How to deal with multi-channel across different data sets?

[brando90]

Hi,

I was curious, how do you deal when one data set has 3 channels (imagenet) but others have 1 (omniglot)?

I've seen others do:

    transform = transforms.Compose([
        lambda x: x.convert("RGB"),
        transforms.Resize(224),
        transforms.ToTensor(),
    ])

in pytorch. Is that what you do?

[lamblin]

We use tf.image.decode_image passing explicitly channels=3, see here.

Then we resize with tf.image.resize_images, and map values to [-1, 1].