neiam /dms

Image: OCR models from Hugging Face, pinned to a commit

e03a73e · Gabe Morell · 2026-10-27 20:04

1 files +9 -2

Message

{commit_body(@commit)}

modified Dockerfile

+9 −2

@@ -70,10 +70,17 @@ RUN cargo build --release \

70	70	# create its output file ("Permission denied", even in a freshly mkdir'd
71	71	# directory), while the builder fetching the URL itself needs no RUN step
72	72	# at all — and so no base image, apt or curl either.
	73	+#
	74	+# From the ocrs models' Hugging Face repo, pinned to a commit: a
	75	+# resolve/<commit>/ URL can't change under us. The S3 bucket ocrs used to
	76	+# point at has moved its files around — the .rten names now 403 (the
	77	+# bucket serves .onnx, which rten 0.21 can't load). These are the same
	78	+# published checkpoints, in the header-less .rten format (V1) rten reads.
73	79	# ---------------------------------------------------------------------------
74	80	FROM scratch AS ocr-models
75		−ADD --chmod=0644 https://ocrs-models.s3-accelerate.amazonaws.com/text-detection.rten /models/text-detection.rten
76		−ADD --chmod=0644 https://ocrs-models.s3-accelerate.amazonaws.com/text-recognition.rten /models/text-recognition.rten
	81	+ARG OCRS_MODELS=https://huggingface.co/robertknight/ocrs/resolve/df0edd170279ab971b53e094c627255a87e1a503
	82	+ADD --chmod=0644 ${OCRS_MODELS}/text-detection-ssfbcj81.rten /models/text-detection.rten
	83	+ADD --chmod=0644 ${OCRS_MODELS}/text-rec-checkpoint-s52qdbqt.rten /models/text-recognition.rten
77	84
78	85	# ---------------------------------------------------------------------------
79	86	# Stage 4 — runtime: minimal Debian slim with the binaries + assets + templates.

Parents: 6228e32