Hi! Thanks for publishing this work, it's a great reference.

I'm trying to integrate a couple of different systems, and I need the model encodings to match. So far, I haven't been able to make that work:

Given this python;

device = "cuda" if torch.cuda.is_available() else "cpu"
model, preprocess = clip.load("ViT-B/32", device=device)

image = preprocess(Image.open("image_1.png")).unsqueeze(0).float().to(device)
text = clip.tokenize(["a face", "a dog", "a cat"]).to(device)

with torch.no_grad():
    image_features = model.encode_image(image)
    print(image_features.tolist()[0])

I'm trying to get the same array of floats out using Clip.mm's - (NSArray<NSNumber*>*)test_uiimagetomat:(UIImage*)image function. Try as I might, they always differ - and I'm not sure what the difference is. I can see that the cvt methods do the same as the image preprocess, then the normalise with the values from clip.

Here's some of the initial values from the python code above:

[0.3502497971057892, 0.0028706961311399937, -0.46749746799468994, -0.14868411421775818, -0.03139263391494751, -0.4536064863204956

And from the Swift:

[0.3193549513816833496, 0.0140316337347030640, -0.4410626888275146484, -0.0908056870102882385, -0.0415024310350418091, -0.4141347408294677734

I used the preview of the quicklook on debugging the iOS code to save the image from the UIImage to ensure the same image is being used. In both cases, I'm using the original vit-b-32 CLIP image encoding. Strangely, the numbers above are kind of similar - but not sure if that's coincidental.

Any advice?