Upload 2 files
Browse files- app.py +334 -0
- requirements.txt +11 -0
app.py
ADDED
@@ -0,0 +1,334 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import json
|
2 |
+
import time
|
3 |
+
|
4 |
+
import gradio as gr
|
5 |
+
import numpy as np
|
6 |
+
import torch
|
7 |
+
# from gradio.themes.Soft import Soft
|
8 |
+
from PIL import Image
|
9 |
+
from qwen_vl_utils import process_vision_info
|
10 |
+
from transformers import (
|
11 |
+
AutoProcessor,
|
12 |
+
Gemma3ForConditionalGeneration,
|
13 |
+
Qwen2_5_VLForConditionalGeneration,
|
14 |
+
)
|
15 |
+
|
16 |
+
from spaces import GPU
|
17 |
+
import supervision as sv
|
18 |
+
|
19 |
+
# --- Config ---
|
20 |
+
# IMPORTANT: Both models are gated. You must be logged in to your Hugging Face account
|
21 |
+
# and have been granted access to use them.
|
22 |
+
# from huggingface_hub import login
|
23 |
+
# login()
|
24 |
+
|
25 |
+
model_qwen_id = "Qwen/Qwen2.5-VL-3B-Instruct"
|
26 |
+
model_gemma_id = "google/gemma-3-4b-it"
|
27 |
+
|
28 |
+
# Load Qwen Model
|
29 |
+
model_qwen = Qwen2_5_VLForConditionalGeneration.from_pretrained(
|
30 |
+
model_qwen_id, torch_dtype="auto", device_map="auto"
|
31 |
+
)
|
32 |
+
min_pixels = 224 * 224
|
33 |
+
max_pixels = 1024 * 1024
|
34 |
+
processor_qwen = AutoProcessor.from_pretrained(
|
35 |
+
model_qwen_id, min_pixels=min_pixels, max_pixels=max_pixels
|
36 |
+
)
|
37 |
+
|
38 |
+
# Load Gemma Model
|
39 |
+
model_gemma = Gemma3ForConditionalGeneration.from_pretrained(
|
40 |
+
model_gemma_id,
|
41 |
+
torch_dtype=torch.bfloat16, # Recommended dtype for Gemma
|
42 |
+
device_map="auto"
|
43 |
+
)
|
44 |
+
processor_gemma = AutoProcessor.from_pretrained(model_gemma_id)
|
45 |
+
|
46 |
+
|
47 |
+
def extract_model_short_name(model_id):
|
48 |
+
return model_id.split("/")[-1].replace("-", " ").replace("_", " ")
|
49 |
+
|
50 |
+
|
51 |
+
model_qwen_name = extract_model_short_name(model_qwen_id) # → "Qwen2.5 VL 3B Instruct"
|
52 |
+
model_gemma_name = extract_model_short_name(model_gemma_id) # → "gemma 3 4b it"
|
53 |
+
|
54 |
+
|
55 |
+
def create_annotated_image(image, json_data, height, width):
|
56 |
+
try:
|
57 |
+
# Standardize parsing for outputs wrapped in markdown
|
58 |
+
if "```json" in json_data:
|
59 |
+
parsed_json_data = json_data.split("```json")[1].split("```")[0]
|
60 |
+
else:
|
61 |
+
parsed_json_data = json_data
|
62 |
+
bbox_data = json.loads(parsed_json_data)
|
63 |
+
except Exception:
|
64 |
+
# If parsing fails, return the original image
|
65 |
+
return image
|
66 |
+
|
67 |
+
# Ensure bbox_data is a list
|
68 |
+
if not isinstance(bbox_data, list):
|
69 |
+
bbox_data = [bbox_data]
|
70 |
+
|
71 |
+
|
72 |
+
original_width, original_height = image.size
|
73 |
+
x_scale = original_width / width
|
74 |
+
y_scale = original_height / height
|
75 |
+
|
76 |
+
points = []
|
77 |
+
point_labels = []
|
78 |
+
|
79 |
+
annotated_image = np.array(image.convert("RGB"))
|
80 |
+
detections_exist = False
|
81 |
+
|
82 |
+
# Check if there are bounding boxes in the data to create detections
|
83 |
+
if any("box_2d" in item for item in bbox_data):
|
84 |
+
detections_exist = True
|
85 |
+
# Use Qwen parser as a generic VLM parser for bounding boxes
|
86 |
+
detections = sv.Detections.from_vlm(vlm = sv.VLM.QWEN_2_5_VL,
|
87 |
+
result=json_data,
|
88 |
+
# resolution_wh is the size model "sees"
|
89 |
+
resolution_wh=(width, height))
|
90 |
+
bounding_box_annotator = sv.BoxAnnotator(color_lookup=sv.ColorLookup.INDEX)
|
91 |
+
label_annotator = sv.LabelAnnotator(color_lookup=sv.ColorLookup.INDEX)
|
92 |
+
|
93 |
+
annotated_image = bounding_box_annotator.annotate(
|
94 |
+
scene=annotated_image, detections=detections
|
95 |
+
)
|
96 |
+
annotated_image = label_annotator.annotate(
|
97 |
+
scene=annotated_image, detections=detections
|
98 |
+
)
|
99 |
+
|
100 |
+
# Handle points separately
|
101 |
+
for item in bbox_data:
|
102 |
+
label = item.get("label", "")
|
103 |
+
if "point_2d" in item:
|
104 |
+
x, y = item["point_2d"]
|
105 |
+
scaled_x = int(x * x_scale)
|
106 |
+
scaled_y = int(y * y_scale)
|
107 |
+
points.append([scaled_x, scaled_y])
|
108 |
+
point_labels.append(label)
|
109 |
+
|
110 |
+
if points:
|
111 |
+
points_array = np.array(points).reshape(1, -1, 2)
|
112 |
+
key_points = sv.KeyPoints(xy=points_array)
|
113 |
+
vertex_annotator = sv.VertexAnnotator(radius=5, color=sv.Color.BLUE)
|
114 |
+
annotated_image = vertex_annotator.annotate(
|
115 |
+
scene=annotated_image, key_points=key_points
|
116 |
+
)
|
117 |
+
|
118 |
+
return Image.fromarray(annotated_image)
|
119 |
+
|
120 |
+
|
121 |
+
@GPU
|
122 |
+
def detect_qwen(image, prompt):
|
123 |
+
messages = [
|
124 |
+
{
|
125 |
+
"role": "user",
|
126 |
+
"content": [
|
127 |
+
{"type": "image", "image": image},
|
128 |
+
{"type": "text", "text": prompt},
|
129 |
+
],
|
130 |
+
}
|
131 |
+
]
|
132 |
+
|
133 |
+
t0 = time.perf_counter()
|
134 |
+
text = processor_qwen.apply_chat_template(
|
135 |
+
messages, tokenize=False, add_generation_prompt=True
|
136 |
+
)
|
137 |
+
image_inputs, video_inputs = process_vision_info(messages)
|
138 |
+
inputs = processor_qwen(
|
139 |
+
text=[text],
|
140 |
+
images=image_inputs,
|
141 |
+
videos=video_inputs,
|
142 |
+
padding=True,
|
143 |
+
return_tensors="pt",
|
144 |
+
).to(model_qwen.device)
|
145 |
+
|
146 |
+
generated_ids = model_qwen.generate(**inputs, max_new_tokens=1024)
|
147 |
+
generated_ids_trimmed = [
|
148 |
+
out_ids[len(in_ids) :]
|
149 |
+
for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
|
150 |
+
]
|
151 |
+
output_text = processor_qwen.batch_decode(
|
152 |
+
generated_ids_trimmed,
|
153 |
+
do_sample=True,
|
154 |
+
skip_special_tokens=True,
|
155 |
+
clean_up_tokenization_spaces=False,
|
156 |
+
)[0]
|
157 |
+
elapsed_ms = (time.perf_counter() - t0) * 1_000
|
158 |
+
|
159 |
+
# These dimensions are specific to how Qwen's processor handles images
|
160 |
+
input_height = inputs["image_grid_thw"][0][1] * 14
|
161 |
+
input_width = inputs["image_grid_thw"][0][2] * 14
|
162 |
+
|
163 |
+
annotated_image = create_annotated_image(
|
164 |
+
image, output_text, input_height, input_width
|
165 |
+
)
|
166 |
+
|
167 |
+
time_taken = f"**Inference time ({model_qwen_name}):** {elapsed_ms:.0f} ms"
|
168 |
+
return annotated_image, output_text, time_taken
|
169 |
+
|
170 |
+
|
171 |
+
@GPU
|
172 |
+
def detect_gemma(image, prompt):
|
173 |
+
messages = [
|
174 |
+
{
|
175 |
+
"role": "user",
|
176 |
+
"content": [
|
177 |
+
{"type": "image", "image": image},
|
178 |
+
{"type": "text", "text": prompt},
|
179 |
+
],
|
180 |
+
}
|
181 |
+
]
|
182 |
+
|
183 |
+
t0 = time.perf_counter()
|
184 |
+
inputs = processor_gemma.apply_chat_template(
|
185 |
+
messages,
|
186 |
+
add_generation_prompt=True,
|
187 |
+
tokenize=True,
|
188 |
+
return_dict=True,
|
189 |
+
return_tensors="pt"
|
190 |
+
).to(model_gemma.device)
|
191 |
+
|
192 |
+
input_len = inputs["input_ids"].shape[-1]
|
193 |
+
|
194 |
+
with torch.inference_mode():
|
195 |
+
generation = model_gemma.generate(**inputs, max_new_tokens=1024, do_sample=False)
|
196 |
+
|
197 |
+
generation_trimmed = generation[0][input_len:]
|
198 |
+
output_text = processor_gemma.decode(generation_trimmed, skip_special_tokens=True)
|
199 |
+
elapsed_ms = (time.perf_counter() - t0) * 1_000
|
200 |
+
|
201 |
+
# Gemma's vision encoder normalizes images to a fixed size (e.g., 896x896)
|
202 |
+
input_height = 896
|
203 |
+
input_width = 896
|
204 |
+
|
205 |
+
annotated_image = create_annotated_image(
|
206 |
+
image, output_text, input_height, input_width
|
207 |
+
)
|
208 |
+
|
209 |
+
time_taken = f"**Inference time ({model_gemma_name}):** {elapsed_ms:.0f} ms"
|
210 |
+
return annotated_image, output_text, time_taken
|
211 |
+
|
212 |
+
|
213 |
+
def detect(image, prompt_model_1, prompt_model_2):
|
214 |
+
STANDARD_SIZE = (1024, 1024)
|
215 |
+
image.thumbnail(STANDARD_SIZE)
|
216 |
+
|
217 |
+
annotated_image_model_1, output_text_model_1, timing_1 = detect_qwen(
|
218 |
+
image, prompt_model_1
|
219 |
+
)
|
220 |
+
annotated_image_model_2, output_text_model_2, timing_2 = detect_gemma(
|
221 |
+
image, prompt_model_2
|
222 |
+
)
|
223 |
+
|
224 |
+
return (
|
225 |
+
annotated_image_model_1,
|
226 |
+
output_text_model_1,
|
227 |
+
timing_1,
|
228 |
+
annotated_image_model_2,
|
229 |
+
output_text_model_2,
|
230 |
+
timing_2,
|
231 |
+
)
|
232 |
+
|
233 |
+
|
234 |
+
css_hide_share = """
|
235 |
+
button#gradio-share-link-button-0 {
|
236 |
+
display: none !important;
|
237 |
+
}
|
238 |
+
"""
|
239 |
+
|
240 |
+
# --- Gradio Interface ---
|
241 |
+
with gr.Blocks(theme=gr.themes.Soft(), css=css_hide_share) as demo:
|
242 |
+
gr.Markdown("# Object Detection & Understanding: Qwen vs. Gemma")
|
243 |
+
gr.Markdown(
|
244 |
+
"### Compare object detection, visual grounding, and keypoint detection using natural language prompts with two leading VLMs."
|
245 |
+
)
|
246 |
+
gr.Markdown("""
|
247 |
+
*Powered by [Qwen2.5-VL 3B](https://huggingface.co/Qwen/Qwen2.5-VL-3B-Instruct) and [Gemma 3 4B IT](https://huggingface.co/google/gemma-3-4b-it). For best results, ask the model to return a JSON list in a markdown block. Inspired by the [HF Team's space](https://huggingface.co/spaces/sergiopaniego/vlm_object_understanding), selecting `detect` for categories with "Object Detection" `point` for the ones with "Keypoint Detection", and reasoning-based querying for all others.*
|
248 |
+
""")
|
249 |
+
|
250 |
+
with gr.Row():
|
251 |
+
with gr.Column(scale=2):
|
252 |
+
image_input = gr.Image(label="Upload an image", type="pil", height=400)
|
253 |
+
prompt_input_model_1 = gr.Textbox(
|
254 |
+
label=f"Enter your prompt for {model_qwen_name}",
|
255 |
+
placeholder="e.g., Detect all red cars. Return a JSON list with 'box_2d' and 'label'.",
|
256 |
+
)
|
257 |
+
prompt_input_model_2 = gr.Textbox(
|
258 |
+
label=f"Enter your prompt for {model_gemma_name}",
|
259 |
+
placeholder="e.g., Detect all red cars. Return a JSON list with 'box_2d' and 'label'.",
|
260 |
+
)
|
261 |
+
generate_btn = gr.Button(value="Generate")
|
262 |
+
|
263 |
+
with gr.Column(scale=1):
|
264 |
+
output_image_model_1 = gr.Image(
|
265 |
+
type="pil", label=f"Annotated image from {model_qwen_name}", height=400
|
266 |
+
)
|
267 |
+
output_textbox_model_1 = gr.Textbox(
|
268 |
+
label=f"Model response from {model_qwen_name}", lines=10
|
269 |
+
)
|
270 |
+
output_time_model_1 = gr.Markdown()
|
271 |
+
|
272 |
+
with gr.Column(scale=1):
|
273 |
+
output_image_model_2 = gr.Image(
|
274 |
+
type="pil",
|
275 |
+
label=f"Annotated image from {model_gemma_name}",
|
276 |
+
height=400,
|
277 |
+
)
|
278 |
+
output_textbox_model_2 = gr.Textbox(
|
279 |
+
label=f"Model response from {model_gemma_name}", lines=10
|
280 |
+
)
|
281 |
+
output_time_model_2 = gr.Markdown()
|
282 |
+
|
283 |
+
gr.Markdown("### Examples")
|
284 |
+
|
285 |
+
prompt_obj_detect = "Detect all objects in this image. For each object, provide a 'box_2d' and a 'label'. Return the output as a JSON list inside a markdown block."
|
286 |
+
prompt_candy_detect = "Detect all individual candies in this image. For each, provide a 'box_2d' and a 'label'. Return the output as a JSON list inside a markdown block."
|
287 |
+
prompt_car_count = "Count the number of red cars in the image."
|
288 |
+
prompt_candy_count = "Count the number of blue candies in the image."
|
289 |
+
prompt_car_keypoint = "Identify the red cars in this image. For each, detect its key points and return their positions as 'point_2d' in a JSON list inside a markdown block."
|
290 |
+
prompt_candy_keypoint = "Identify the blue candies in this image. For each, detect its key points and return their positions as 'point_2d' in a JSON list inside a markdown block."
|
291 |
+
prompt_car_ground = "Detect the red car that is leading in this image. Return its location with 'box_2d' and 'label' in a JSON list inside a markdown block."
|
292 |
+
prompt_candy_ground = "Detect the blue candy at the top of the group. Return its location with 'box_2d' and 'label' in a JSON list inside a markdown block."
|
293 |
+
|
294 |
+
|
295 |
+
example_prompts = [
|
296 |
+
["examples/example_1.jpg", prompt_obj_detect, prompt_obj_detect],
|
297 |
+
["examples/example_2.JPG", prompt_candy_detect, prompt_candy_detect],
|
298 |
+
["examples/example_1.jpg", prompt_car_count, prompt_car_count],
|
299 |
+
["examples/example_2.JPG", prompt_candy_count, prompt_candy_count],
|
300 |
+
["examples/example_1.jpg", prompt_car_keypoint, prompt_car_keypoint],
|
301 |
+
["examples/example_2.JPG", prompt_candy_keypoint, prompt_candy_keypoint],
|
302 |
+
["examples/example_1.jpg", prompt_car_ground, prompt_car_ground],
|
303 |
+
["examples/example_2.JPG", prompt_candy_ground, prompt_candy_ground],
|
304 |
+
]
|
305 |
+
|
306 |
+
gr.Examples(
|
307 |
+
examples=example_prompts,
|
308 |
+
inputs=[
|
309 |
+
image_input,
|
310 |
+
prompt_input_model_1,
|
311 |
+
prompt_input_model_2,
|
312 |
+
],
|
313 |
+
label="Click an example to populate the input",
|
314 |
+
)
|
315 |
+
|
316 |
+
generate_btn.click(
|
317 |
+
fn=detect,
|
318 |
+
inputs=[
|
319 |
+
image_input,
|
320 |
+
prompt_input_model_1,
|
321 |
+
prompt_input_model_2,
|
322 |
+
],
|
323 |
+
outputs=[
|
324 |
+
output_image_model_1,
|
325 |
+
output_textbox_model_1,
|
326 |
+
output_time_model_1,
|
327 |
+
output_image_model_2,
|
328 |
+
output_textbox_model_2,
|
329 |
+
output_time_model_2,
|
330 |
+
],
|
331 |
+
)
|
332 |
+
|
333 |
+
if __name__ == "__main__":
|
334 |
+
demo.launch()
|
requirements.txt
ADDED
@@ -0,0 +1,11 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
torch
|
2 |
+
transformers
|
3 |
+
datasets
|
4 |
+
bitsandbytes
|
5 |
+
Pillow
|
6 |
+
gradio
|
7 |
+
accelerate
|
8 |
+
qwen-vl-utils
|
9 |
+
torchvision
|
10 |
+
matplotlib
|
11 |
+
supervision
|