requirements.txt

# ══════════════════════════════════════════════════════════════════════════════
# Blueprint Room Extractor — requirements.txt
# Target: Hugging Face Spaces (NVIDIA T4 / A10G GPU runtime)
#
# Install order matters — torch must come before torchvision/torchaudio and
# before segment-anything / easyocr so they all link against the same libcuda.
#
# To run locally with GPU:
#   pip install -r requirements.txt
#
# To run locally CPU-only (fallback):
#   pip install -r requirements.txt --extra-index-url https://download.pytorch.org/whl/cpu
# ══════════════════════════════════════════════════════════════════════════════

# ── Web framework ─────────────────────────────────────────────────────────────
flask>=3.0.0
werkzeug>=3.0.0

# ── Core numerics ─────────────────────────────────────────────────────────────
numpy>=1.26.0,<3.0
scipy>=1.12.0
scikit-image>=0.22.0
pillow>=10.0.0

# ── OpenCV (CUDA-enabled build via headless + contrib) ────────────────────────
# opencv-python-headless is the standard HF Spaces variant (no GTK/Qt deps).
# The opencv-contrib-python-headless build ships ximgproc (thinning) and
# cuda* modules when the underlying system has libcuda.so.
# On CPU-only machines both packages fall back gracefully; cv2.cuda
# functions will raise RuntimeError which we catch and route to CPU.
opencv-contrib-python-headless>=4.9.0.80

# ── PyTorch — CUDA 12.1 wheel (HF Spaces default CUDA version) ───────────────
# HF Spaces hardware: NVIDIA T4 (sm_75) or A10G (sm_86), CUDA 12.x
torch>=2.3.0
torchvision>=0.18.0
torchaudio>=2.3.0

# ── CuPy — GPU NumPy (matches CUDA 12.x on HF Spaces) ────────────────────────
# Provides cp.asarray / cp.asnumpy for zero-copy GPU↔CPU transfers and
# replaces NumPy hotspots (chroma erase, RLE encoding, gap analysis) on GPU.
cupy-cuda12x>=13.0.0

# ── Segment Anything Model (SAM) ──────────────────────────────────────────────
# Facebook's original SAM — predictor.predict() runs on CUDA when available.
# Checkpoint (vit_h ~2.4 GB) is downloaded at runtime from HF Hub.
git+https://github.com/facebookresearch/segment-anything.git

# ── OCR — EasyOCR with GPU support ───────────────────────────────────────────
# EasyOCR uses its own torch-based CRNN; when torch.cuda.is_available() it
# automatically moves tensors to GPU without any code change.
easyocr>=1.7.1

# ── ONNXRuntime GPU (optional accelerator for EasyOCR detection backbone) ─────
# EasyOCR's detection model can run via onnxruntime-gpu if present.
onnxruntime-gpu>=1.18.0

# ── Hugging Face integration ──────────────────────────────────────────────────
# huggingface_hub: model download (SAM checkpoint)
# spaces: @spaces.GPU decorator for zero-GPU billing on HF Spaces free tier
huggingface_hub>=0.23.0
spaces>=0.29.0

# ── Enchant spell checker (room-label validation) ─────────────────────────────
# Requires system package: apt-get install libenchant-2-dev (add to packages.txt)
pyenchant>=3.2.2

# ── Miscellaneous utilities ───────────────────────────────────────────────────
# imageio: used by skimage internally
imageio>=2.34.0
# tqdm: progress bars inside EasyOCR / SAM
tqdm>=4.66.0
# requests: SAM checkpoint download fallback
requests>=2.31.0
openpyxl