requirements.txt

# DiffShield v2 requirements.
# STATUS: PINNED to the Kaggle GPU image recorded by J00 (2026-10-14, results/J00/env_probe/pip_freeze.txt;
# image gcr.io/kaggle-gpu-images/python@sha256:2757e0c7d1e0a9cb43da657b97e223c321a98f5014bdf64f44f2f6b083ad2b2f,
# Python 3.13.15, CUDA 12.8, cuDNN 9.19, 2x Tesla T4). J01 ran v1 end to end with exactly these versions.
# Check an environment with: python scripts/check_env.py
#
# --- provided by the Kaggle image (pinned for check_env.py, NOT reinstalled by the job runner) ---
# Kaggle reports torch/torchvision as 2.11.0+cu128 / 0.26.0+cu128; check_env.py ignores the +local part.
torch==2.11.0
torchvision==0.26.0
numpy==2.1.3
pillow==12.3.0
safetensors==0.8.0
scipy==1.16.3
matplotlib==3.10.0
#
# === INSTALL (core: job runner pip-installs everything below this line, up to the next section) ===
diffusers==0.40.0
transformers==5.16.1
accelerate==1.14.0
huggingface_hub==1.29.0
kornia==0.8.3
lpips==0.1.4
kneed==0.8.6
scikit-image==0.25.2
scikit-learn==1.6.1
pandas==2.3.3
pyyaml==6.0.3
tqdm==4.67.3
optuna==5.0.0
# === INSTALL[eval] (only for jobs that set extras: [eval]; face scoring) ===
insightface==0.7.3          # not in the Kaggle image; compiles Cython at install time; buffalo_l downloads on first use (unverified on Kaggle)
onnxruntime-gpu==1.22.0    # not in the Kaggle image. S05: the unpinned latest wheel needs CUDA 13 (libcublasLt.so.13) and fell back to CPU; 1.22.0 is a CUDA 12 build (UNVERIFIED on this image; faces.py calls preload_dlls and falls back to CPU)
opencv-python-headless==4.14.0.94
# === OPTIONAL (never auto-installed) ===
# facenet-pytorch (L_id / second recogniser): pip install --no-deps facenet-pytorch ; onnx2torch (optional L_id backend)