from mlxim.model import create_model
from mlxim.io import read_rgb
from mlxim.transform import ImageNetTransform

transform = ImageNetTransform(train=False, img_size=518)
x = transform(read_rgb("cat.png"))
x = mx.expand_dims(x, 0)

model = create_model("vit_large_patch14_518.dinov2")
model.eval()

logits, attn_masks = model(x, attn_masks=True)

You can also use the embeds from layer before head:

python

from mlxim.model import create_model
from mlxim.io import read_rgb
from mlxim.transform import ImageNetTransform

transform = ImageNetTransform(train=False, img_size=512)
x = transform(read_rgb("cat.png"))
x = mx.expand_dims(x, 0)

# first option
model = create_model("vit_large_patch14_518.dinov2", num_classes=0)
model.eval()

embeds = model(x)

# second option
model = create_model("vit_large_patch14_518.dinov2")
model.eval()

embeds, attn_masks = model.get_features(x)

Attention maps

You can visualize the attention maps using the attn_masks returned by the model. Go check the mlx-image notebook.

mlx-vision/vit_large_patch14_518.dinov2-mlxim

vit_large_patch14_518.dinov2

How to use

Attention maps

mlx-vision/
vit_large_patch14_518.dinov2-mlxim