How to get the feature map by using pretrained model.

[JiuqingDong]

Hi, recently I try to use the DINO and DINOv2 pre-train model to get the feature map. I use the pre-trained model to get the feature, but only got a vector. It seems like the output of the last average pooling layer.
Could you give me some suggestions to get the pyramid feature maps?

[Catchip]

I tried to use the following code to extract 2D image features. According to my expectations, I should have obtained a feature map of [1, 14*14, 384], but in reality, I got a feature map of [1, 197, 384]. This makes it difficult for me to figure out how to reshape this feature map into [1, 14, 14, 384].

preprocess = T.Compose([
    T.Resize((224, 224)), 
    T.ToTensor(),  
    T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]), 
])


img = Image.open("sample.jpg")
img_tensor = preprocess(img)
img_tensor = img_tensor.unsqueeze(0)

vits16 = torch.hub.load('facebookresearch/dino:main', 'dino_vits16')
feature = vits16.get_intermediate_layers(img_tensor)  # [1, 197, 384]

[Catchip]

#252 (comment)