Update app.py

app.py

@@ -1,32 +1,38 @@
-import os, math, json, io
+import os, json, time, threading, io
 import numpy as np
 import torch
 import torch.nn as nn
 import requests
-import imageio.v3 as iio
-import gradio as gr
-from PIL import Image, ImageDraw, ImageFont
+from PIL import Image
 from matplotlib.colors import LinearSegmentedColormap
 from huggingface_hub import hf_hub_download, HfApi, create_repo
 from pysteps.motion.lucaskanade import dense_lucaskanade
 from pysteps.extrapolation.semilagrangian import extrapolate
 from pyproj import Proj
 from datetime import datetime, timedelta
+import gradio as gr
+from fastapi import FastAPI
+from fastapi.staticfiles import StaticFiles
+import uvicorn
 
 HF_TOKEN=os.environ.get("HF_TOKEN","")
 MODEL_REPO="NovatasticRoScript/himawari-nowcast"
 DATASET_REPO="NovatasticRoScript/himawari-live-cache"
 SEQ=9; PRED=18; LK=4; W=H=640; IN_CH=6
-DISPLAY_BBOX=(103.,-3.,139.,35.)
+DISPLAY_BBOX=(103.,-3.,139.,35.)  # lon_min, lat_min, lon_max, lat_max
 PAR_POLY=[(115.,5.),(115.,15.),(120.,21.),(120.,25.),(135.,25.),(135.,5.)]
-NEIGHBOR_COUNTRIES=["Philippines","Taiwan","Vietnam","Malaysia","Indonesia",
-                    "China","Japan","Brunei","Palau"]
 SLIDER="https://slider.cira.colostate.edu"
-ZOOM=3  # tile zoom level for regional crop (0=full disk thumbnail, higher=sharper, more tiles)
+ZOOM=3
+FRAMES_DIR="/tmp/frames"
+UPDATE_INTERVAL_SEC=600  # 10 minutes
+
+os.makedirs(FRAMES_DIR, exist_ok=True)
 
-# --- Himawari full-disk geostationary projection ---
+# =============================================================================
+# Himawari full-disk geostationary projection (for cropping to DISPLAY_BBOX)
+# =============================================================================
 _HIMA_PROJ = Proj(proj='geos', h=35785863.0, lon_0=140.7, a=6378137.0, b=6356752.3, sweep='x')
-_FULL_DISK_EXTENT = 5500000.035308  # meters, half-width/height of full-disk image
+_FULL_DISK_EXTENT = 5500000.035308
 
 def _lonlat_to_frac(lon, lat):
     x, y = _HIMA_PROJ(lon, lat)
@@ -44,7 +50,6 @@ def _bbox_frac_range():
     return min(fxs), max(fxs), min(fys), max(fys)
 
 def _crop_resize_to_region(full_disk_arr):
-    """Fallback: crop the single full-disk zoom-0 image (low-res)."""
     h, w = full_disk_arr.shape
     fx0, fx1, fy0, fy1 = _bbox_frac_range()
     x0, x1 = int(fx0*w), int(fx1*w)
@@ -61,37 +66,9 @@ def _bbox_tile_range(zoom):
     row0,row1 = max(0,int(fy0*n)), min(n-1,int(fy1*n))
     return row0,row1,col0,col1,n
 
-# --- Country borders/coastlines ---
-def _fetch_geojson():
-    try:
-        return requests.get(
-            "https://raw.githubusercontent.com/nvkelso/natural-earth-vector/master/geojson/ne_110m_admin_0_countries.geojson",
-            timeout=15).json()
-    except Exception:
-        return None
-_GEOJSON = _fetch_geojson()
-
-def _rasterize_land_mask(w, h, bbox, names):
-    lon_min,lat_min,lon_max,lat_max = bbox
-    img = Image.new("L",(w,h),0)
-    if not _GEOJSON:
-        return np.zeros((h,w),dtype=np.float32)
-    canvas = ImageDraw.Draw(img)
-    wanted = set(names)
-    for f in _GEOJSON.get("features",[]):
-        if f.get("properties",{}).get("NAME") not in wanted: continue
-        geom = f.get("geometry",{})
-        coords = [geom.get("coordinates",[])[0]] if geom.get("type")=="Polygon" else [p[0] for p in geom.get("coordinates",[])]
-        for ring in coords:
-            pts = [((lon-lon_min)/(lon_max-lon_min)*w,(lat_max-lat)/(lat_max-lat_min)*h) for lon,lat in ring]
-            if len(pts)>=3: canvas.polygon(pts,fill=255)
-    return np.array(img,dtype=np.float32)/255.
-
-_WIDE_MASK = _rasterize_land_mask(W,H,DISPLAY_BBOX,NEIGHBOR_COUNTRIES)
-_COASTLINE = np.zeros_like(_WIDE_MASK,dtype=bool)
-_COASTLINE[:-1,:]|=(_WIDE_MASK[:-1,:]>0.5)!=(_WIDE_MASK[1:,:]>0.5)
-_COASTLINE[:,:-1]|=(_WIDE_MASK[:,:-1]>0.5)!=(_WIDE_MASK[:,1:]>0.5)
-
+# =============================================================================
+# Colormap + model
+# =============================================================================
 ir_colors=[(0.00,(0.05,0.05,0.05)),(0.30,(0.15,0.20,0.35)),(0.50,(0.00,0.65,0.90)),
            (0.65,(0.00,0.75,0.00)),(0.80,(1.00,0.85,0.00)),(0.92,(0.90,0.10,0.00)),
            (1.00,(1.00,1.00,1.00))]
@@ -119,6 +96,11 @@ except Exception as e:
     print(f"Model load failed: {e}")
 model.eval()
 
+zeros_t=torch.zeros(1,1,H,W); coast_t=torch.zeros(1,1,H,W)
+
+# =============================================================================
+# SLIDER fetch (zoom-tile mosaic, falls back to full-disk crop)
+# =============================================================================
 def slider_ts(product="band_13"):
     url=f"{SLIDER}/data/json/himawari/full_disk/{product}/latest_times.json"
     r=requests.get(url,timeout=10)
@@ -154,7 +136,6 @@ def fetch_slider_region(ts, zoom=ZOOM, product="band_13"):
                 arr=np.array(img,dtype=np.float32)/255.
                 tiles[(row,col)]=arr
                 if tile_h is None: tile_h,tile_w=arr.shape
-                print(f"TILE z{zoom} {row:03d}_{col:03d} OK {arr.shape}")
             except Exception as e:
                 print(f"TILE z{zoom} {row:03d}_{col:03d} FAILED: {e}")
     if not tiles:
@@ -202,74 +183,173 @@ def build_live_seq():
         frames.insert(0,frames[0])
     return np.array(frames[-SEQ:]),base
 
-zeros_t=torch.zeros(1,1,H,W); coast_t=torch.zeros(1,1,H,W)
-aoi=[((lon-DISPLAY_BBOX[0])/(DISPLAY_BBOX[2]-DISPLAY_BBOX[0])*W,
-       (DISPLAY_BBOX[3]-lat)/(DISPLAY_BBOX[3]-DISPLAY_BBOX[1])*H) for lon,lat in PAR_POLY]
-try:
-    font=ImageFont.truetype("/usr/share/fonts/truetype/dejavu/DejaVuSansMono-Bold.ttf",14)
-except:
-    font=ImageFont.load_default()
-
-def run_forecast(uploaded=None):
-    if uploaded:
-        try: raw=list(iio.imiter(uploaded))
-        except: raw=[iio.imread(uploaded)]
-        imgs=[np.array(Image.fromarray(f).convert("L").resize((W,H)),dtype=np.float32)/255. for f in raw]
-        while len(imgs)<SEQ: imgs.append(imgs[-1])
-        history=np.array(imgs[-SEQ:]); base=datetime.utcnow(); mode="UPLOAD"
-    else:
-        history,base=build_live_seq(); mode=f"SLIDER LIVE {base:%H:%M}Z"
-
-    v=dense_lucaskanade(history[-LK:],verbose=False)
-    hor=np.array(extrapolate(history[-1],v,timesteps=PRED,outval="min"))
-
-    curr=torch.tensor(history[-1],dtype=torch.float32).unsqueeze(0).unsqueeze(0)
-    last=torch.tensor(history[-2],dtype=torch.float32).unsqueeze(0).unsqueeze(0)
-    preds=[]
-    with torch.no_grad():
-        for i in range(PRED):
-            prior=torch.tensor(hor[i],dtype=torch.float32).unsqueeze(0).unsqueeze(0)
-            ref=model(curr,zeros_t,zeros_t,prior,curr-last,coast_t)
-            preds.append(ref.squeeze().numpy()); last=curr; curr=ref
-
-    def frame(data,cap,fcst=False):
-        rgb=(cmap(np.clip(data,0,1))[...,:3]*255).astype(np.uint8)
-        rgb[_COASTLINE]=[0,255,80]
-        img=Image.fromarray(rgb); d=ImageDraw.Draw(img)
-        d.line(aoi+[aoi[0]],fill=(255,0,0),width=2)
-        d.rectangle([0,0,W,20],fill=(0,0,0))
-        d.text((6,2),cap,fill=(255,255,255) if not fcst else (255,215,0),font=font)
-        return np.array(img)
-
-    frames=[frame(history[i],f"OBS T-{(SEQ-1-i)*10:02d}min | {mode}") for i in range(SEQ)]
-    frames+=[frame(preds[i],f"FCST T+{(i+1)*10:02d}min",True) for i in range(PRED)]
-    out="/tmp/forecast.gif"
-    iio.imwrite(out,frames,plugin="pillow",extension=".gif",loop=0,duration=220)
-
-    if HF_TOKEN and not uploaded:
+def save_frame_png(data, path):
+    rgb = (cmap(np.clip(data,0,1))[...,:3]*255).astype(np.uint8)
+    Image.fromarray(rgb).save(path)
+
+# =============================================================================
+# Background update loop: runs forever, refreshes frames every 10 minutes
+# =============================================================================
+def update_loop():
+    while True:
         try:
-            api=HfApi(token=HF_TOKEN)
-            create_repo(DATASET_REPO,repo_type="dataset",token=HF_TOKEN,exist_ok=True)
-            fname=f"live_{base:%Y%m%d_%H%M}.npy"
-            npy=f"/tmp/{fname}"; np.save(npy,history[-1])
-            api.upload_file(path_or_fileobj=npy,path_in_repo=f"frames/{fname}",
-                            repo_id=DATASET_REPO,repo_type="dataset",token=HF_TOKEN)
+            history, base = build_live_seq()
+            v = dense_lucaskanade(history[-LK:], verbose=False)
+            hor = np.array(extrapolate(history[-1], v, timesteps=PRED, outval="min"))
+
+            curr = torch.tensor(history[-1], dtype=torch.float32).unsqueeze(0).unsqueeze(0)
+            last = torch.tensor(history[-2], dtype=torch.float32).unsqueeze(0).unsqueeze(0)
+            preds=[]
+            with torch.no_grad():
+                for i in range(PRED):
+                    prior = torch.tensor(hor[i], dtype=torch.float32).unsqueeze(0).unsqueeze(0)
+                    ref = model(curr, zeros_t, zeros_t, prior, curr-last, coast_t)
+                    preds.append(ref.squeeze().numpy()); last=curr; curr=ref
+
+            for i in range(SEQ):
+                save_frame_png(history[i], os.path.join(FRAMES_DIR, f"obs_{i}.png"))
+            for i in range(PRED):
+                save_frame_png(preds[i], os.path.join(FRAMES_DIR, f"fcst_{i}.png"))
+
+            manifest = {
+                "ready": True,
+                "base_time": base.strftime("%Y-%m-%dT%H:%M:%SZ"),
+                "obs_count": SEQ, "fcst_count": PRED,
+                "obs_step_min": 10, "fcst_step_min": 10,
+                "version": int(time.time()),
+            }
+            with open(os.path.join(FRAMES_DIR,"manifest.json"),"w") as f:
+                json.dump(manifest, f)
+            print(f"✅ Updated frames @ {base} UTC")
+
+            if HF_TOKEN:
+                try:
+                    api = HfApi(token=HF_TOKEN)
+                    create_repo(DATASET_REPO, repo_type="dataset", token=HF_TOKEN, exist_ok=True)
+                    fname = f"live_{base:%Y%m%d_%H%M}.npy"
+                    npy_path = f"/tmp/{fname}"
+                    np.save(npy_path, history[-1])
+                    api.upload_file(path_or_fileobj=npy_path, path_in_repo=f"frames/{fname}",
+                                     repo_id=DATASET_REPO, repo_type="dataset", token=HF_TOKEN)
+                except Exception as e:
+                    print(f"Dataset push failed: {e}")
+
         except Exception as e:
-            print(f"Dataset push: {e}")
-    return out
-
-with gr.Blocks(theme=gr.themes.Monochrome()) as demo:
-    gr.Markdown("# 🌀 Himawari-9 PySTEPS + Neural Correction Nowcast")
-    gr.Markdown("CIRA SLIDER IR frames → pysteps optical-flow → CNN residual correction. "
-                "3hr forecast, official PAR hexagon, real country borders. "
-                "Each live forecast auto-saves to HF Dataset for Colab self-improvement retraining.")
-    with gr.Tab("Live Forecast"):
-        btn=gr.Button("⚡ Generate Forecast",variant="primary")
-        out=gr.Image(label="Forecast GIF")
-        btn.click(fn=lambda:run_forecast(None),inputs=[],outputs=[out])
-    with gr.Tab("Upload Your Own"):
-        f=gr.File(label="IR image or GIF",type="filepath")
-        btn2=gr.Button("🧠 Predict",variant="primary")
-        out2=gr.Image(label="Forecast GIF")
-        btn2.click(fn=run_forecast,inputs=[f],outputs=[out2])
-demo.launch()
+            print(f"⚠️ Update loop error: {e}")
+
+        time.sleep(UPDATE_INTERVAL_SEC)
+
+if not os.path.exists(os.path.join(FRAMES_DIR,"manifest.json")):
+    with open(os.path.join(FRAMES_DIR,"manifest.json"),"w") as f:
+        json.dump({"ready": False, "version": 0}, f)
+
+threading.Thread(target=update_loop, daemon=True).start()
+
+# =============================================================================
+# Map UI (Leaflet, served as static HTML/JS inside a Gradio Blocks page)
+# =============================================================================
+_bounds_js = json.dumps([[DISPLAY_BBOX[1],DISPLAY_BBOX[0]],[DISPLAY_BBOX[3],DISPLAY_BBOX[2]]])
+_par_js = json.dumps([[lat,lon] for lon,lat in PAR_POLY])
+
+_MAP_HTML_TEMPLATE = r"""
+<div id="liveMap" style="width:100%;height:560px;border-radius:8px;"></div>
+<div style="margin-top:8px;display:flex;align-items:center;gap:10px;font-family:monospace;">
+  <button id="liveBtn" style="padding:6px 14px;border-radius:6px;border:none;background:#e63946;color:white;font-weight:bold;cursor:pointer;">🔴 LIVE</button>
+  <span id="frameLabel" style="min-width:260px;">Loading...</span>
+  <input id="frameSlider" type="range" min="0" max="1" value="0" style="flex:1;">
+</div>
+<link rel="stylesheet" href="https://cdnjs.cloudflare.com/ajax/libs/leaflet/1.9.4/leaflet.css"/>
+<script src="https://cdnjs.cloudflare.com/ajax/libs/leaflet/1.9.4/leaflet.js"></script>
+<script>
+(function(){
+  const BOUNDS = __BOUNDS__;
+  const PAR = __PAR__;
+  const map = L.map('liveMap', {zoomControl:true});
+  map.fitBounds(BOUNDS);
+  L.tileLayer('https://{s}.basemaps.cartocdn.com/dark_all/{z}/{x}/{y}{r}.png', {
+    attribution: '&copy; CARTO &copy; OpenStreetMap contributors',
+    subdomains: 'abcd', maxZoom: 19
+  }).addTo(map);
+  L.rectangle(BOUNDS, {color:'#888', weight:1, dashArray:'4', fill:false}).addTo(map);
+  L.polygon(PAR, {color:'#ff3b30', weight:2, fill:false}).addTo(map);
+
+  let overlay = L.imageOverlay('/frames/obs_0.png', BOUNDS, {opacity:0.78}).addTo(map);
+
+  let manifest = null;
+  let live = true;
+  let lastVersion = -1;
+
+  const slider = document.getElementById('frameSlider');
+  const label = document.getElementById('frameLabel');
+  const liveBtn = document.getElementById('liveBtn');
+
+  function frameName(idx){
+    if (!manifest) return null;
+    if (idx < manifest.obs_count) {
+      const minsAgo = (manifest.obs_count - 1 - idx) * manifest.obs_step_min;
+      return {file: 'obs_' + idx + '.png', label: 'OBS  T-' + minsAgo + 'min'};
+    } else {
+      const fi = idx - manifest.obs_count;
+      const minsFwd = (fi + 1) * manifest.fcst_step_min;
+      return {file: 'fcst_' + fi + '.png', label: 'FCST T+' + minsFwd + 'min'};
+    }
+  }
+
+  function showFrame(idx){
+    const fr = frameName(idx);
+    if (!fr) return;
+    overlay.setUrl('/frames/' + fr.file + '?t=' + Date.now());
+    label.textContent = fr.label + (manifest.base_time ? ('  |  base ' + manifest.base_time) : '');
+    slider.value = idx;
+  }
+
+  function applyManifest(m){
+    manifest = m;
+    const total = m.obs_count + m.fcst_count;
+    slider.max = total - 1;
+    if (live) showFrame(m.obs_count - 1);
+  }
+
+  function poll(){
+    fetch('/frames/manifest.json?t=' + Date.now())
+      .then(r => r.json())
+      .then(m => {
+        if (!m.ready) return;
+        if (m.version !== lastVersion) {
+          lastVersion = m.version;
+          applyManifest(m);
+        }
+      })
+      .catch(()=>{});
+  }
+
+  slider.addEventListener('input', function(){
+    live = false;
+    liveBtn.style.opacity = 0.45;
+    showFrame(parseInt(slider.value));
+  });
+
+  liveBtn.addEventListener('click', function(){
+    live = true;
+    liveBtn.style.opacity = 1;
+    if (manifest) showFrame(manifest.obs_count - 1);
+  });
+
+  poll();
+  setInterval(poll, 20000);
+})();
+</script>
+"""
+
+MAP_HTML = _MAP_HTML_TEMPLATE.replace("__BOUNDS__", _bounds_js).replace("__PAR__", _par_js)
+
+with gr.Blocks(title="Himawari PAR Nowcast") as demo:
+    gr.Markdown("### 🌀 Himawari-9 Live Nowcast — Philippine Area of Responsibility")
+    gr.HTML(MAP_HTML)
+
+fastapi_app = FastAPI()
+fastapi_app.mount("/frames", StaticFiles(directory=FRAMES_DIR), name="frames")
+fastapi_app = gr.mount_gradio_app(fastapi_app, demo, path="/")
+
+if __name__ == "__main__":
+    uvicorn.run(fastapi_app, host="0.0.0.0", port=7860)