Dockerfile

# ============================================================
# DetectifAI Backend — Hugging Face Spaces (Docker SDK, CPU)
# ============================================================
FROM python:3.11-slim

# ---- Non-interactive, UTF-8 ----
ENV DEBIAN_FRONTEND=noninteractive \
    PYTHONUNBUFFERED=1 \
    PIP_NO_CACHE_DIR=1 \
    PORT=7860

WORKDIR /app

# ---- System deps (OpenCV, WeasyPrint, ffmpeg) ----
RUN apt-get update && apt-get install -y --no-install-recommends \
    build-essential \
    libgl1 libglib2.0-0 libsm6 libxext6 libxrender-dev \
    libpango-1.0-0 libpangocairo-1.0-0 libgdk-pixbuf-xlib-2.0-0 \
    libffi-dev shared-mime-info \
    ffmpeg \
    git \
    && rm -rf /var/lib/apt/lists/*

# ---- Install PyTorch CPU-only first (saves ~1 GB vs CUDA) ----
RUN pip install torch torchvision --index-url https://download.pytorch.org/whl/cpu

# ---- Python deps (torch excluded — installed above as CPU-only) ----
COPY requirements-docker.txt .
# IMPORTANT: --extra-index-url ensures any transitive torch deps resolve to CPU, not CUDA
RUN pip install --no-cache-dir -r requirements-docker.txt \
    --extra-index-url https://download.pytorch.org/whl/cpu

# ---- Copy application code ----
# Cache-bust: forces COPY and all downstream layers to rebuild
ARG DETECTIFAI_BUILD_TS=0
# Core application files
COPY app.py config.py main_pipeline.py database_video_service.py \
     object_detection.py behavior_analysis_integrator.py \
     video_captioning_integrator.py event_aggregation.py \
     video_segmentation.py highlight_reel.py video_compression.py \
     json_reports.py detectifai_events.py facial_recognition.py \
     stripe_service.py subscription_middleware.py subscription_routes.py \
     alert_routes.py real_time_alerts.py event_clip_generator.py \
     extract_upload_keyframes.py live_stream_processor.py \
     start_detectifai.py ./

# Sub-packages
COPY core/ core/
COPY database/ database/
COPY report_generation/ report_generation/
COPY video_captioning/ video_captioning/
COPY behavior_analysis/ behavior_analysis/
COPY nlp_search/ nlp_search/
COPY DetectifAI_db/ DetectifAI_db/

# Small model files (<50 MB each) — ship in image
COPY models/fire_YOLO11.pt models/fire_YOLO11.pt
COPY models/weapon_YOLO11.pt models/weapon_YOLO11.pt
COPY models/merged_fire_knife_gun.pt models/merged_fire_knife_gun.pt
COPY ["models/best (2).pt", "models/best (2).pt"]
COPY models/classifier_svm.pkl models/classifier_svm.pkl
COPY models/label_encoder.pkl models/label_encoder.pkl
COPY models/metadata.json models/metadata.json

# Copy the top-level model/ directory (FAISS/SVM face index)
COPY model/ /app/model/

# ---- Pre-create writable directories ----
RUN mkdir -p /app/uploads /app/video_processing_outputs /app/logs \
    /app/temp_faces /app/report_generation/models \
    && chmod -R 777 /app/uploads /app/video_processing_outputs /app/logs /app/temp_faces

# ---- Download large models at build time (cached in Docker layer) ----
# fight_detection.pt & accident_detection.pt (~127 MB each)
# Qwen2.5-3B GGUF (~2 GB)
# This runs once during build; layer is cached on HF Spaces.
RUN python -c "\
from huggingface_hub import hf_hub_download; \
print('Downloading fight_detection.pt...'); \
hf_hub_download('blacksinisterx/detectifai-models', 'fight_detection.pt', local_dir='/app/behavior_analysis', local_dir_use_symlinks=False); \
print('Downloading accident_detection.pt...'); \
hf_hub_download('blacksinisterx/detectifai-models', 'accident_detection.pt', local_dir='/app/behavior_analysis', local_dir_use_symlinks=False); \
print('Done with behavior models.'); \
" || echo "WARNING: Could not download behavior models — will retry at startup"

# NOTE: llama-cpp-python and Qwen GGUF skipped — too slow to build/download on free tier.
# Report generation will use template-based fallback.

EXPOSE 7860

# ---- Start Flask ----
CMD ["python", "app.py"]