Installation

pip install "specsr[hub]"

Extras select what comes with it: hub adds weight download from the Hugging Face Hub, train adds the training stack (wandb, pandas), zcompare adds the alternative redshift estimators used for the independent cross-check, and all is everything.

PyTorch and CUDA

specsr depends on PyTorch but deliberately does not pin a CUDA build. On a GPU machine, install the torch build matching your driver first, from the official index:

pip install torch --index-url https://download.pytorch.org/whl/cu128
pip install "specsr[hub]"

To track main rather than a release:

pip install "git+https://github.com/aryana-haghjoo/specsr.git#egg=specsr[hub]"

Warning

Installing specsr into an environment with a pre-existing CUDA-enabled torch can cause pip to resolve a different set of nvidia-* runtime wheels, producing errors such as undefined symbol: ncclCommWindowRegister at import time. If that happens, reinstall the nvidia-* pins listed in your torch distribution’s metadata:

grep -E "^Requires-Dist: nvidia" \
  $(python -c "import torch,pathlib;print(pathlib.Path(torch.__file__).parent.parent)")/torch-*.dist-info/METADATA

Development install

git clone https://github.com/aryana-haghjoo/specsr
cd specsr
pip install -e ".[dev,train,hub]"
pytest

Environment variables

Variable	Purpose	Default
SPECSR_JADES_ROOT	Root of the raw JADES tree (DR3/, DR4/, …)	unset — required to build datasets
SPECSR_DATA_DIR	Where derived products are written	./data
SPECSR_OUTPUT_DIR	Where figures, predictions and evaluation tables are written	./outputs
SPECSR_CACHE_DIR	Downloaded weights and caches	~/.cache/specsr
SPECSR_CHECKPOINT_DIR	Load weights from a local directory instead of the Hub	unset
SPECSR_CHECKPOINT_REPO	Override the Hub model repo	aryana-haghjoo/specsr
SPECSR_CHECKPOINT_REVISION	Pin a Hub branch, tag or commit	main

Optional: run notifications

Training runs can email you when they start and finish. It is opt-in and uses your own address and mail server – nothing is configured by default and nothing is sent until you ask for it:

./scripts/notify-run --setup

That asks for the address to notify, the SMTP server to send through, and the credentials if the server needs them, then writes ~/.specsr_notify.conf at mode 600 and offers to send a test message. Use an app-specific password rather than your account password: it is scoped to sending mail and can be revoked on its own.

To do it by hand, or in CI and containers where a file on disk is the awkward part, every setting is equally an environment variable:

SPECSR_NOTIFY_TO=you@example.edu
SPECSR_NOTIFY_SMTP_HOST=smtp.example.edu
SPECSR_NOTIFY_SMTP_PORT=587
SPECSR_NOTIFY_SMTP_USER=you@example.edu
SPECSR_NOTIFY_SMTP_PASS=your-app-password

./scripts/notify-run --check reports what is configured, without printing the password. See the training guide for what the messages contain.