Jadebks's picture
Role 3 prototype bundle: app, minimal code, fold-0 baseline (no dataset images)
be666c7 verified
Raw History Blame Contribute Delete
4.81 kB
"""Hugging Face Space entry point for the Role 3 Gradio prototype.
Same app as app/app.py in the GitHub repository
(jadebeeks/face-verification-real-world-conditions); this file only adds the
Space setup below.
"""
import subprocess
import sys
from pathlib import Path
# facenet-pytorch (MTCNN face detector) pins old torch / numpy / Pillow versions,
# so it is installed without its dependencies, exactly as in the repository README.
try:
import facenet_pytorch # noqa: F401
except ImportError:
subprocess.check_call([sys.executable, "-m", "pip", "install", "--no-deps", "facenet-pytorch==2.6.0"])
sys.path.insert(0, str(Path(__file__).parent / "src"))
import matplotlib
matplotlib.use("Agg") # draw charts off-screen inside the web server
import gradio as gr
from face_verification.conditions import MAX_LEVEL
from face_verification.prototype import (
CONDITION_LABELS, available_models, get_model, prepare, sweep, sweep_figure, verify,
)
LABEL_TO_KEY = {v: k for k, v in CONDITION_LABELS.items()}
INTRO = """# Face verification under real-world conditions
Upload a **reference photo** (like a passport photo) and a **second photo** of the same or a different person.
The second photo is degraded with the condition you choose, as if captured at a gate in poor conditions.
The model compares the two faces and decides **same person** or **different people** using a threshold
chosen on validation people.
*Research prototype for a university project (IE University, AI/ML): not an identity-check system. Uploaded photos are used only to compute the comparison and are deleted from the server within an hour. Please only upload photos you have the right to use.*"""
def on_upload(img):
"""Detect and crop a face once per upload; keep it in state."""
if img is None:
return None, None
crop, landmarks, warning = prepare(img)
return {"crop": crop, "landmarks": landmarks}, (warning or "")
def run(state1, state2, model_name, condition_label, level):
if state1 is None or state2 is None:
return None, None, "Upload both photos to compare them.", None
if not model_name:
return None, None, "No trained model found. See README: train the baseline first.", None
model = get_model(model_name)
cond = LABEL_TO_KEY[condition_label]
level = int(level)
r = verify(model, state1["crop"], state2["crop"], cond, level, state2["landmarks"])
scores = sweep(model, state1["crop"], state2["crop"], cond, state2["landmarks"])
verdict = "✅ **Same person**" if r["same"] else "❌ **Different people**"
margin = r["score"] - r["threshold"]
text = (f"### {verdict}\n"
f"Score **{r['score']:.3f}** vs threshold **{r['threshold']:.3f}** "
f"(margin {margin:+.3f}) \n"
f"Condition on photo 2: {condition_label}, level {level} ({r['condition_value']}) \n"
f"Model: {model_name}")
fig = sweep_figure(scores, r["threshold"], level, cond)
return state1["crop"], r["degraded"], text, fig
def build() -> gr.Blocks:
models = available_models()
with gr.Blocks(title="Face verification under real-world conditions",
delete_cache=(1800, 3600)) as demo:
gr.Markdown(INTRO)
state1, state2 = gr.State(None), gr.State(None)
with gr.Row():
with gr.Column():
img1 = gr.Image(label="Photo 1: reference (kept clean)", type="numpy")
warn1 = gr.Markdown()
with gr.Column():
img2 = gr.Image(label="Photo 2: live capture (gets the condition)", type="numpy")
warn2 = gr.Markdown()
with gr.Row():
model_dd = gr.Dropdown(models, value=models[0] if models else None, label="Model")
cond_dd = gr.Dropdown(list(CONDITION_LABELS.values()), value=list(CONDITION_LABELS.values())[0],
label="Condition on photo 2")
level_sl = gr.Slider(0, MAX_LEVEL, value=0, step=1, label="Strength (0 = none, 5 = strongest)")
result = gr.Markdown()
with gr.Row():
crop1_out = gr.Image(label="Face 1 as the model sees it", height=220)
crop2_out = gr.Image(label="Face 2 with the condition", height=220)
sweep_out = gr.Plot(label="Score at every strength level (above the dashed line = same person)")
inputs = [state1, state2, model_dd, cond_dd, level_sl]
outputs = [crop1_out, crop2_out, result, sweep_out]
img1.change(on_upload, img1, [state1, warn1]).then(run, inputs, outputs)
img2.change(on_upload, img2, [state2, warn2]).then(run, inputs, outputs)
for ctrl in (model_dd, cond_dd, level_sl):
ctrl.change(run, inputs, outputs)
return demo
if __name__ == "__main__":
build().launch()