Fixed documentation for Zero-Shot Prediction

[apkuhar]

norm must be done with dim=1 not dim=-1

[shinian-023]

dim=1 and dim=-1 are the same

[apkuhar]

@GentelmanTenyeaers
Not strictly, it depends on the number of dimensions of the input tensor.
The problem is if someone blindly copies the examples and modified it's for his own use, no realizing the -1 is not the intended dimension.

Plus the code internally actually uses dim=1

CLIP/clip/model.py

Line 363 in a9b1bf5

image_features = image_features / image_features.norm(dim=1, keepdim=True)