Question for the dataset.py

[halbielee]

Hello, thanks for sharing your excellent work!

I have a question about your code, especially coco_train_dataset.py.

PiCIE/data/coco_train_dataset.py

Lines 97 to 106 in 62d2748

	# Invariance transform.
	image1 = self.transform_inv(index, image, 0)
	image1 = self.transform_tensor(image)
	if self.mode == 'baseline_train':
	return (image1, )
	image2 = self.transform_inv(index, image, 1)
	image2 = TF.resize(image2, self.res1, Image.BILINEAR)
	image2 = self.transform_tensor(image2)

When applying the transform_tensor function to the image, it applies to the variable image not image1.

Is there any issue for this?

Thanks!

[janghyuncho]

Hi halbielee,

Thank you for your interest in our work! I am happy to help you.

Indeed, you are absolutely correct that it appears to be my mistake during refactoring. The code for the result from the paper did not have this bug (only during refactoring). However, we did try this version before and it makes very small difference; in fact, many contrastive learning papers choose this way (contrasting between x <-> T(x) instead of T1(x) <-> T2(x)).

I fixed the code just now.

Thank you!

Jang Hyun Cho

[janghyuncho]

Hi halbielee,

Here is the training log of PiCIE with the updated code.

ACC - All: 51.5297
mIOU - All: 14.7917
ACC - Thing: 76.6696
mIOU - Thing: 18.1085
ACC - Stuff: 64.4714
mIOU - Stuff: 26.3870

Best,

Jang Hyun Cho

[halbielee]

Thank you for the quick and kind reply :)