Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
|
@@ -14,16 +14,28 @@ from transformers import AutoProcessor, AutoModelForZeroShotObjectDetection
|
|
| 14 |
# Add supervision for better visualization
|
| 15 |
import supervision as sv
|
| 16 |
|
| 17 |
-
# Model
|
| 18 |
-
|
|
|
|
|
|
|
|
|
|
| 19 |
|
| 20 |
-
#
|
| 21 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
| 22 |
-
|
| 23 |
-
|
|
|
|
| 24 |
|
| 25 |
@spaces.GPU
|
| 26 |
-
def run_grounding(input_image, grounding_caption, box_threshold, text_threshold):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 27 |
# Convert numpy array to PIL Image if needed
|
| 28 |
if isinstance(input_image, np.ndarray):
|
| 29 |
if input_image.ndim == 3:
|
|
@@ -63,8 +75,6 @@ def run_grounding(input_image, grounding_caption, box_threshold, text_threshold)
|
|
| 63 |
|
| 64 |
for i, (box, score, label) in enumerate(zip(result["boxes"], result["scores"], result["labels"])):
|
| 65 |
# box is xyxy format [xmin, ymin, xmax, ymax]
|
| 66 |
-
if label.strip() == "":
|
| 67 |
-
continue
|
| 68 |
xyxy = box.tolist()
|
| 69 |
boxes.append(xyxy)
|
| 70 |
labels.append(label)
|
|
@@ -144,12 +154,18 @@ if __name__ == "__main__":
|
|
| 144 |
}
|
| 145 |
"""
|
| 146 |
with gr.Blocks(css=css) as demo:
|
| 147 |
-
gr.Markdown("<h1><center>MM Grounding DINO Base<h1><center>")
|
| 148 |
-
gr.Markdown("<h3><center>Open-World Detection with
|
| 149 |
|
| 150 |
with gr.Row():
|
| 151 |
with gr.Column():
|
| 152 |
input_image = gr.Image(label="Input Image", type="pil")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 153 |
grounding_caption = gr.Textbox(
|
| 154 |
label="Detection Prompt (lowercase + each ends with a dot)",
|
| 155 |
value="a person. a car."
|
|
@@ -181,16 +197,16 @@ if __name__ == "__main__":
|
|
| 181 |
|
| 182 |
run_button.click(
|
| 183 |
fn=run_grounding,
|
| 184 |
-
inputs=[input_image, grounding_caption, box_threshold, text_threshold],
|
| 185 |
outputs=[gallery, det_text]
|
| 186 |
)
|
| 187 |
|
| 188 |
gr.Examples(
|
| 189 |
examples=[
|
| 190 |
-
["000000039769.jpg", "a cat. a remote control.", 0.3, 0.25],
|
| 191 |
-
["KakaoTalk_20250430_163200504.jpg", "cup. screen. hand.", 0.3, 0.25]
|
| 192 |
],
|
| 193 |
-
inputs=[input_image, grounding_caption, box_threshold, text_threshold],
|
| 194 |
outputs=[gallery, det_text],
|
| 195 |
fn=run_grounding,
|
| 196 |
cache_examples=True,
|
|
|
|
| 14 |
# Add supervision for better visualization
|
| 15 |
import supervision as sv
|
| 16 |
|
| 17 |
+
# Model IDs for Hugging Face
|
| 18 |
+
MODEL_IDS = {
|
| 19 |
+
"MM Grounding DINO Large": "rziga/mm_grounding_dino_large_all",
|
| 20 |
+
"MM Grounding DINO Base": "rziga/mm_grounding_dino_base_all"
|
| 21 |
+
}
|
| 22 |
|
| 23 |
+
# Global variables for model caching
|
| 24 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
| 25 |
+
loaded_model_name = None
|
| 26 |
+
processor = None
|
| 27 |
+
model = None
|
| 28 |
|
| 29 |
@spaces.GPU
|
| 30 |
+
def run_grounding(input_image, grounding_caption, model_choice, box_threshold, text_threshold):
|
| 31 |
+
global loaded_model_name, processor, model
|
| 32 |
+
|
| 33 |
+
# Load or reload model if changed
|
| 34 |
+
if loaded_model_name != model_choice:
|
| 35 |
+
model_id = MODEL_IDS[model_choice]
|
| 36 |
+
processor = AutoProcessor.from_pretrained(model_id)
|
| 37 |
+
model = AutoModelForZeroShotObjectDetection.from_pretrained(model_id).to(device)
|
| 38 |
+
loaded_model_name = model_choice
|
| 39 |
# Convert numpy array to PIL Image if needed
|
| 40 |
if isinstance(input_image, np.ndarray):
|
| 41 |
if input_image.ndim == 3:
|
|
|
|
| 75 |
|
| 76 |
for i, (box, score, label) in enumerate(zip(result["boxes"], result["scores"], result["labels"])):
|
| 77 |
# box is xyxy format [xmin, ymin, xmax, ymax]
|
|
|
|
|
|
|
| 78 |
xyxy = box.tolist()
|
| 79 |
boxes.append(xyxy)
|
| 80 |
labels.append(label)
|
|
|
|
| 154 |
}
|
| 155 |
"""
|
| 156 |
with gr.Blocks(css=css) as demo:
|
| 157 |
+
gr.Markdown("<h1><center>MM Grounding DINO (Large & Base)<h1><center>")
|
| 158 |
+
gr.Markdown("<h3><center>Open-World Detection with MM Grounding DINO Models<h3><center>")
|
| 159 |
|
| 160 |
with gr.Row():
|
| 161 |
with gr.Column():
|
| 162 |
input_image = gr.Image(label="Input Image", type="pil")
|
| 163 |
+
model_choice = gr.Radio(
|
| 164 |
+
choices=list(MODEL_IDS.keys()),
|
| 165 |
+
value="MM Grounding DINO Large",
|
| 166 |
+
label="Select Model",
|
| 167 |
+
info="Choose between Large (better performance) or Base (faster) model"
|
| 168 |
+
)
|
| 169 |
grounding_caption = gr.Textbox(
|
| 170 |
label="Detection Prompt (lowercase + each ends with a dot)",
|
| 171 |
value="a person. a car."
|
|
|
|
| 197 |
|
| 198 |
run_button.click(
|
| 199 |
fn=run_grounding,
|
| 200 |
+
inputs=[input_image, grounding_caption, model_choice, box_threshold, text_threshold],
|
| 201 |
outputs=[gallery, det_text]
|
| 202 |
)
|
| 203 |
|
| 204 |
gr.Examples(
|
| 205 |
examples=[
|
| 206 |
+
["000000039769.jpg", "a cat. a remote control.", "MM Grounding DINO Large", 0.3, 0.25],
|
| 207 |
+
["KakaoTalk_20250430_163200504.jpg", "cup. screen. hand.", "MM Grounding DINO Base", 0.3, 0.25]
|
| 208 |
],
|
| 209 |
+
inputs=[input_image, grounding_caption, model_choice, box_threshold, text_threshold],
|
| 210 |
outputs=[gallery, det_text],
|
| 211 |
fn=run_grounding,
|
| 212 |
cache_examples=True,
|