I have slight differences between obtaining features on the server and on the local machine.

[bbhxwl]

I have slight differences between obtaining features on the server and on the local machine.
0.0037348971236497164
0.0037349374033510685

model, preprocess = clip.load("ViT-B/32")

def extract_features(image_path):
    # 如果是URL链接，下载图片
    if isinstance(image_path, Image.Image):
        img = image_path.convert('RGB')
    elif image_path.startswith('http://') or image_path.startswith('https://'):
        response = requests.get(image_path)
        img = Image.open(BytesIO(response.content)).convert('RGB')
    elif os.path.exists(image_path):  
        img = Image.open(image_path).convert('RGB')

    else:
        raise ValueError("Invalid image path or URL.")
    img_tensor = preprocess(img).unsqueeze(0)
    with torch.no_grad():
        emb = model.encode_image(img_tensor)
        emb = emb / emb.norm(dim=-1, keepdim=True)
        features = emb[0].tolist()

    return features

[bbhxwl]

Is there a different version of ViT-B/32 that is automatically downloaded? Is my server different from the local version?

[AloscariS]

It can be perfectly normal since the code will run on different hardware (remote and local) or even different versions of PyTorch/cuDNN. This will result in slightly different treatments of the floating-point operations.

These small deviations are harmless and don’t indicate a bug.