-
Notifications
You must be signed in to change notification settings - Fork 46
/
detect.py
165 lines (129 loc) · 4.69 KB
/
detect.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
import torch
from datasets import load_dataset
from kornia.metrics import mean_average_precision
from torchvision.ops import nms
from transformers import AutoModelForZeroShotObjectDetection, AutoProcessor
from kangas import DataGrid, Image
PRETRAINED_MODEL_NAME = "google/owlvit-base-patch16"
DATASET_NAME = "detection-datasets/fashionpedia_4_categories"
DATASET_SPLIT = "val"
N_SAMPLES = 500
IOU_THRESHOLD = 0.5
SCORE_THRESHOLD = 0.2
if torch.cuda.is_available():
device = torch.device("cuda")
else:
device = torch.device("cpu")
processor = AutoProcessor.from_pretrained(PRETRAINED_MODEL_NAME)
model = AutoModelForZeroShotObjectDetection.from_pretrained(PRETRAINED_MODEL_NAME).to(
device
)
model.eval()
dataset = load_dataset(f"{DATASET_NAME}", split=f"{DATASET_SPLIT}", streaming=True)
dataset = dataset.shuffle(seed=42)
dataset = dataset.take(N_SAMPLES)
class_names = dataset.features["objects"].feature["category"].names
columns = ["id", "image", "mAP"]
columns.extend([cn for cn in class_names])
columns.extend([f"score_{cn}" for cn in class_names])
dg = DataGrid(
name="fashionpedia",
columns=columns,
)
def get_class_pred_score(cn, labels, pred_scores):
return pred_scores[labels.index(cn)].item()
def apply_nms(pred_bboxes, pred_scores, pred_labels, threshold):
indices = nms(pred_bboxes, pred_scores, threshold)
return pred_bboxes[indices], pred_scores[indices], pred_labels[indices]
def apply_score_threshold(pred_bboxes, pred_scores, pred_labels, threshold):
indices = torch.where(pred_scores >= threshold)
if len(indices) != 0:
return pred_bboxes[indices], pred_scores[indices], pred_labels[indices]
else:
return None
def annotate_datagrid_image(dg_image, bboxes, labels, scores=None):
if scores is None:
for category, bbox in zip(labels, bboxes):
dg_image.add_bounding_boxes(
f"gt_{class_names[category]}",
[(bbox[0], bbox[1]), (bbox[2], bbox[3])],
score=1.0,
)
else:
for pred_bbox, pred_score, pred_label in zip(bboxes, scores, labels):
pred_bbox, pred_score, pred_label = map(
lambda x: x.tolist(), [pred_bbox, pred_score, pred_label]
)
dg_image.add_bounding_boxes(
f"pred_{class_names[pred_label]}",
[
(pred_bbox[0], pred_bbox[1]),
(pred_bbox[2], pred_bbox[3]),
],
score=pred_score,
)
for idx, example in enumerate(dataset):
image_id = example["image_id"]
image = example["image"]
dg_image = Image(image)
objects = example["objects"]
labels = objects["category"]
bboxes = objects["bbox"]
annotate_datagrid_image(dg_image, bboxes, labels)
inputs = processor(
text=[f"a photograph of {cn}" for cn in class_names],
images=image,
return_tensors="pt",
).to(device)
with torch.no_grad():
outputs = model(**inputs)
outputs.logits = outputs.logits.cpu()
outputs.pred_boxes = outputs.pred_boxes.cpu()
target_sizes = torch.Tensor([image.size[::-1]])
results = processor.post_process(outputs=outputs, target_sizes=target_sizes)
result = results[0]
pred_bboxes, pred_scores, pred_labels = (
result["boxes"],
result["scores"],
result["labels"],
)
filtered_preds = apply_score_threshold(
pred_bboxes, pred_scores, pred_labels, SCORE_THRESHOLD
)
# Skip if predictions scores are too low
if filtered_preds is None:
continue
else:
pred_bboxes, pred_scores, pred_labels = filtered_preds
pred_bboxes, pred_scores, pred_labels = apply_nms(
pred_bboxes, pred_scores, pred_labels, IOU_THRESHOLD
)
map_score, _ = mean_average_precision(
[pred_bboxes],
[pred_labels],
[pred_scores],
[torch.tensor(bboxes)],
[torch.tensor(labels)],
len(class_names),
threshold=IOU_THRESHOLD,
)
annotate_datagrid_image(dg_image, pred_bboxes.long(), pred_labels, pred_scores)
row = {
"id": image_id,
"image": dg_image,
"mAP": map_score.item(),
}
# update columns based on whether a class is present in an image
label_names = list(map(lambda x: class_names[x], labels))
row.update({cn: 1 if cn in label_names else 0 for cn in class_names})
pred_label_names = list(map(lambda x: class_names[x], pred_labels))
row.update(
{
f"score_{cn}": get_class_pred_score(cn, pred_label_names, pred_scores)
if cn in pred_label_names
else 0.0
for cn in class_names
}
)
dg.append(row)
dg.save()