app.py

import gradio as gr
import torch
import timm
import torch.nn as nn
from torchvision import transforms
from PIL import Image
from huggingface_hub import hf_hub_download

# ---------------------------
# Device configuration
# ---------------------------

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# ---------------------------
# Model definition
# ---------------------------

def build_model():
    """Re-create the network architecture used during training."""
    backbone = timm.create_model(
        "convnext_small", pretrained=False, num_classes=0, global_pool="avg"
    )
    model = nn.Sequential(
        backbone,
        nn.Linear(backbone.num_features, 128),
        nn.ReLU(),
        nn.Dropout(0.4),
        nn.Linear(128, 1),
    )
    return model

# ---------------------------
# Load pretrained weights
# ---------------------------
model_path = hf_hub_download(
    repo_id="lmedz/ips-model-weights",  # your Model Hub repo
    filename="model.pth",
)

model = build_model()
state_dict = torch.load(model_path, map_location=device)
model.load_state_dict(state_dict)
model.to(device)
model.eval()

# ---------------------------
# Pre-processing pipeline
# ---------------------------
transform = transforms.Compose(
    [
        transforms.Resize((224, 224)),
        transforms.ToTensor(),
        transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
    ]
)

# ---------------------------
# Inference function & temperature setting
# ---------------------------

TEMPERATURE = 3.5  # fixed temperature (between 3 and 4) for confidence calibration


def predict(img: Image.Image, magnification: int, ra_conc: float):
    """Return probabilities for High/Low CPM classes.

    Args:
        img: Microscopy image.
        magnification: Tag for objective magnification (×4/10/20).
        ra_conc: Tag for RA concentration (µM).
    """
    img_tensor = transform(img).unsqueeze(0).to(device)
    with torch.no_grad():
        logit = model(img_tensor)
        # Apply fixed temperature scaling to mitigate over‑confidence
        logit_scaled = logit / TEMPERATURE
        prob_high = torch.sigmoid(logit_scaled).item()

    prob_low = 1.0 - prob_high

    return {
        "High CPM Score": prob_high,
        "Low CPM Score": prob_low,
    }

# ---------------------------
# Gradio interface
# ---------------------------

demo = gr.Interface(
    fn=predict,
    inputs=[
        gr.Image(type="pil", label="Microscopy Image"),
        gr.Dropdown(choices=[4, 10, 20], value=10, label="Magnification (×)"),
        gr.Dropdown(choices=[0.1, 0.5, 1.0], value=0.1, label="RA Concentration (µM)"),
    ],
    outputs=gr.Label(num_top_classes=2, label="Predicted CPM Class & Probability"),
    title="iPS Cell Quality Classifier",
    description=(
        "Upload a microscopy image, choose magnification & RA concentration "
        "(metadata only). Probabilities have been temperature‑scaled for more "
        "realistic confidence estimates."
    ),
)

# ---------------------------
# Launch locally or on Spaces
# ---------------------------
if __name__ == "__main__":
    demo.launch()