RSReasonNet / ui.py
sosppxo's picture
Replace highway service area example with storage tank 04077
b036ed3 verified
Raw History Blame Contribute Delete
7.57 kB
"""Gradio layout and bundled examples for the RSReasonNet demo."""
import json
from pathlib import Path
import gradio as gr
EXAMPLE_DIR = Path(__file__).resolve().parent / "example"
EXAMPLE_TITLES = {
"03261.jpg": "Ship · Maritime logistics",
"04077.jpg": "Storage tank · Industrial infrastructure",
"02026.jpg": "Tennis court · Everyday language",
}
def load_examples():
"""Keep each image paired with its original annotated instruction."""
with (EXAMPLE_DIR / "annotations.json").open(encoding="utf-8") as handle:
annotations = json.load(handle)
examples = []
for annotation in annotations:
image_path = EXAMPLE_DIR / annotation["id"]
if not image_path.is_file():
raise FileNotFoundError(f"Example image is missing: {image_path.name}")
examples.append((str(image_path), annotation["sentence"]))
return examples
CSS = """
.gradio-container { max-width: 1120px !important; margin: auto; }
#rs-header { text-align: center; padding: 22px 16px 18px; }
#rs-header .eyebrow {
display: inline-block; padding: 5px 12px; border-radius: 999px;
border: 1px solid var(--border-color-primary);
color: var(--body-text-color-subdued); font-size: 12px; letter-spacing: .05em;
}
#rs-header h1 {
margin: 12px 0 8px; font-size: clamp(30px, 5vw, 42px);
font-weight: 700; letter-spacing: -.04em; color: var(--body-text-color);
}
#rs-header p {
margin: 0 auto; max-width: 640px; color: var(--body-text-color-subdued);
font-size: 15px; line-height: 1.6;
}
.rs-panel { padding: 18px !important; border-radius: 16px !important; }
.rs-section h2 { font-size: 18px !important; margin: 0 0 4px !important; }
.rs-section p { color: var(--body-text-color-subdued); font-size: 13px; }
#rs-run { min-height: 44px; font-weight: 600; }
#rs-examples-heading { margin-top: 12px; }
#rs-examples { border-radius: 14px; }
#rs-examples .grid-wrap { height: auto; min-height: 0; max-height: none; }
#rs-examples .grid-container {
grid-template-columns: repeat(3, minmax(0, 1fr));
grid-template-rows: none;
}
#rs-examples .thumbnail-item {
cursor: pointer; aspect-ratio: auto; height: 210px;
}
#rs-examples .thumbnail-item:hover { border-color: var(--color-accent); }
#rs-examples .caption-label { font-size: 13px; }
#rs-note { color: var(--body-text-color-subdued); font-size: 12px; }
@media (max-width: 640px) {
#rs-header { padding: 14px 4px; }
.rs-panel { padding: 12px !important; }
#rs-examples .grid-container { grid-template-columns: minmax(0, 1fr); }
}
"""
def build_demo(predict):
examples = load_examples()
theme = gr.themes.Soft(
primary_hue="blue", secondary_hue="slate", neutral_hue="slate",
font=["system-ui", "sans-serif"],
)
with gr.Blocks(
title="RSReasonNet · Reasoning Segmentation",
theme=theme,
css=CSS,
analytics_enabled=False,
) as demo:
gr.HTML("""
<header id="rs-header">
<span class="eyebrow">🛰️ REMOTE SENSING · REASONING SEGMENTATION</span>
<h1>RSReasonNet</h1>
<p>Describe what you are looking for. Explore how language guides
segmentation in satellite imagery.</p>
</header>
""")
with gr.Row(equal_height=True):
with gr.Column(min_width=300, variant="panel", elem_classes="rs-panel"):
gr.Markdown(
"## 1. Image & instruction\nUpload an image or choose an example below.",
elem_classes="rs-section",
)
input_image = gr.Image(
value=examples[0][0], type="filepath", sources=["upload"],
label="Satellite image", height=320, elem_id="rs-input-image",
show_share_button=False,
)
instruction = gr.Textbox(
value=examples[0][1], label="What should the model find?",
placeholder="Describe an object, its purpose, or a clue in English…",
lines=3, max_lines=5,
info="Use an English instruction (up to 1,024 characters).",
)
with gr.Row():
run_button = gr.Button(
"Run segmentation", variant="primary", scale=3, elem_id="rs-run",
)
clear_button = gr.ClearButton(value="Clear", scale=1)
with gr.Column(min_width=300, variant="panel", elem_classes="rs-panel"):
gr.Markdown(
"## 2. Segmentation & response\nRun the model to see the region it identifies.",
elem_classes="rs-section",
)
output_image = gr.Image(
type="numpy", label="Segmentation overlay", height=320,
interactive=False, show_share_button=False, elem_id="rs-output-image",
)
text_output = gr.Textbox(
label="Model response", lines=3, max_lines=6,
interactive=False, show_copy_button=True,
placeholder="The model's text response will appear here.",
)
gr.Markdown(
"🔴 **Red overlay** marks the predicted region. "
"Some instructions may return text without a mask.",
elem_id="rs-note",
)
gr.Markdown(
"## Try an example\nClick a thumbnail to load its image and original instruction, "
"then select **Run segmentation**.",
elem_id="rs-examples-heading", elem_classes="rs-section",
)
gallery = gr.Gallery(
value=[(path, EXAMPLE_TITLES.get(Path(path).name, Path(path).stem))
for path, _ in examples],
label="Example images", show_label=False,
columns=3, rows=1,
object_fit="contain", allow_preview=False, interactive=False,
show_share_button=False, show_download_button=False,
show_fullscreen_button=False, elem_id="rs-examples",
)
# Keep the existing API argument order: instruction first, image second.
run_event = gr.on(
triggers=[run_button.click, instruction.submit], fn=predict,
inputs=[instruction, input_image], outputs=[output_image, text_output],
api_name="predict", concurrency_limit=1, concurrency_id="segmentation",
show_progress="full",
)
def select_example(event: gr.SelectData):
image_path, prompt = examples[event.index]
# Allow the same card to be selected again after clearing or editing.
return image_path, prompt, None, "", gr.Gallery(selected_index=None)
gallery.select(
select_example, inputs=None,
outputs=[input_image, instruction, output_image, text_output, gallery],
queue=False, api_name=False, cancels=[run_event],
)
clear_button.add([input_image, instruction, output_image, text_output])
clear_button.click(fn=None, cancels=[run_event], queue=False, api_name=False)
# A result belongs to its input pair; remove it when that pair is edited.
gr.on(
triggers=[input_image.input, instruction.input], fn=lambda: (None, ""),
inputs=None, outputs=[output_image, text_output], queue=False,
api_name=False, cancels=[run_event],
)
return demo