOO

[rockerBOO]

I see the sizes are set as None but I'm not following how that is stretching the images. In further reading it picks up on the None and tries to get the image size later in the flow.

[feffy380]

The validation split function shuffles and truncates the list of image paths but not the corresponding list of sizes. So when we access sizes[i] later we get a random image size from the dataset, which gets passed to the image transforms and causes the stretching

[feffy380]

The easiest fix is probably to zip the paths and sizes before doing the validation split and then separate them afterwards