Output with shape [1, 28, 28] doesn't match the broadcast shape [3, 28, 28]

vision
1

output with shape [1, 28, 28] doesn’t match the broadcast shape [3, 28, 28], this is the error how can I resolve this??

Screenshot_1
Screenshot_11365×589 97.1 KB

2

Based on the error message it looks like (some) images might be grayscale images and thus only contain one channel.
If that’s the case for all images, just use a single value for the mean and std arguments:

transforms.Normalize((0.5,), (0.5,))

Alternatively, if just a few images are gray while others are RGB images, you could convert these images after loading using:

img = Image.open(path).convert('RGB')
1 Like
3

I have the same problem, and I am converting the image to RGB as you suggest. Although really I want a grayscale image.

t_ = None
    if Hyper.is_grayscale:
        t_ = transforms.Compose([
            transforms.Grayscale(num_output_channels=1),
            transforms.ToTensor(),
            transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 5])
            ])
    else:
        t_ = transforms.Compose([
            transforms.ToTensor(),
             transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
            ])
    test_img1 = Image.open("test_examples/fast_train.jpg").convert("RGB")
    test_img1 =  t_(test_img1).unsqueeze(0)
4

If you want to keep the image tensors in the grayscale format, you could use the Normalize transformation from my previous post with a single mean and std value and it should work.
Currently you are passing 3 values for these stats, which will most likely yield a shape mismatch error.

5

Yes, I have since noticed that for Grayscale images I need only 1 value for the mean and 1 value for the std in the normalisation transformation, so I have fixed that now