nielsr HF staff commited on
Commit
74b5ea6
1 Parent(s): f830025

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +27 -10
README.md CHANGED
@@ -31,32 +31,49 @@ The model uses a CLIP backbone with a ViT-L/14 Transformer architecture as an im
31
 
32
  ### Use with Transformers
33
 
34
- ```python3
35
  import requests
36
  from PIL import Image
 
37
  import torch
 
 
38
 
39
- from transformers import Owlv2Processor, Owlv2ForObjectDetection
40
-
41
- processor = Owlv2Processor.from_pretrained("google/owlv2-large-patch14")
42
  model = Owlv2ForObjectDetection.from_pretrained("google/owlv2-large-patch14")
43
 
44
  url = "http://images.cocodataset.org/val2017/000000039769.jpg"
45
  image = Image.open(requests.get(url, stream=True).raw)
46
  texts = [["a photo of a cat", "a photo of a dog"]]
47
  inputs = processor(text=texts, images=image, return_tensors="pt")
48
- outputs = model(**inputs)
49
 
50
- # Target image sizes (height, width) to rescale box predictions [batch_size, 2]
51
- target_sizes = torch.Tensor([image.size[::-1]])
52
- # Convert outputs (bounding boxes and class logits) to COCO API
53
- results = processor.post_process_object_detection(outputs=outputs, threshold=0.1, target_sizes=target_sizes)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
54
 
55
  i = 0 # Retrieve predictions for the first image for the corresponding text queries
56
  text = texts[i]
57
  boxes, scores, labels = results[i]["boxes"], results[i]["scores"], results[i]["labels"]
58
 
59
- # Print detected objects and rescaled box coordinates
60
  for box, score, label in zip(boxes, scores, labels):
61
  box = [round(i, 2) for i in box.tolist()]
62
  print(f"Detected {text[label]} with confidence {round(score.item(), 3)} at location {box}")
 
31
 
32
  ### Use with Transformers
33
 
34
+ ```python
35
  import requests
36
  from PIL import Image
37
+ import numpy as np
38
  import torch
39
+ from transformers import AutoProcessor, Owlv2ForObjectDetection
40
+ from transformers.utils.constants import OPENAI_CLIP_MEAN, OPENAI_CLIP_STD
41
 
42
+ processor = AutoProcessor.from_pretrained("google/owlv2-large-patch14")
 
 
43
  model = Owlv2ForObjectDetection.from_pretrained("google/owlv2-large-patch14")
44
 
45
  url = "http://images.cocodataset.org/val2017/000000039769.jpg"
46
  image = Image.open(requests.get(url, stream=True).raw)
47
  texts = [["a photo of a cat", "a photo of a dog"]]
48
  inputs = processor(text=texts, images=image, return_tensors="pt")
 
49
 
50
+ # forward pass
51
+ with torch.no_grad():
52
+ outputs = model(**inputs)
53
+
54
+ # Note: boxes need to be visualized on the padded, unnormalized image
55
+ # hence we'll set the target image sizes (height, width) based on that
56
+
57
+ def get_preprocessed_image(pixel_values):
58
+ pixel_values = pixel_values.squeeze().numpy()
59
+ unnormalized_image = (pixel_values * np.array(OPENAI_CLIP_STD)[:, None, None]) + np.array(OPENAI_CLIP_MEAN)[:, None, None]
60
+ unnormalized_image = (unnormalized_image * 255).astype(np.uint8)
61
+ unnormalized_image = np.moveaxis(unnormalized_image, 0, -1)
62
+ unnormalized_image = Image.fromarray(unnormalized_image)
63
+ return unnormalized_image
64
+
65
+ unnormalized_image = get_preprocessed_image(inputs.pixel_values)
66
+
67
+ target_sizes = torch.Tensor([unnormalized_image.size[::-1]])
68
+ # Convert outputs (bounding boxes and class logits) to final bounding boxes and scores
69
+ results = processor.post_process_object_detection(
70
+ outputs=outputs, threshold=0.2, target_sizes=target_sizes
71
+ )
72
 
73
  i = 0 # Retrieve predictions for the first image for the corresponding text queries
74
  text = texts[i]
75
  boxes, scores, labels = results[i]["boxes"], results[i]["scores"], results[i]["labels"]
76
 
 
77
  for box, score, label in zip(boxes, scores, labels):
78
  box = [round(i, 2) for i in box.tolist()]
79
  print(f"Detected {text[label]} with confidence {round(score.item(), 3)} at location {box}")