Installation¶

pip install specsr-roman                 # inference + training
pip install "specsr-roman[hub]"          # + published checkpoints and data
pip install "specsr-roman[extract]"      # + dataset building
pip install "specsr-roman[all]"          # everything

Python 3.10 or newer.

To track main instead of the last release:

pip install "specsr-roman[hub] @ git+https://github.com/aryana-haghjoo/specsr-roman"

Extras¶

Extra	Pulls in	Needed for
(base)	numpy, scipy, torch, astropy, matplotlib	inference, training
`hub`	huggingface_hub	fetching published checkpoints and data
`train`	wandb	metric syncing during training
`extract`	grizli, photutils, h5py, pyarrow	building datasets from raw sims
`docs`	sphinx, furo, myst-parser	building these pages
`dev`	pytest, ruff	the test suite

The extraction stack is heavy and only needed if you are re-extracting rather than downloading the released dataset. Everything else in the package works without it.

Checking an install¶

specsr-roman info

reports the version, whether CUDA is visible, which optional extras are present, and the canonical checkpoint names.

GPU¶

Training wants a GPU; inference on a handful of spectra is fine on CPU.

On Blackwell cards (RTX 50-series with CUDA 12.8) the fused attention kernels return NaN in backward for SR2’s line-token transformer, so specsr-roman forces the math SDPA backend during SR2 training. The transformer is 98 tokens wide, so the cost is negligible.

Configuring paths¶

Data locations are environment-overridable and default to the working directory:

Variable	Default	What
`SPECSR_ROMAN_DATA`	`./data`	raw and prepared simulation products
`SPECSR_ROMAN_DATASETS`	`$SPECSR_ROMAN_DATA/dataset`	built training datasets
`SPECSR_ROMAN_RUNS`	`./runs`	checkpoints and predictions
`SPECSR_ROMAN_HUB_REPO`	`aryana-haghjoo/roman-spectral-superresolution`	checkpoint source
`GRIZLI`	`./grizli_conf`	grizli `CONF` tree (extraction only)