app.py

import streamlit as st
import tensorflow as tf
import numpy as np
from PIL import Image
import os

# === Fix font/matplotlib warnings for Hugging Face ===
os.environ["MPLCONFIGDIR"] = "/tmp/matplotlib"
os.environ["XDG_CACHE_HOME"] = "/tmp"

# === Custom loss and metrics ===
def weighted_dice_loss(y_true, y_pred):
    smooth = 1e-6
    y_true_f = tf.reshape(y_true, [-1])
    y_pred_f = tf.reshape(y_pred, [-1])
    intersection = tf.reduce_sum(y_true_f * y_pred_f)
    return 1 - ((2. * intersection + smooth) /
                (tf.reduce_sum(y_true_f) + tf.reduce_sum(y_pred_f) + smooth))

def iou_metric(y_true, y_pred):
    y_true = tf.cast(y_true > 0.5, tf.float32)
    y_pred = tf.cast(y_pred > 0.5, tf.float32)
    intersection = tf.reduce_sum(y_true * y_pred)
    union = tf.reduce_sum(y_true) + tf.reduce_sum(y_pred) - intersection
    return intersection / (union + 1e-6)

def bce_loss(y_true, y_pred):
    return tf.keras.losses.binary_crossentropy(y_true, y_pred)

# === Load model ===
model_path = "final_model_after_third_iteration_WDL0.07_0.5155/"
@st.cache_resource
def load_model():
    return tf.keras.models.load_model(
        model_path,
        custom_objects={
            "weighted_dice_loss": weighted_dice_loss,
            "iou_metric": iou_metric,
            "bce_loss": bce_loss
        }
    )

model = load_model()

# === Title ===
st.title("🕳️ Sinkhole Segmentation with EffV2-UNet")

# === File uploader ===
uploaded_image = st.file_uploader("Upload an image", type=["png", "jpg", "jpeg", "tif", "tiff"])

# === Example selector ===
example_dir = "examples"
example_files = sorted([
    f for f in os.listdir(example_dir)
    if f.lower().endswith((".jpg", ".jpeg", ".png", ".tif", ".tiff"))
])

if example_files:
    st.subheader("🖼️ Try with an Example Image")
    cols = st.columns(min(len(example_files), 4))  # up to 4 per row

    for i, file in enumerate(example_files):
        img_path = os.path.join(example_dir, file)
        example_img = Image.open(img_path)
        with cols[i % len(cols)]:
            if st.button(file, key=file):
                uploaded_image = img_path  # simulate upload
                image = example_img.convert("RGB")
                st.image(image, caption=f"Example: {file}", use_column_width=True)

# === Prediction ===
if uploaded_image:
    if isinstance(uploaded_image, str):
        image = Image.open(uploaded_image).convert("RGB")
    else:
        image = Image.open(uploaded_image).convert("RGB")

    st.image(image, caption="Input Image", use_column_width=True)

    # Confidence threshold slider
    threshold = st.slider("Confidence Threshold", 0.0, 1.0, 0.5, step=0.01)

    # Preprocess and predict
    resized = image.resize((512, 512))
    x = np.expand_dims(np.array(resized), axis=0)
    y = model.predict(x)[0, :, :, 0]

    st.text(f"Prediction min/max: {y.min():.5f} / {y.max():.5f}")

    # Apply thresholding if needed
    mask_bin = (y > threshold).astype(np.uint8) * 255
    mask_image = Image.fromarray(mask_bin)

    st.image(mask_image, caption=f"Segmentation Mask (Threshold = {threshold:.2f})", use_column_width=True)