perf: aggressive build optimizations - remove albumentations, use plain uvicorn, multi-stage docker build

Dockerfile

@@ -1,28 +1,47 @@
+# ═══════════════════════════════════════════════════════════════════════════
+# Multi-stage build: Smaller final image, faster rebuilds
+# ═══════════════════════════════════════════════════════════════════════════
+
+# ─── Stage 1: Build dependencies ──────────────────────────────────────────
+FROM python:3.10-slim as builder
+
+WORKDIR /build
+
+# Copy and install Python packages
+COPY requirements.txt .
+RUN pip install --no-cache-dir --upgrade pip setuptools wheel && \
+    pip install --no-cache-dir --default-timeout=1000 --prefix=/install -r requirements.txt
+
+# ─── Stage 2: Runtime (minimal) ───────────────────────────────────────────
 FROM python:3.10-slim
 
-# Install system deps needed by OpenCV (minimal set)
+# Install only runtime system deps (minimal OpenCV requirements)
 RUN apt-get update && apt-get install -y --no-install-recommends \
     libglib2.0-0 \
     libsm6 \
     libxext6 \
-    libxrender-dev \
+    libxrender1 \
+    libgomp1 \
     libgl1 \
-    && rm -rf /var/lib/apt/lists/* /tmp/* /var/tmp/*
+    && rm -rf /var/lib/apt/lists/* /tmp/* /var/tmp/* /var/cache/apt/*
 
 WORKDIR /app
 
-# ── Install dependencies FIRST (most stable layer) ──
-COPY requirements.txt .
-RUN pip install --no-cache-dir --upgrade pip setuptools wheel && \
-    pip install --no-cache-dir --default-timeout=1000 -r requirements.txt
+# Copy installed packages from builder stage
+COPY --from=builder /install /usr/local
 
-# ── Copy model files SECOND (slower to change than code) ──
+# Copy model files (cached unless model changes)
 COPY trainedmodels/ ./trainedmodels/
 
-# ── Copy app code LAST (fastest layer, changes most often) ──
+# Copy application code (changes most frequently)
 COPY app.py inference.py ./
 
-# HuggingFace Spaces requires port 7860
+# Environment optimizations
+ENV PYTHONUNBUFFERED=1 \
+    PYTHONDONTWRITEBYTECODE=1 \
+    PIP_NO_CACHE_DIR=1
+
 EXPOSE 7860
 
-CMD ["uvicorn", "app:app", "--host", "0.0.0.0", "--port", "7860"]
+# Use exec form for better signal handling
+CMD ["uvicorn", "app:app", "--host", "0.0.0.0", "--port", "7860", "--workers", "1"]

inference.py

@@ -5,7 +5,6 @@ import torch
 import torch.nn as nn
 import torch.nn.functional as F
 from torchvision import models
-import albumentations as A
 from pathlib import Path
 
 # ─── Architecture (must match training notebook exactly) ───────────────────────
@@ -165,12 +164,27 @@ def apply_jet_colormap(gray_img):
     return (np.stack([r, g, b], axis=-1) * 255).astype(np.uint8)
 
 
-def get_transform(image_size=512):
-    return A.Compose([
-        A.CLAHE(clip_limit=2.0, tile_grid_size=(8, 8), p=1.0),
-        A.CenterCrop(height=350, width=350),
-        A.Resize(height=image_size, width=image_size),
-    ])
+def apply_clahe_cv2(image, clip_limit=2.0, tile_grid_size=(8, 8)):
+    """Apply CLAHE using OpenCV (replaces albumentations)."""
+    clahe = cv2.createCLAHE(clipLimit=clip_limit, tileGridSize=tile_grid_size)
+    return clahe.apply(image)
+
+
+def preprocess_with_cv2(image, image_size=512):
+    """Preprocessing pipeline using pure OpenCV (replaces albumentations)."""
+    # CLAHE
+    image = apply_clahe_cv2(image, clip_limit=2.0, tile_grid_size=(8, 8))
+    
+    # Center crop 350x350
+    h, w = image.shape[:2]
+    start_y = (h - 350) // 2
+    start_x = (w - 350) // 2
+    image = image[start_y:start_y+350, start_x:start_x+350]
+    
+    # Resize to target size
+    image = cv2.resize(image, (image_size, image_size), interpolation=cv2.INTER_LINEAR)
+    
+    return image
 
 
 def preprocess_image(image_path: str, meta: dict) -> torch.Tensor:
@@ -179,8 +193,8 @@ def preprocess_image(image_path: str, meta: dict) -> torch.Tensor:
     if image is None:
         raise ValueError(f"Could not read image: {image_path}")
 
-    transform = get_transform(meta["image_size"])
-    image = transform(image=image)["image"]
+    # Use OpenCV preprocessing instead of albumentations
+    image = preprocess_with_cv2(image, image_size=meta["image_size"])
 
     image = image.astype(np.float32) / 255.0
     image = (image - meta["global_mean"]) / meta["global_std"]

requirements.txt

@@ -4,9 +4,8 @@ torch==2.0.1+cpu
 torchvision==0.15.2+cpu
 
 fastapi==0.111.0
-uvicorn[standard]==0.29.0
+uvicorn==0.29.0
 python-multipart==0.0.9
 opencv-python-headless==4.9.0.80
-albumentations==1.4.2
 numpy==1.26.4
 Pillow==10.3.0