Upload app.py
Browse files
app.py
ADDED
@@ -0,0 +1,49 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from PIL import Image
|
3 |
+
from transformers import DetrImageProcessor, DetrForObjectDetection
|
4 |
+
import gradio as gr
|
5 |
+
import matplotlib.pyplot as plt
|
6 |
+
import matplotlib.patches as patches
|
7 |
+
import io
|
8 |
+
|
9 |
+
# Load model and processor once
|
10 |
+
processor = DetrImageProcessor.from_pretrained("facebook/detr-resnet-50")
|
11 |
+
model = DetrForObjectDetection.from_pretrained("facebook/detr-resnet-50")
|
12 |
+
|
13 |
+
def detect_objects(image):
|
14 |
+
# Run DETR model
|
15 |
+
inputs = processor(images=image, return_tensors="pt")
|
16 |
+
outputs = model(**inputs)
|
17 |
+
target_sizes = torch.tensor([image.size[::-1]])
|
18 |
+
results = processor.post_process_object_detection(outputs, target_sizes=target_sizes, threshold=0.9)[0]
|
19 |
+
|
20 |
+
# Draw boxes on image
|
21 |
+
fig, ax = plt.subplots(1)
|
22 |
+
ax.imshow(image)
|
23 |
+
|
24 |
+
for score, label, box in zip(results["scores"], results["labels"], results["boxes"]):
|
25 |
+
xmin, ymin, xmax, ymax = box.tolist()
|
26 |
+
ax.add_patch(patches.Rectangle(
|
27 |
+
(xmin, ymin), xmax - xmin, ymax - ymin,
|
28 |
+
linewidth=2, edgecolor='red', facecolor='none'))
|
29 |
+
ax.text(xmin, ymin, f"{model.config.id2label[label.item()]}: {round(score.item(), 2)}",
|
30 |
+
bbox=dict(facecolor='yellow', alpha=0.5), fontsize=8)
|
31 |
+
|
32 |
+
# Save output to bytes buffer
|
33 |
+
buf = io.BytesIO()
|
34 |
+
plt.axis("off")
|
35 |
+
plt.savefig(buf, format='png', bbox_inches='tight', pad_inches=0)
|
36 |
+
plt.close(fig)
|
37 |
+
buf.seek(0)
|
38 |
+
return Image.open(buf)
|
39 |
+
|
40 |
+
# Create Gradio interface
|
41 |
+
interface = gr.Interface(fn=detect_objects,
|
42 |
+
inputs=gr.Image(type="pil"),
|
43 |
+
outputs="image",
|
44 |
+
title="DETR Object Detection",
|
45 |
+
description="Upload an image to detect objects using Facebook's DETR model.")
|
46 |
+
|
47 |
+
# Launch the app locally
|
48 |
+
if __name__ == "__main__":
|
49 |
+
interface.launch(share=True)
|