Training issues.

[LWShowTime]

SelfReformer/data/__init__.py

Line 54 in c95d07e

MASK = color.rgb2gray(MASK) # shape of [h, w]

Here I got an exception.
ValueError: the input array must have size 3 along channel_axis, got (284, 400)

[LWShowTime]

I think there is no need to do with MASK here, so I just comment this line out and everything goes on well.

[BarCodeReader]

Thanks for your info and PR.
For some of the benchmark datasets, their masks are in RGB format so we still need to make sure they are all in grayscale.

I feel a better way to do this is to use PIL and then convert the grayscale to np.array:

mask = Image.open("mask_file_path").convert("L")
mask = np.array(mask)

[LWShowTime]

@BarCodeReader Yes, I found some mask have three dimensions in DUTS-TR, you are right. I guess it's the version of skimage I use. My versions will give out an exception when this line of code met with two dimension. Thanks you for your advice. I finnally handle this by using a if-else statement.

[BarCodeReader]

@LWShowTime I will update the code accordingly. Thanks for your info again.