[FEATURE] DINOv2 Feature Map extraction

[FLamefiREz]

How can I extract feature maps using DNIOv2？

[FLamefiREz]

here， I found an example

from transformers import Dinov2Backbone
import torch

# model_ = torch.load("weights/dinov2_s_cow_front_face_fine_tuned_model.pth")
model = Dinov2Backbone.from_pretrained("facebook/dinov2-base")

pixel_values = torch.randn(1, 3, 224, 224)

outputs = model(pixel_values)

for feature_map in outputs.feature_maps:
    print(feature_map.shape)

[rwightman]

@FLamefiREz this is timm, not transformers... timm options

model = timm.create_model('vit_small_patch14_dinov2.lvd142m', pretrained=True, global_pool='')

print('final features, incl cls token', model(torch.randn(1, 3, 518, 518)).shape)

output = model.get_intermediate_layers(torch.randn(1, 3, 518, 518), 4)
print('last 4 features, no cls token')
for o in output:
    print(o.shape)

output = model.get_intermediate_layers(torch.randn(1, 3, 518, 518), 4, reshape=True)
print('last 4 features, 2D spatial shape')
for o in output:
    print(o.shape)
    
output = model.get_intermediate_layers(torch.randn(1, 3, 518, 518), 99, return_prefix_tokens=True)
print('all features, w/ class token')
for o in output:
    print(o[0].shape, o[1].shape)
    
final features, incl cls token torch.Size([1, 1370, 384])

last 4 features, no cls token
torch.Size([1, 1369, 384])
torch.Size([1, 1369, 384])
torch.Size([1, 1369, 384])
torch.Size([1, 1369, 384])

last 4 features, 2D spatial shape
torch.Size([1, 384, 37, 37])
torch.Size([1, 384, 37, 37])
torch.Size([1, 384, 37, 37])
torch.Size([1, 384, 37, 37])

all features, w/ class token
torch.Size([1, 1369, 384]) torch.Size([1, 1, 384])
torch.Size([1, 1369, 384]) torch.Size([1, 1, 384])
torch.Size([1, 1369, 384]) torch.Size([1, 1, 384])
torch.Size([1, 1369, 384]) torch.Size([1, 1, 384])
torch.Size([1, 1369, 384]) torch.Size([1, 1, 384])
torch.Size([1, 1369, 384]) torch.Size([1, 1, 384])
torch.Size([1, 1369, 384]) torch.Size([1, 1, 384])
torch.Size([1, 1369, 384]) torch.Size([1, 1, 384])
torch.Size([1, 1369, 384]) torch.Size([1, 1, 384])
torch.Size([1, 1369, 384]) torch.Size([1, 1, 384])
torch.Size([1, 1369, 384]) torch.Size([1, 1, 384])
torch.Size([1, 1369, 384]) torch.Size([1, 1, 384])