training with acid

[ChengLiDuoJi]

屏幕截图 2024-10-27 215743
I use acid too. I found that after 60k steps, it often skip bad samples. Should I stop training? But I found your config file wrote that mvsplat should train 300k steps.

[ChengLiDuoJi]

[donydchen]

Hi @ChengLiDuoJi, the skipping operation is as expected, and there is nothing to be worried about. This is because the training mainly intends to use raw data with image resolution 360x640, those raw images with other resolutions will be automatically skipped. Related technical details are in

mvsplat/src/dataset/dataset_re10k.py

Lines 156 to 164 in 31ac688

	context_image_invalid = context_images.shape[1:] != (3, 360, 640)
	target_image_invalid = target_images.shape[1:] != (3, 360, 640)
	if self.cfg.skip_bad_shape and (context_image_invalid or target_image_invalid):
	print(
	f"Skipped bad example {example['key']}. Context shape was "
	f"{context_images.shape} and target shape was "
	f"{target_images.shape}."
	)
	continue

In short, ignore this warning message and continue training to 300K steps.