initial setup
Browse files- app.py +140 -4
- requirements.txt +8 -0
app.py
CHANGED
@@ -1,7 +1,143 @@
|
|
1 |
import gradio as gr
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2 |
|
3 |
-
def greet(name):
|
4 |
-
return "Hello " + name + "!!"
|
5 |
|
6 |
-
|
7 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
import gradio as gr
|
2 |
+
import spaces
|
3 |
+
from transformers import Qwen2VLForConditionalGeneration, AutoTokenizer, AutoProcessor
|
4 |
+
from qwen_vl_utils import process_vision_info
|
5 |
+
import torch
|
6 |
+
import base64
|
7 |
+
from PIL import Image, ImageDraw
|
8 |
+
from io import BytesIO
|
9 |
+
import re
|
10 |
|
|
|
|
|
11 |
|
12 |
+
models = {
|
13 |
+
"OS-Copilot/OS-Atlas-Base-7B": Qwen2VLForConditionalGeneration.from_pretrained("OS-Copilot/OS-Atlas-Base-7B", torch_dtype="auto", device_map="auto"),
|
14 |
+
}
|
15 |
+
|
16 |
+
processors = {
|
17 |
+
"OS-Copilot/OS-Atlas-Base-7B": AutoProcessor.from_pretrained("OS-Copilot/OS-Atlas-Base-7B")
|
18 |
+
}
|
19 |
+
|
20 |
+
|
21 |
+
def image_to_base64(image):
|
22 |
+
buffered = BytesIO()
|
23 |
+
image.save(buffered, format="PNG")
|
24 |
+
img_str = base64.b64encode(buffered.getvalue()).decode("utf-8")
|
25 |
+
return img_str
|
26 |
+
|
27 |
+
|
28 |
+
def draw_bounding_boxes(image, bounding_boxes, outline_color="red", line_width=2):
|
29 |
+
draw = ImageDraw.Draw(image)
|
30 |
+
for box in bounding_boxes:
|
31 |
+
xmin, ymin, xmax, ymax = box
|
32 |
+
draw.rectangle([xmin, ymin, xmax, ymax], outline=outline_color, width=line_width)
|
33 |
+
return image
|
34 |
+
|
35 |
+
|
36 |
+
def rescale_bounding_boxes(bounding_boxes, original_width, original_height, scaled_width=1000, scaled_height=1000):
|
37 |
+
x_scale = original_width / scaled_width
|
38 |
+
y_scale = original_height / scaled_height
|
39 |
+
rescaled_boxes = []
|
40 |
+
for box in bounding_boxes:
|
41 |
+
xmin, ymin, xmax, ymax = box
|
42 |
+
rescaled_box = [
|
43 |
+
xmin * x_scale,
|
44 |
+
ymin * y_scale,
|
45 |
+
xmax * x_scale,
|
46 |
+
ymax * y_scale
|
47 |
+
]
|
48 |
+
rescaled_boxes.append(rescaled_box)
|
49 |
+
return rescaled_boxes
|
50 |
+
|
51 |
+
|
52 |
+
@spaces.GPU
|
53 |
+
def run_example(image, text_input, system_prompt, model_id="OS-Copilot/OS-Atlas-Base-7B"):
|
54 |
+
model = models[model_id].eval()
|
55 |
+
processor = processors[model_id]
|
56 |
+
|
57 |
+
messages = [
|
58 |
+
{
|
59 |
+
"role": "user",
|
60 |
+
"content": [
|
61 |
+
{"type": "image", "image": f"data:image;base64,{image_to_base64(image)}"},
|
62 |
+
{"type": "text", "text": text_input},
|
63 |
+
],
|
64 |
+
}
|
65 |
+
]
|
66 |
+
|
67 |
+
text = processor.apply_chat_template(
|
68 |
+
messages, tokenize=False, add_generation_prompt=True
|
69 |
+
)
|
70 |
+
image_inputs, video_inputs = process_vision_info(messages)
|
71 |
+
inputs = processor(
|
72 |
+
text=[text],
|
73 |
+
images=image_inputs,
|
74 |
+
videos=video_inputs,
|
75 |
+
padding=True,
|
76 |
+
return_tensors="pt",
|
77 |
+
)
|
78 |
+
inputs = inputs.to("cuda")
|
79 |
+
|
80 |
+
generated_ids = model.generate(**inputs, max_new_tokens=128)
|
81 |
+
generated_ids_trimmed = [
|
82 |
+
out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
|
83 |
+
]
|
84 |
+
output_text = processor.batch_decode(
|
85 |
+
generated_ids_trimmed, skip_special_tokens=False, clean_up_tokenization_spaces=False
|
86 |
+
)
|
87 |
+
print(output_text)
|
88 |
+
|
89 |
+
object_ref_pattern = r"<\|object_ref_start\|>(.*?)<\|object_ref_end\|>"
|
90 |
+
box_pattern = r"<\|box_start\|>(.*?)<\|box_end\|>"
|
91 |
+
|
92 |
+
object_ref = re.search(object_ref_pattern, text).group(1)
|
93 |
+
box_content = re.search(box_pattern, text).group(1)
|
94 |
+
|
95 |
+
boxes = [tuple(map(int, pair.strip("()").split(','))) for pair in box_content.split("),(")]
|
96 |
+
boxes = [boxes[0][0], boxes[0][1], boxes[1][0], boxes[1][1]]
|
97 |
+
|
98 |
+
scaled_boxes = rescale_bounding_boxes(boxes, image.width, image.height)
|
99 |
+
return output_text, boxes, draw_bounding_boxes(image, scaled_boxes)
|
100 |
+
|
101 |
+
css = """
|
102 |
+
#output {
|
103 |
+
height: 500px;
|
104 |
+
overflow: auto;
|
105 |
+
border: 1px solid #ccc;
|
106 |
+
}
|
107 |
+
"""
|
108 |
+
default_system_prompt = "You are a helpfull assistant to detect objects in images. When asked to detect elements based on a description you return bounding boxes for all elements in the form of [xmin, ymin, xmax, ymax] whith the values beeing scaled to 1000 by 1000 pixels. When there are more than one result, answer with a list of bounding boxes in the form of [[xmin, ymin, xmax, ymax], [xmin, ymin, xmax, ymax], ...]."
|
109 |
+
|
110 |
+
with gr.Blocks(css=css) as demo:
|
111 |
+
gr.Markdown(
|
112 |
+
"""
|
113 |
+
# OS-Atlas Demo
|
114 |
+
""")
|
115 |
+
with gr.Tab(label="OS-Atlas Input"):
|
116 |
+
with gr.Row():
|
117 |
+
with gr.Column():
|
118 |
+
input_img = gr.Image(label="Input Image", type="pil")
|
119 |
+
model_selector = gr.Dropdown(choices=list(models.keys()), label="Model", value="OS-Copilot/OS-Atlas-Base-7B")
|
120 |
+
system_prompt = gr.Textbox(label="System Prompt", value=default_system_prompt)
|
121 |
+
text_input = gr.Textbox(label="User Prompt")
|
122 |
+
submit_btn = gr.Button(value="Submit")
|
123 |
+
with gr.Column():
|
124 |
+
model_output_text = gr.Textbox(label="Model Output Text")
|
125 |
+
parsed_boxes = gr.Textbox(label="Parsed Boxes")
|
126 |
+
annotated_image = gr.Image(label="Annotated Image")
|
127 |
+
|
128 |
+
gr.Examples(
|
129 |
+
examples=[
|
130 |
+
["assets/image1.jpg", "detect goats", default_system_prompt],
|
131 |
+
["assets/image2.jpg", "detect blue button", default_system_prompt],
|
132 |
+
["assets/image3.jpg", "detect person on bike", default_system_prompt],
|
133 |
+
],
|
134 |
+
inputs=[input_img, text_input, system_prompt],
|
135 |
+
outputs=[model_output_text, parsed_boxes, annotated_image],
|
136 |
+
fn=run_example,
|
137 |
+
cache_examples=True,
|
138 |
+
label="Try examples"
|
139 |
+
)
|
140 |
+
|
141 |
+
submit_btn.click(run_example, [input_img, text_input, system_prompt, model_selector], [model_output_text, parsed_boxes, annotated_image])
|
142 |
+
|
143 |
+
demo.launch(debug=True)
|
requirements.txt
ADDED
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
numpy==1.24.4
|
2 |
+
Pillow==10.3.0
|
3 |
+
Requests==2.31.0
|
4 |
+
torch
|
5 |
+
torchvision
|
6 |
+
transformers
|
7 |
+
accelerate==0.30.0
|
8 |
+
qwen-vl-utils
|