GPU-accelerated phash + fix discovery/takeout hang

GPU: - Switch Dockerfile base to pytorch/pytorch:2.3.1-cuda12.1-cudnn8-runtime - Add gpu_hasher.py: batched 2D DCT on GPU via PyTorch matrix multiply, 256 images/batch, produces imagehash-compatible 64-bit hex hashes, auto-falls back to CPU when CUDA unavailable - Replace per-image phash loop in scanner.py with phasher.hash_files() - docker-compose.yml: add nvidia GPU device reservation Hang fix: - takeout.is_takeout_folder() now caps at 50 directories (was walking entire tree — blocked for minutes on 65k+ file libraries) - Add "Not a Takeout folder" status message so takeout phase is never silent Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-04-05 01:37:28 -04:00
parent 1d46b9945d
commit c110a8e4f9
6 changed files with 222 additions and 20 deletions
--- a/6
+++ b/6
@@ -1,7 +1,9 @@
-FROM python:3.12-slim
+# PyTorch + CUDA 12.1 base — matches Ubuntu 22.04 with NVIDIA driver 525+
+FROM pytorch/pytorch:2.3.1-cuda12.1-cudnn8-runtime

 RUN apt-get update && apt-get install -y \
-    libheif-dev libjpeg-dev libpng-dev libtiff-dev libwebp-dev exiftool \
+    libheif-dev libjpeg-dev libpng-dev libtiff-dev libwebp-dev \
+    libgl1 libglib2.0-0 exiftool ffmpeg \
    && rm -rf /var/lib/apt/lists/*

 WORKDIR /app
--- a/app/gpu_hasher.py
+++ b/app/gpu_hasher.py
@@ -0,0 +1,162 @@
+"""
+GPU-accelerated perceptual hashing via PyTorch + CUDA.
+
+Implements the same pHash algorithm as the `imagehash` library (DCT-II,
+8×8 low-frequency block, 64-bit hash) so hashes produced here are
+directly comparable with any existing imagehash-generated hashes in the DB.
+
+Falls back to CPU if CUDA is not available — no code changes needed.
+"""
+
+import logging
+import math
+from pathlib import Path
+
+import numpy as np
+import torch
+from PIL import Image, UnidentifiedImageError
+
+try:
+    from pillow_heif import register_heif_opener
+    register_heif_opener()
+except ImportError:
+    pass
+
+log = logging.getLogger(__name__)
+
+# Must match imagehash defaults:  hash_size=8, highfreq_factor=4
+HASH_SIZE  = 8
+IMG_SIZE   = HASH_SIZE * 4   # 32
+BATCH_SIZE = 256              # images per GPU batch; lower if VRAM is tight
+
+
+class GpuPhasher:
+    """
+    Batched perceptual hasher.  Uses CUDA when available, CPU otherwise.
+
+    The DCT is implemented as two matrix multiplications:
+        DCT2D(X) = D @ X @ Dᵀ
+    where D is the precomputed orthonormal DCT-II matrix of size IMG_SIZE.
+    This runs entirely on-GPU for the full batch.
+    """
+
+    def __init__(self, batch_size: int = BATCH_SIZE):
+        self.batch_size = batch_size
+        if torch.cuda.is_available():
+            self.device = torch.device("cuda")
+            dev_name = torch.cuda.get_device_name(0)
+            log.info("GpuPhasher: using CUDA device — %s", dev_name)
+        else:
+            self.device = torch.device("cpu")
+            log.info("GpuPhasher: CUDA not available, using CPU")
+
+        # Precompute orthonormal DCT-II matrix (IMG_SIZE × IMG_SIZE)
+        self._dct = self._build_dct_matrix(IMG_SIZE).to(self.device)
+
+    # ── DCT matrix ────────────────────────────────────────────────────────────
+
+    @staticmethod
+    def _build_dct_matrix(n: int) -> torch.Tensor:
+        """Orthonormal DCT-II matrix of size n×n."""
+        k = torch.arange(n, dtype=torch.float32).unsqueeze(1)   # (n, 1)
+        i = torch.arange(n, dtype=torch.float32).unsqueeze(0)   # (1, n)
+        mat = torch.cos(math.pi * k * (2.0 * i + 1.0) / (2.0 * n))  # (n, n)
+        mat[0]  *= 1.0 / math.sqrt(n)
+        mat[1:] *= math.sqrt(2.0 / n)
+        return mat  # (n, n)
+
+    # ── Image loading ─────────────────────────────────────────────────────────
+
+    @staticmethod
+    def _load_image(path: str) -> np.ndarray | None:
+        """Load image → greyscale float32 numpy array of shape (IMG_SIZE, IMG_SIZE)."""
+        try:
+            img = (
+                Image.open(path)
+                .convert("L")
+                .resize((IMG_SIZE, IMG_SIZE), Image.Resampling.LANCZOS)
+            )
+            return np.asarray(img, dtype=np.float32)
+        except (UnidentifiedImageError, OSError, Exception):
+            return None
+
+    # ── Core GPU batch ────────────────────────────────────────────────────────
+
+    def _phash_batch(self, arrays: list[np.ndarray]) -> list[str]:
+        """
+        Compute pHash for a list of (IMG_SIZE, IMG_SIZE) float32 numpy arrays.
+        Returns a list of 16-char hex strings (64-bit hashes).
+        """
+        # Stack into GPU tensor (B, H, W)
+        batch = torch.from_numpy(np.stack(arrays)).to(self.device)  # (B, 32, 32)
+
+        # 2D DCT:  D @ X @ Dᵀ
+        dct2d = self._dct @ batch @ self._dct.T          # (B, 32, 32)
+
+        # Keep only top-left HASH_SIZE × HASH_SIZE block
+        low   = dct2d[:, :HASH_SIZE, :HASH_SIZE]         # (B, 8, 8)
+        flat  = low.reshape(low.shape[0], -1)            # (B, 64)
+
+        # Each bit: is value > row mean?
+        means = flat.mean(dim=1, keepdim=True)
+        bits  = (flat > means).cpu().numpy()             # (B, 64) bool
+
+        # Pack bits → bytes → hex  (matches imagehash's __str__ format)
+        return [np.packbits(b).tobytes().hex() for b in bits]
+
+    # ── Public API ────────────────────────────────────────────────────────────
+
+    def hash_files(
+        self,
+        paths: list[str],
+        progress_cb=None,
+    ) -> dict[str, str]:
+        """
+        Compute pHash for every path in `paths`.
+
+        Returns {path: hex_hash_string}.  Paths that fail to open are omitted.
+        progress_cb(n_done: int) is called after each batch.
+        """
+        results: dict[str, str] = {}
+        done = 0
+
+        for i in range(0, len(paths), self.batch_size):
+            chunk = paths[i : i + self.batch_size]
+
+            arrays: list[np.ndarray] = []
+            valid:  list[str]        = []
+
+            for p in chunk:
+                arr = self._load_image(p)
+                if arr is not None:
+                    arrays.append(arr)
+                    valid.append(p)
+
+            if arrays:
+                try:
+                    hashes = self._phash_batch(arrays)
+                    results.update(zip(valid, hashes))
+                except Exception as exc:
+                    log.warning("GPU batch failed (%s); skipping batch", exc)
+
+            done += len(chunk)
+            if progress_cb:
+                progress_cb(done)
+
+        return results
+
+    @property
+    def using_gpu(self) -> bool:
+        return self.device.type == "cuda"
+
+
+# ── Module-level singleton (created once, reused across scan phases) ──────────
+
+_phasher: GpuPhasher | None = None
+
+
+def get_phasher() -> GpuPhasher:
+    global _phasher
+    if _phasher is None:
+        _phasher = GpuPhasher()
+    return _phasher
--- a/app/scanner.py
+++ b/app/scanner.py
@@ -20,6 +20,7 @@ except ImportError:
    pass

 from takeout import is_takeout_folder, process_takeout
+from gpu_hasher import get_phasher


 PHOTO_EXT = {
@@ -516,10 +517,14 @@ def run_scan(folder_path: str, scan_id: int, mode: str = "incremental"):
            con.commit()

        # ── Phase: takeout pre-processing ─────────────────────────────────
-        scan_state.update(phase="takeout", message="Checking for Google Takeout structure...")
+        # Detection samples ≤50 dirs so it never blocks on large libraries
+        scan_state.update(phase="takeout",
+                          message="Checking for Google Takeout structure (sampling)...")
        if is_takeout_folder(folder_path):
            scan_state["message"] = "Processing Google Takeout sidecars..."
            process_takeout(folder_path, DB_PATH)
+        else:
+            scan_state["message"] = "Not a Takeout folder — skipping"

        if scan_state["cancel_requested"]:
            _mark_scan(cur, scan_id, "cancelled")
@@ -607,8 +612,10 @@ def run_scan(folder_path: str, scan_id: int, mode: str = "incremental"):
        con.commit()

        # ── Phase: phash ──────────────────────────────────────────────────
+        phasher = get_phasher()
+        hw_label = "GPU" if phasher.using_gpu else "CPU"
        scan_state.update(phase="phash", progress=0,
-                          message="Computing perceptual hashes...")
+                          message=f"Computing perceptual hashes ({hw_label})...")

        cur.execute("""
            SELECT id, path FROM files
@@ -621,19 +628,35 @@ def run_scan(folder_path: str, scan_id: int, mode: str = "incremental"):
        photo_rows = cur.fetchall()
        scan_state["total"] = len(photo_rows)

-        for i, row in enumerate(photo_rows):
-            if scan_state["cancel_requested"]:
-                _mark_scan(cur, scan_id, "cancelled")
-                con.commit()
-                scan_state["status"] = "cancelled"
-                return
+        if photo_rows:
+            # Build id lookup so we can write results back efficiently
+            path_to_id = {row["path"]: row["id"] for row in photo_rows}
+            all_paths  = list(path_to_id.keys())

-            scan_state["progress"] = i + 1
-            scan_state["message"] = f"Phash: {Path(row['path']).name}"
-            ph = _phash(row["path"])
-            if ph:
-                cur.execute("UPDATE files SET phash=? WHERE id=?", (ph, row["id"]))
-            if (i + 1) % 200 == 0:
+            def _phash_progress(n_done: int):
+                if scan_state["cancel_requested"]:
+                    return
+                scan_state["progress"] = n_done
+                scan_state["message"]  = (
+                    f"Phash ({hw_label}): {n_done:,} / {len(all_paths):,}"
+                )
+
+            results = phasher.hash_files(all_paths, progress_cb=_phash_progress)
+
+            # Bulk write to DB in chunks of 500
+            items = list(results.items())
+            for chunk_start in range(0, len(items), 500):
+                if scan_state["cancel_requested"]:
+                    _mark_scan(cur, scan_id, "cancelled")
+                    con.commit()
+                    scan_state["status"] = "cancelled"
+                    return
+                for path, ph in items[chunk_start : chunk_start + 500]:
+                    fid = path_to_id.get(path)
+                    if fid and ph:
+                        cur.execute(
+                            "UPDATE files SET phash=? WHERE id=?", (ph, fid)
+                        )
                con.commit()

        con.commit()
--- a/app/takeout.py
+++ b/app/takeout.py
@@ -50,14 +50,19 @@ def is_takeout_folder(folder_path: str) -> bool:
    adjacent media files. If we find at least 5 such pairs, call it Takeout.
    """
    count = 0
+    dirs_checked = 0
+    MAX_DIRS = 50   # sample at most 50 directories — fast on any library size
+
    for root, dirs, files in os.walk(folder_path):
-        # Skip hidden dirs
        dirs[:] = [d for d in dirs if not d.startswith(".")]
+        dirs_checked += 1
+        if dirs_checked > MAX_DIRS:
+            break
+
        file_set = set(files)
        for f in files:
            if not f.endswith(".json"):
                continue
-            # Check if a media file exists that this could be a sidecar for
            base = f[:-5]  # strip .json
            if base in file_set:
                count += 1
--- a/docker-compose.yml
+++ b/docker-compose.yml
@@ -13,5 +13,10 @@ services:
    deploy:
      resources:
        limits:
-          cpus: "2.0"
-          memory: 2G
+          cpus: "4.0"
+          memory: 4G
+        reservations:
+          devices:
+            - driver: nvidia
+              count: 1
+              capabilities: [gpu]
--- a/requirements.txt
+++ b/requirements.txt
@@ -1,3 +1,7 @@
+# torch + torchvision come pre-installed in the pytorch/pytorch base image
+# (torchvision needed for image transforms)
+torchvision==0.18.1
+
 fastapi==0.115.6
 uvicorn==0.32.1
 Pillow==11.0.0
@@ -5,3 +9,4 @@ imagehash==4.3.1
 pillow-heif==0.21.0
 jinja2==3.1.4
 aiofiles==24.1.0
+numpy==1.26.4