app.py

#!/usr/bin/env python3
"""
app.py — Gray Leaf Spot Colony Segmentation Pipeline
"""

import csv, json, math, os, re, logging, tempfile, zipfile, io
import datetime as dt
from pathlib import Path

import cv2
import gradio as gr
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import torch
import torch.nn as nn
import torch.nn.functional as F
from PIL import Image, ExifTags
from huggingface_hub import hf_hub_download

logging.basicConfig(level=logging.INFO, format="%(asctime)s %(message)s", datefmt="%H:%M:%S")
log = logging.getLogger("app")

IMAGE_EXTS = {".jpg", ".jpeg", ".png", ".tif", ".tiff", ".bmp", ".webp"}
THUMB_SIZE = (160, 160)
DATE_RE = re.compile(r"(20\d{2})(0[1-9]|1[0-2])(0[1-9]|[12]\d|3[01])")
MAX_IMAGES = 50
MODEL_REPO = "rotsl/grayleafspot-segmentation"
MODEL_FILE = "best_area_w_0.7.pt"
DISH_MM = 90.0
MODEL_SZ = 256
HF_TOKEN = os.environ.get("HF_TOKEN")
CSS = ".gallery-wrap{max-height:65vh;overflow-y:auto} .footer-text{text-align:center;margin-top:8px}"

# ── SmallUNet — exact architecture from model_small_unet.py ──

class ConvBlock(nn.Module):
    def __init__(self, in_channels: int, out_channels: int) -> None:
        super().__init__()
        self.block = nn.Sequential(
            nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1, bias=False),
            nn.ReLU(inplace=True),
            nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1, bias=False),
            nn.ReLU(inplace=True),
        )

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return self.block(x)


class DownBlock(nn.Module):
    def __init__(self, in_channels: int, out_channels: int) -> None:
        super().__init__()
        self.pool = nn.MaxPool2d(kernel_size=2, stride=2)
        self.conv = ConvBlock(in_channels, out_channels)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return self.conv(self.pool(x))


class UpBlock(nn.Module):
    def __init__(self, in_channels: int, out_channels: int) -> None:
        super().__init__()
        self.up = nn.Upsample(scale_factor=2, mode="bilinear", align_corners=False)
        self.conv = ConvBlock(in_channels, out_channels)

    def forward(self, x: torch.Tensor, skip: torch.Tensor) -> torch.Tensor:
        x = self.up(x)
        if x.shape[-2:] != skip.shape[-2:]:
            x = F.interpolate(x, size=skip.shape[-2:], mode="bilinear", align_corners=False)
        x = torch.cat([skip, x], dim=1)
        return self.conv(x)


class SmallUNet(nn.Module):
    def __init__(self, in_channels=3, out_channels=1, base_channels=16):
        super().__init__()
        c1 = base_channels
        c2 = base_channels * 2
        c3 = base_channels * 4
        c4 = base_channels * 8
        bottleneck = base_channels * 16
        self.enc1 = ConvBlock(in_channels, c1)
        self.enc2 = DownBlock(c1, c2)
        self.enc3 = DownBlock(c2, c3)
        self.enc4 = DownBlock(c3, c4)
        self.bottleneck = DownBlock(c4, bottleneck)
        self.up4 = UpBlock(bottleneck + c4, c4)
        self.up3 = UpBlock(c4 + c3, c3)
        self.up2 = UpBlock(c3 + c2, c2)
        self.up1 = UpBlock(c2 + c1, c1)
        self.head = nn.Conv2d(c1, out_channels, kernel_size=1)
        self.activation = nn.Sigmoid()

    def forward(self, x):
        s1 = self.enc1(x)
        s2 = self.enc2(s1)
        s3 = self.enc3(s2)
        s4 = self.enc4(s3)
        b = self.bottleneck(s4)
        x = self.up4(b, s4)
        x = self.up3(x, s3)
        x = self.up2(x, s2)
        x = self.up1(x, s1)
        x = self.head(x)
        return self.activation(x)

# ── Model loading ──
_model = None
def load_model():
    global _model
    if _model is None:
        p = hf_hub_download(MODEL_REPO, MODEL_FILE, token=HF_TOKEN)
        _model = SmallUNet(in_channels=3, out_channels=1, base_channels=16)
        ckpt = torch.load(p, map_location="cpu", weights_only=False)
        sd = ckpt["model_state_dict"] if isinstance(ckpt, dict) and "model_state_dict" in ckpt else ckpt
        _model.load_state_dict(sd, strict=True); _model.eval()
        log.info("Model loaded: SmallUNet (%s)", MODEL_FILE)
    return _model

# ── Core inference ──
def infer_image(img_pil, threshold):
    model = load_model()
    img_arr = np.array(img_pil.convert("RGB"))
    img_resized = cv2.resize(img_arr, (MODEL_SZ, MODEL_SZ))
    x = torch.from_numpy(img_resized.transpose(2, 0, 1)).float() / 255.0
    x = x.unsqueeze(0)
    with torch.no_grad():
        prob = model(x)[0, 0].detach().cpu().numpy()
    log.info("  output range: [%.4f, %.4f] mean=%.4f  >0.5:%d >0.3:%d >0.1:%d",
             prob.min(), prob.max(), prob.mean(),
             (prob > 0.5).sum(), (prob > 0.3).sum(), (prob > 0.1).sum())
    mask = (prob > threshold).astype(np.uint8) * 255
    mask = cv2.resize(mask, (img_pil.width, img_pil.height), interpolation=cv2.INTER_NEAREST)
    overlay = img_arr.copy()
    overlay[mask > 0] = (overlay[mask > 0] * 0.5 + np.array([255, 0, 0]) * 0.5).astype(np.uint8)
    return Image.fromarray(overlay), Image.fromarray(mask)

# ── Helpers ──
def make_thumbnail(p):
    try: im = Image.open(p); im.thumbnail(THUMB_SIZE, Image.LANCZOS); return im
    except: return Image.new("RGB", THUMB_SIZE, (200, 200, 200))

def detect_image_date(p):
    m = DATE_RE.search(Path(p).stem)
    if m:
        try: return dt.date(int(m[1]), int(m[2]), int(m[3])).isoformat()
        except: pass
    try:
        im = Image.open(p); ex = im.getexif()
        if ex:
            for tid, tn in ExifTags.TAGS.items():
                if tn == "DateTimeOriginal":
                    v = ex.get(tid)
                    if v: return dt.datetime.strptime(v, "%Y:%m:%d %H:%M:%S").date().isoformat()
    except: pass
    try: return dt.date.fromtimestamp(os.path.getmtime(p)).isoformat()
    except: return dt.date.today().isoformat()

def day_code(img_d, exp_d):
    try: d = (dt.date.fromisoformat(img_d) - dt.date.fromisoformat(exp_d)).days + 1; return f"d{max(d,1):02d}"
    except: return "d??"

def write_ics(rems, path):
    L = ["BEGIN:VCALENDAR","VERSION:2.0","PRODID:-//FungalPipeline//EN"]
    for r in rems:
        uid = r["image_path"].replace("/","_")
        ds = r["remind_me"].replace("-","").replace(" ","T").replace(":","") + "00"
        L += ["BEGIN:VEVENT",f"UID:{uid}@fp",f"DTSTART:{ds}",
              f"SUMMARY:Reminder - {r['experiment_name']}: {Path(r['image_path']).name}","END:VEVENT"]
    L.append("END:VCALENDAR")
    with open(path,"w") as f: f.write("\r\n".join(L))

def fig_to_pil(fig):
    buf = io.BytesIO(); fig.savefig(buf, format="png", dpi=120, bbox_inches="tight", facecolor="white")
    buf.seek(0); img = Image.open(buf).copy(); buf.close(); plt.close(fig); return img

# ── Full pipeline helpers (lazy skimage) ──
def _load_skimage():
    from skimage import filters, measure, morphology
    from skimage.filters import frangi, meijering
    from skimage.morphology import skeletonize, disk, opening, closing, erosion, dilation
    return filters, measure, morphology, frangi, meijering, skeletonize, disk, opening, closing, erosion, dilation

def detect_dish(img_bgr):
    try:
        gray = cv2.cvtColor(img_bgr, cv2.COLOR_BGR2GRAY)
        blurred = cv2.GaussianBlur(gray, (9,9), 2); h, w = gray.shape
        mn, mx = int(min(h,w)*0.25), int(min(h,w)*0.52)
        circles = cv2.HoughCircles(blurred, cv2.HOUGH_GRADIENT, dp=1.2, minDist=min(h,w)//2,
                                    param1=100, param2=40, minRadius=mn, maxRadius=mx)
        if circles is None: return None
        circles = np.round(circles[0]).astype(int); ic,jc = w/2, h/2; bi,bs = 0,-1
        for i,(cx,cy,r) in enumerate(circles):
            s = r / (1 + math.hypot(cx-ic, cy-jc)/100)
            if s > bs: bs=s; bi=i
        cx,cy,r = int(circles[bi][0]), int(circles[bi][1]), int(circles[bi][2])
        return cx, cy, r, DISH_MM/(2*r)
    except: return None

def detect_cracks(gray, colony_mask):
    filters,measure,_,_,_,_,disk,opening,_,erosion,_ = _load_skimage()
    if colony_mask.sum() < 100: return np.zeros_like(colony_mask, dtype=bool)
    interior = gray.copy(); interior[~colony_mask] = 0; er = erosion(colony_mask, disk(5))
    iu = (interior*255 if interior.max()<=1 else interior).astype(np.uint8)
    lt = filters.threshold_local(iu, block_size=51, method="gaussian")
    dk = (iu < (lt-15)) & er; dk = opening(dk, disk(1)); lb = measure.label(dk)
    cm = np.zeros_like(dk, dtype=bool)
    for rp in measure.regionprops(lb):
        if rp.area < 10: continue
        if rp.major_axis_length > 0 and rp.minor_axis_length > 0:
            if rp.major_axis_length/rp.minor_axis_length > 2.5 or rp.eccentricity > 0.85:
                cm[lb==rp.label] = True
    return cm

def detect_hyphae(gray, colony_mask):
    _,_,_,frangi,meijering,skeletonize,disk,_,_,_,dilation = _load_skimage()
    if colony_mask.sum() < 100:
        z = np.zeros_like(colony_mask, dtype=bool); return z, z.copy(), z.copy()
    g = gray.astype(np.float64); ex = dilation(colony_mask, disk(20))
    fr = frangi(g, sigmas=range(1,5), black_ridges=False); fr[~ex]=0
    th = fr[ex].mean()+2*fr[ex].std() if ex.sum()>0 else .01; fs = skeletonize(fr>th)
    mr = meijering(g, sigmas=range(1,5), black_ridges=False); mr[~ex]=0
    th2 = mr[ex].mean()+2*mr[ex].std() if ex.sum()>0 else .01; ms = skeletonize(mr>th2)
    return fs, ms, fs|ms

def compute_metrics(mask_bool, gray, px2mm, dcx, dcy, crack_mask, hyph_f, hyph_m, hyph_h):
    filters,measure,morphology,_,_,_,_,_,_,_,_ = _load_skimage()
    mm2 = px2mm**2
    if mask_bool.sum() < 50:
        return {k:0 for k in ["area_mm2","diameter_mm","perimeter_mm","eccentricity","edge_roughness",
                               "centre_delta_mm","entropy","texture_std","crack_px","crack_area_mm2",
                               "crack_coverage_pct","crack_count","hyph_frangi_mm","hyph_meijering_mm","hyph_hybrid_mm"]}
    pr = measure.regionprops(mask_bool.astype(np.uint8))[0]; R = {}
    R["area_mm2"]=round(pr.area*mm2,4); pm=measure.perimeter(mask_bool)
    R["perimeter_mm"]=round(pm*px2mm,4); R["diameter_mm"]=round(pr.equivalent_diameter_area*px2mm,4)
    R["eccentricity"]=round(pr.eccentricity,6); eq=math.pi*pr.equivalent_diameter_area
    R["edge_roughness"]=round(pm/eq,6) if eq>0 else 0; cy,cx=pr.centroid
    R["centre_delta_mm"]=round(math.hypot(cx-dcx,cy-dcy)*px2mm,4)
    gu8=(gray*255).astype(np.uint8) if gray.max()<=1 else gray.astype(np.uint8)
    R["entropy"]=round(float(filters.rank.entropy(gu8,morphology.disk(5),mask=mask_bool)[mask_bool].mean()),6) if pr.area>100 else 0
    R["texture_std"]=round(float(gray[mask_bool].std()),6)
    R["crack_px"]=int(crack_mask.sum()); R["crack_area_mm2"]=round(crack_mask.sum()*mm2,6)
    R["crack_coverage_pct"]=round(100*crack_mask.sum()/pr.area,4) if pr.area>0 else 0
    R["crack_count"]=int(measure.label(crack_mask).max())
    R["hyph_frangi_mm"]=round(int(hyph_f.sum())*px2mm,4)
    R["hyph_meijering_mm"]=round(int(hyph_m.sum())*px2mm,4)
    R["hyph_hybrid_mm"]=round(int(hyph_h.sum())*px2mm,4)
    return R

def create_full_overlays(img_bgr, colony_mask, crack_mask, hyph_hybrid, dish_info, fname):
    img_rgb = cv2.cvtColor(img_bgr, cv2.COLOR_BGR2RGB); h,w = img_bgr.shape[:2]
    dcx,dcy,dr = (dish_info[0],dish_info[1],dish_info[2]) if dish_info else (w//2,h//2,min(h,w)//2)
    p1=img_rgb.copy()
    if dish_info: cv2.circle(p1,(dcx,dcy),dr,(0,255,0),3)
    cts,_ = cv2.findContours(colony_mask.astype(np.uint8), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
    cv2.drawContours(p1,cts,-1,(255,0,0),2)
    p2=np.zeros_like(img_rgb); p2[colony_mask]=[255,255,255]
    p3=img_rgb.copy()
    if colony_mask.sum()>0: p3[colony_mask]=(p3[colony_mask].astype(np.float32)*0.5+np.array([255,0,0],dtype=np.float32)*0.5).astype(np.uint8)
    if dish_info: cv2.circle(p3,(dcx,dcy),dr,(0,255,0),2)
    p4=img_rgb.copy()
    if crack_mask.sum()>0:
        ck=cv2.dilate(crack_mask.astype(np.uint8),cv2.getStructuringElement(cv2.MORPH_ELLIPSE,(3,3)))>0
        p4[ck]=(p4[ck].astype(np.float32)*0.3+np.array([255,255,0],dtype=np.float32)*0.7).astype(np.uint8)
    if dish_info: cv2.circle(p4,(dcx,dcy),dr,(0,255,0),2); cv2.drawContours(p4,cts,-1,(255,0,0),1)
    p5=img_rgb.copy()
    if hyph_hybrid.sum()>0:
        hy=cv2.dilate(hyph_hybrid.astype(np.uint8),cv2.getStructuringElement(cv2.MORPH_ELLIPSE,(3,3)))>0
        p5[hy]=(p5[hy].astype(np.float32)*0.3+np.array([0,255,255],dtype=np.float32)*0.7).astype(np.uint8)
    if dish_info: cv2.circle(p5,(dcx,dcy),dr,(0,255,0),2); cv2.drawContours(p5,cts,-1,(255,0,0),1)
    p6=img_rgb.copy()
    if colony_mask.sum()>0: p6[colony_mask]=(p6[colony_mask].astype(np.float32)*0.6+np.array([255,0,0],dtype=np.float32)*0.4).astype(np.uint8)
    if crack_mask.sum()>0:
        ck2=cv2.dilate(crack_mask.astype(np.uint8),cv2.getStructuringElement(cv2.MORPH_ELLIPSE,(3,3)))>0; p6[ck2]=[255,255,0]
    if hyph_hybrid.sum()>0:
        hy2=cv2.dilate(hyph_hybrid.astype(np.uint8),cv2.getStructuringElement(cv2.MORPH_ELLIPSE,(3,3)))>0; p6[hy2]=[0,255,255]
    if dish_info: cv2.circle(p6,(dcx,dcy),dr,(0,255,0),2)
    return [(Image.fromarray(p1),f"{fname} — Raw+Dish"),(Image.fromarray(p2),f"{fname} — Mask"),
            (Image.fromarray(p3),f"{fname} — Colony"),(Image.fromarray(p4),f"{fname} — Cracks"),
            (Image.fromarray(p5),f"{fname} — Hyphae"),(Image.fromarray(p6),f"{fname} — Combined")]

def make_growth_charts(results):
    """Generate time-series charts for every morphometric parameter.

    All spatial metrics are already in mm (or mm²) via the per-image
    px_to_mm calibration computed from dish detection, so images of
    different resolutions are correctly comparable.
    """
    if len(results) < 2:
        return []
    df = pd.DataFrame(results)
    if "error" in df.columns:
        df = df[df["error"].fillna("").astype(str).str.strip() == ""].copy()
    if len(df) < 2:
        return []

    # Coerce every plottable column to numeric
    numeric_cols = [
        "days_since_start", "area_mm2", "diameter_mm", "perimeter_mm",
        "eccentricity", "edge_roughness", "centre_delta_mm",
        "entropy", "texture_std",
        "crack_area_mm2", "crack_coverage_pct", "crack_count",
        "hyph_frangi_mm", "hyph_meijering_mm", "hyph_hybrid_mm",
        "rgr_per_day", "relative_growth_per_day",
    ]
    for c in numeric_cols:
        if c in df.columns:
            df[c] = pd.to_numeric(df[c], errors="coerce")

    df = df.sort_values("days_since_start").reset_index(drop=True)
    charts = []

    # ── Chart definitions ──────────────────────────────────────────────
    # (column, y-label, title, colour, marker, fill_under_curve)
    chart_defs = [
        # Colony geometry
        ("area_mm2",                "Area (mm²)",              "Colony Area",                 "#e74c3c", "o", True),
        ("diameter_mm",             "Diameter (mm)",           "Colony Diameter",             "#2980b9", "s", False),
        ("perimeter_mm",            "Perimeter (mm)",          "Colony Perimeter",            "#8e44ad", "^", False),
        # Shape descriptors
        ("eccentricity",            "Eccentricity",            "Colony Eccentricity",         "#e67e22", "D", False),
        ("edge_roughness",          "Edge Roughness",          "Edge Roughness (P / πd)",     "#16a085", "v", False),
        ("centre_delta_mm",         "Centre Offset (mm)",      "Colony Centre Offset",        "#d35400", "p", False),
        # Texture
        ("entropy",                 "Entropy",                 "Colony Texture Entropy",      "#7f8c8d", "h", False),
        ("texture_std",             "Texture Std Dev",         "Colony Texture Std Dev",      "#2c3e50", "*", False),
        # Cracks
        ("crack_area_mm2",          "Crack Area (mm²)",        "Crack Area",                  "#f1c40f", "o", True),
        ("crack_coverage_pct",      "Crack Coverage (%)",      "Crack Coverage",              "#d4ac0d", "s", False),
        ("crack_count",             "Crack Count",             "Number of Cracks",            "#b7950b", "^", False),
        # Hyphae
        ("hyph_frangi_mm",          "Length (mm)",             "Hyphae Length — Frangi",      "#1abc9c", "o", False),
        ("hyph_meijering_mm",       "Length (mm)",             "Hyphae Length — Meijering",   "#3498db", "s", False),
        ("hyph_hybrid_mm",          "Length (mm)",             "Hyphae Length — Hybrid",      "#2ecc71", "D", False),
        # Growth rates (only present from image 2 onward)
        ("rgr_per_day",             "RGR (ln mm² / day)",      "Relative Growth Rate",        "#c0392b", "o", False),
        ("relative_growth_per_day", "Growth (mm² / day)",      "Absolute Growth Rate",        "#e74c3c", "s", False),
    ]

    for col, ylabel, title, color, marker, fill in chart_defs:
        if col not in df.columns:
            continue
        valid = df[col].notna()
        # Also drop rows where the value was left as empty string
        valid = valid & (df[col].astype(str).str.strip() != "")
        if valid.sum() < 2:
            continue
        sub = df.loc[valid].copy()
        fig, ax = plt.subplots(figsize=(8, 4))
        ax.plot(sub["days_since_start"], sub[col], f"{marker}-",
                color=color, lw=2, ms=8)
        if fill:
            ax.fill_between(sub["days_since_start"], 0, sub[col],
                            alpha=0.15, color=color)
        ax.set(xlabel="Days", ylabel=ylabel, title=title)
        ax.grid(True, alpha=0.3)
        charts.append((fig_to_pil(fig), title))

    return charts

# ═══════════════════════════════════════════════════════════════════════════
# Gradio UI
# ═══════════════════════════════════════════════════════════════════════════
with gr.Blocks(title="Gray Leaf Spot Segmentation", css=CSS) as demo:
    paths_st=gr.State([]); dates_st=gr.State({}); rems_st=gr.State({}); cur_idx=gr.State(-1); results_st=gr.State([])

    gr.Markdown("# 🔬 Gray Leaf Spot Colony Segmentation\n"
        "Upload → **Run Inference** → instant results  |  Toggle *Full Pipeline* for morphometrics\n\n"
        "Model: [`rotsl/grayleafspot-segmentation/best_area_w_0.7.pt`]"
        "(https://huggingface.co/rotsl/grayleafspot-segmentation) · SmallUNet (area-consistency w=0.7)")

    with gr.Accordion("📂 Step 1 — Upload Images", open=True):
        upload = gr.File(label="Drag & drop petri dish images", file_count="multiple",
                         file_types=["image"])
        up_st = gr.Markdown("")

    with gr.Accordion("⚙️ Step 2 — Settings", open=True):
        with gr.Row():
            threshold_slider = gr.Slider(label="Mask confidence threshold", minimum=0.0, maximum=1.0, value=0.5, step=0.01)
            full_pipeline_cb = gr.Checkbox(label="Full Pipeline (slower: dish, cracks, hyphae, morphometrics)", value=False)
        with gr.Row():
            exp_name = gr.Textbox(label="Experiment Name", placeholder="MagExp01")
            exp_date = gr.Textbox(label="Experiment Date", placeholder="2025-04-01")
            user_name = gr.Textbox(label="User Name", placeholder="Your name")
            plates_count = gr.Number(label="Plates", value=1, minimum=1, maximum=200, precision=0)

    with gr.Accordion("🖼️ Step 3 — Review & Edit Dates", open=False):
        gr.Markdown("*Click thumbnail → edit date → Save*")
        with gr.Row():
            with gr.Column(scale=2):
                gallery = gr.Gallery(label="Images", columns=4, height=400, object_fit="contain", allow_preview=False, interactive=False)
            with gr.Column(scale=1):
                sel_img=gr.Image(label="Selected",height=200,interactive=False)
                sel_fn=gr.Textbox(label="Filename",interactive=False)
                sel_dt=gr.Textbox(label="Image Date",interactive=True)
                sel_dc=gr.Textbox(label="Day Code",interactive=False)
                sel_rm=gr.Textbox(label="Remind Me",placeholder="YYYY-MM-DD HH:MM",interactive=True)
                sv_btn=gr.Button("💾 Save Date",variant="primary"); sv_st=gr.Markdown("")

    with gr.Accordion("📥 Step 4 — Export Metadata", open=False):
        exp_btn=gr.Button("📥 Export CSV / JSON / ICS",variant="primary"); exp_st=gr.Markdown("")
        meta_preview=gr.Dataframe(label="image_metadata.csv",interactive=False,wrap=True)
        meta_dl=gr.File(label="⬇️ Download metadata zip",interactive=False)

    with gr.Accordion("🚀 Step 5 — Run Inference", open=True):
        run_btn=gr.Button("🚀 Run Inference",variant="primary",size="lg"); run_st=gr.Markdown("")
        gr.Markdown("### Results")
        overlay_gallery=gr.Gallery(label="Segmentation results",columns=3,height=500,object_fit="contain",allow_preview=True)
        gr.Markdown("### Growth Charts (full pipeline, ≥2 images)")
        chart_gallery=gr.Gallery(label="Growth curves",columns=3,height=400,object_fit="contain",allow_preview=True)
        gr.Markdown("### Results Table (full pipeline)")
        results_df=gr.Dataframe(label="analysis_full.csv",interactive=False,wrap=True)
        results_dl=gr.File(label="⬇️ Download analysis zip",interactive=False)

    # ── Footer ──
    gr.Markdown("<div class='footer-text'>\n\n---\nDeveloped by [Rohan R](https://rotsl.github.io/)\n</div>")

    # ── Handlers ──
    def on_upload(files):
        if not files: return [],[],{},[],"",-1
        paths=[str(f) for f in files if Path(str(f)).suffix.lower() in IMAGE_EXTS][:MAX_IMAGES]
        if not paths: return [],[],{},[],"",-1
        dates={p:detect_image_date(p) for p in paths}; rems={p:"" for p in paths}
        return paths,dates,rems,[(p,Path(p).name) for p in paths],f"✅ **{len(paths)}** images loaded.",-1
    upload.upload(on_upload,[upload],[paths_st,dates_st,rems_st,gallery,up_st,cur_idx])

    def on_sel(paths,dates,rems,ed,evt:gr.SelectData):
        i=evt.index
        if i<0 or i>=len(paths): return -1,None,"","","",""
        p=paths[i]; return i,make_thumbnail(p),Path(p).name,dates.get(p,""),day_code(dates.get(p,""),ed) if ed else "",rems.get(p,"")
    gallery.select(on_sel,[paths_st,dates_st,rems_st,exp_date],[cur_idx,sel_img,sel_fn,sel_dt,sel_dc,sel_rm])

    def on_save(paths,dates,rems,i,nd,nr,ed):
        if i<0 or i>=len(paths): return dates,rems,"","⚠️ Select image."
        p=paths[i]; dates=dict(dates); rems=dict(rems); dates[p]=nd; rems[p]=nr
        return dates,rems,day_code(nd,ed) if ed else "",f"✅ **{Path(p).name}** → {nd}"
    sv_btn.click(on_save,[paths_st,dates_st,rems_st,cur_idx,sel_dt,sel_rm,exp_date],[dates_st,rems_st,sel_dc,sv_st])

    def on_export(paths,dates,rems,en,ed,un,pc):
        if not paths: return "⚠️ Upload first.",None,None
        tmp=tempfile.mkdtemp(); rows=[]; rl=[]
        for p in paths:
            imd=dates.get(p,detect_image_date(p)); rm=rems.get(p,"")
            row=dict(image_path=Path(p).name,experiment_name=en or"",experiment_date=ed or"",
                     image_date=imd,day_code=day_code(imd,ed) if ed else"",user_name=un or"",
                     plates_count=int(pc) if pc else 1,remind_me=rm)
            rows.append(row)
            if rm.strip(): rl.append({**row})
        cp=Path(tmp)/"image_metadata.csv"
        with open(cp,"w",newline="") as f: w=csv.DictWriter(f,fieldnames=list(rows[0].keys())); w.writeheader(); w.writerows(rows)
        jp=Path(tmp)/"image_metadata.json"
        with open(jp,"w") as f: json.dump(rows,f,indent=2)
        zf=[cp,jp]
        if rl: ip=Path(tmp)/"reminders.ics"; write_ics(rl,str(ip)); zf.append(ip)
        zp=Path(tmp)/"image_metadata.zip"
        with zipfile.ZipFile(zp,"w") as z:
            for f2 in zf: z.write(f2,f2.name)
        return f"✅ Exported **{len(rows)}** images.",pd.DataFrame(rows),str(zp)
    exp_btn.click(on_export,[paths_st,dates_st,rems_st,exp_name,exp_date,user_name,plates_count],[exp_st,meta_preview,meta_dl])

    def on_run(paths,dates,en,ed,un,pc,thresh,full_pipeline,progress=gr.Progress()):
        if not paths: return "⚠️ Upload images first.",[],[],None,None,[]
        try: load_model()
        except Exception as e: return f"❌ Model failed: {e}",[],[],None,None,[]

        results=[]; vis=[]; errors=[]

        if not full_pipeline:
            for p in progress.tqdm(paths, desc="Segmenting"):
                try:
                    img=Image.open(p).convert("RGB")
                    overlay,mask=infer_image(img,thresh)
                    mask_px=np.sum(np.array(mask)>0)
                    vis.append((img,f"{Path(p).name} — Raw"))
                    vis.append((mask,f"{Path(p).name} — Mask"))
                    vis.append((overlay,f"{Path(p).name} — Overlay"))
                    log.info("%s: done (mask_pixels=%d, threshold=%.2f)", Path(p).name, mask_px, thresh)
                except Exception as e:
                    log.error("%s: %s",Path(p).name,e); errors.append(f"{Path(p).name}: {e}")
            em=f"\n\n⚠️ Errors: {'; '.join(errors)}" if errors else ""
            ok=len(paths)-len(errors)
            return f"✅ **{ok}/{len(paths)}** segmented (fast mode, threshold={thresh:.2f}).{em}",vis,[],None,None,[]

        # Full pipeline
        for p in progress.tqdm(paths, desc="Full pipeline"):
            imd=dates.get(p,detect_image_date(p))
            try:
                img_bgr=cv2.imread(str(p))
                if img_bgr is None: raise RuntimeError(f"Cannot read: {p}")
                model=load_model(); img_rgb=cv2.cvtColor(img_bgr,cv2.COLOR_BGR2RGB)
                img_resized=cv2.resize(img_rgb,(MODEL_SZ,MODEL_SZ))
                x=torch.from_numpy(img_resized.transpose(2,0,1)).float()/255.0; x=x.unsqueeze(0)
                with torch.no_grad(): prob=model(x)[0,0].detach().cpu().numpy()
                mask_small=(prob>thresh).astype(np.uint8)*255; h,w=img_bgr.shape[:2]
                colony_mask=cv2.resize(mask_small,(w,h),interpolation=cv2.INTER_NEAREST)>0
                dish_info=detect_dish(img_bgr)
                gray=cv2.cvtColor(img_bgr,cv2.COLOR_BGR2GRAY).astype(np.float64)/255.0
                crack_mask=detect_cracks(gray,colony_mask); hyph_f,hyph_m,hyph_h=detect_hyphae(gray,colony_mask)
                if dish_info: dcx,dcy,dr,px2mm=dish_info
                else: dcx,dcy=w//2,h//2; dr=min(h,w)//2; px2mm=1.0
                metrics=compute_metrics(colony_mask,gray,px2mm,dcx,dcy,crack_mask,hyph_f,hyph_m,hyph_h)
                metrics.update(colony_pixels=int(colony_mask.sum()),dish_detected=dish_info is not None,
                               dish_radius_px=dr,px_to_mm=round(px2mm,6),
                               calibration_diameter_mm=round(2*dr*px2mm,4),
                               calibration_error_pct=round(abs(2*dr*px2mm-90)/90*100,4) if dish_info else 0,
                               image_path=Path(p).name,experiment_name=en or"",experiment_date=ed or"",
                               image_date=imd,day_code=day_code(imd,ed) if ed else"",
                               user_name=un or"",plates_count=int(pc) if pc else 1)
                results.append(metrics)
                panels=create_full_overlays(img_bgr,colony_mask,crack_mask,hyph_h,dish_info,Path(p).name)
                vis.extend(panels); log.info("%s: area=%.1f mm²",Path(p).name,metrics["area_mm2"])
            except Exception as e:
                log.error("%s: %s",Path(p).name,e); errors.append(f"{Path(p).name}: {e}")
                results.append({"image_path":Path(p).name,"error":str(e)})

        ok_results=[r for r in results if not r.get("error")]
        if len(ok_results)>1:
            ok_results.sort(key=lambda r:r.get("image_date",""))
            try: base=dt.date.fromisoformat(ok_results[0].get("image_date",""))
            except: base=None
            for i,r in enumerate(ok_results):
                try: r["days_since_start"]=(dt.date.fromisoformat(r.get("image_date",""))-base).days if base else 0
                except: r["days_since_start"]=0
                if i==0: r["rgr_per_day"]=""; r["relative_growth_per_day"]=""; continue
                prev=ok_results[i-1]
                try:
                    dd=(dt.date.fromisoformat(r["image_date"])-dt.date.fromisoformat(prev["image_date"])).days
                    a2,a1=float(r.get("area_mm2",0)),float(prev.get("area_mm2",0))
                    if dd>0 and a1>0 and a2>0:
                        r["rgr_per_day"]=round((math.log(a2)-math.log(a1))/dd,6)
                        r["relative_growth_per_day"]=round((a2-a1)/dd,4)
                    else: r["rgr_per_day"]=""; r["relative_growth_per_day"]=""
                except: r["rgr_per_day"]=""; r["relative_growth_per_day"]=""
        chart_items=make_growth_charts(ok_results) if len(ok_results)>=2 else []
        tmp=tempfile.mkdtemp(); all_results=ok_results+[r for r in results if r.get("error")]
        cp=Path(tmp)/"analysis_full.csv"
        if all_results:
            ks=list(all_results[0].keys())
            with open(cp,"w",newline="") as f: w=csv.DictWriter(f,fieldnames=ks,extrasaction="ignore"); w.writeheader(); w.writerows(all_results)
        jp=Path(tmp)/"analysis_full.json"
        with open(jp,"w") as f: json.dump(all_results,f,indent=2,default=str)
        for i,(cimg,cap) in enumerate(chart_items): cimg.save(str(Path(tmp)/f"chart_{i}.png"))
        zp=Path(tmp)/"analysis_full.zip"
        with zipfile.ZipFile(zp,"w") as z:
            for fp in Path(tmp).glob("*"):
                if fp.name!="analysis_full.zip": z.write(fp,fp.name)
        em=f"\n\n⚠️ Errors: {'; '.join(errors)}" if errors else ""
        cm=f"\n\n📊 **{len(chart_items)} charts**" if chart_items else ""
        return (f"✅ **{len(ok_results)}/{len(results)}** analyzed.{cm}{em}",
                vis,chart_items,pd.DataFrame(all_results),str(zp),all_results)

    run_btn.click(on_run,[paths_st,dates_st,exp_name,exp_date,user_name,plates_count,threshold_slider,full_pipeline_cb],
                  [run_st,overlay_gallery,chart_gallery,results_df,results_dl,results_st])

if __name__=="__main__":
    demo.launch(server_name="0.0.0.0",server_port=7860)