Update README.md
Browse files
README.md
CHANGED
@@ -28,17 +28,18 @@ import torch
|
|
28 |
import torch.nn.functional as F
|
29 |
from urllib.request import urlopen
|
30 |
from PIL import Image
|
31 |
-
from open_clip import create_model_from_pretrained, get_tokenizer
|
32 |
|
33 |
-
model, preprocess = create_model_from_pretrained('hf-hub:ViT-L-16-SigLIP-384')
|
34 |
-
tokenizer = get_tokenizer('hf-hub:ViT-L-16-SigLIP-384')
|
35 |
|
36 |
image = Image.open(urlopen(
|
37 |
'https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/beignets-task-guide.png'
|
38 |
))
|
39 |
image = preprocess(image).unsqueeze(0)
|
40 |
|
41 |
-
|
|
|
42 |
|
43 |
with torch.no_grad(), torch.cuda.amp.autocast():
|
44 |
image_features = model.encode_image(image)
|
@@ -46,9 +47,10 @@ with torch.no_grad(), torch.cuda.amp.autocast():
|
|
46 |
image_features = F.normalize(image_features, dim=-1)
|
47 |
text_features = F.normalize(text_features, dim=-1)
|
48 |
|
49 |
-
text_probs = (
|
50 |
|
51 |
-
|
|
|
52 |
```
|
53 |
|
54 |
### With `timm` (for image embeddings)
|
|
|
28 |
import torch.nn.functional as F
|
29 |
from urllib.request import urlopen
|
30 |
from PIL import Image
|
31 |
+
from open_clip import create_model_from_pretrained, get_tokenizer # works on open-clip-torch>=2.23.0, timm>=0.9.8
|
32 |
|
33 |
+
model, preprocess = create_model_from_pretrained('hf-hub:timm/ViT-L-16-SigLIP-384')
|
34 |
+
tokenizer = get_tokenizer('hf-hub:timm/ViT-L-16-SigLIP-384')
|
35 |
|
36 |
image = Image.open(urlopen(
|
37 |
'https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/beignets-task-guide.png'
|
38 |
))
|
39 |
image = preprocess(image).unsqueeze(0)
|
40 |
|
41 |
+
labels_list = ["a dog", "a cat", "a donut", "a beignet"]
|
42 |
+
text = tokenizer(labels_list, context_length=model.context_length)
|
43 |
|
44 |
with torch.no_grad(), torch.cuda.amp.autocast():
|
45 |
image_features = model.encode_image(image)
|
|
|
47 |
image_features = F.normalize(image_features, dim=-1)
|
48 |
text_features = F.normalize(text_features, dim=-1)
|
49 |
|
50 |
+
text_probs = torch.sigmoid(image_features @ text_features.T * model.logit_scale.exp() + model.logit_bias)
|
51 |
|
52 |
+
zipped_list = list(zip(labels_list, [round(p.item(), 3) for p in text_probs[0]]))
|
53 |
+
print("Label probabilities: ", zipped_list)
|
54 |
```
|
55 |
|
56 |
### With `timm` (for image embeddings)
|