model-scaffold

Name: model-scaffold
Author: Aperivue

208stars53forks

How to install model-scaffold?

aperivue/medsci-skills/model-scaffold

$npx -y skills add Aperivue/medsci-skills --skill model-scaffold

Installs into the current project.

Summary

Generate a reproducible, runnable PyTorch training repo for a medical-imaging task — segmentation, classification, detection, image-to-image synthesis, self-supervised pretraining, or fine-tuning a pretrained backbone (transfer learning) — the missing middle link between choosing

Files · 1

View on GitHub

SKILL.md

1	# Model-Scaffold Skill
2
3	## Purpose
4
5	This skill stamps out a runnable PyTorch training repo for a medical-imaging task — `--task`
6	segmentation (U-Net), classification (CNN / `timm` backbone), detection (torchvision Faster
7	R-CNN / FPN), synthesis (Pix2Pix generator + PatchGAN), ssl (SimCLR encoder), or finetune
8	(transfer-learning a pretrained backbone with a frozen→unfrozen schedule + a provenance record) —
9	with the reproducibility guarantees baked in by construction — so the build is leakage-safe and
10	reproducible before a single epoch runs. It is the imaging analogue of how `/analyze-stats` generates
11	runnable statistical code: the generator produces the repo, you run the training on your GPU / Colab,
12	and the lane's deterministic gates verify the network-free parts.
13
14	It is the missing middle link in the lane: `/architecture-zoo` (choose) → model-scaffold (build)
15	→ `/model-validation` (validate the split / design) → `/model-evaluation` + `/analyze-stats` (metrics)
16	→ `/write-paper` + `/check-reporting` (publish). It integrates MONAI / nnU-Net / TorchIO (referenced
17	in the generated `requirements.txt`); it does not reimplement them.
18
19	## When to use
20	- You have a data manifest (one row per image, with a patient/subject ID) and want a reproducible,
21	leakage-safe starting repo for a segmentation model.
22	- You want to fine-tune a pretrained backbone (transfer learning — the common clinician workflow:
23	a `timm` / MONAI / MedSAM checkpoint adapted to your collected clinical data) with the freeze schedule,
24	discriminative learning rates, and pretrained-weight provenance recorded (`--task finetune`).
25
26	## When NOT to use
27	- Auditing an already-trained model's validation design → `/model-validation`.
28	- Held-out metrics / calibration / bootstrap CIs → `/model-evaluation` then `/analyze-stats`.
29	- Choosing the architecture for the research question → `/architecture-zoo` (when available).
30	- Reimplementing MONAI / nnU-Net → out of scope (the scaffold integrates them).
31	- LLM / MLLM evaluation → `/mllm-eval`.
32
33	## Workflow
34
35	### Phase 1 — Prepare the manifest
36	A CSV with one row per image and a patient/subject ID column (`patient_id` / `subject_id` /
37	`case_id`), plus image and label path columns. The ID column is load-bearing: the split is done at the
38	patient level off this column.
39
40	### Phase 2 — Generate the repo
41	```bash
42	python3 ${CLAUDE_SKILL_DIR}/scripts/scaffold.py \
43	--manifest <manifest.csv> --task segmentation --out model_repo --seed 42 \
44	--in-channels 1 --out-channels 1
45	# --task = segmentation \| classification \| detection \| synthesis \| ssl \| finetune
46	# (out-channels = num classes for classification/finetune, target channels for synthesis)
47	# fine-tuning a pretrained backbone (transfer learning) on collected clinical data:
48	python3 ${CLAUDE_SKILL_DIR}/scripts/scaffold.py \
49	--manifest <manifest.csv> --task finetune --out model_repo --seed 42 \
50	--out-channels <num_classes> --from-pretrained timm:resnet50.a1_in1k
51	# emits PRETRAINED.md (provenance) + a frozen→unfrozen train.py with discriminative LRs;
52	# record the exact pretrained source so the fine-tune is reproducible.
53	```
54	This writes `model_repo/` with `config.yaml`, `model.py` (the task's model — U-Net / CNN / Faster R-CNN
55	/ Pix2Pix / SimCLR encoder), `dataset.py` (reads the frozen split), `losses.py` (task-appropriate),
56	`train.py`, `evaluate.py`, `requirements.txt`,
57	`REPRODUCIBILITY.md`, `methods_st

People who liked this also viewed

6 picks

Type

model-scaffold

208stars53forks

View on GitHub

How to install model-scaffold?

aperivue/medsci-skills/model-scaffold

$npx -y skills add Aperivue/medsci-skills --skill model-scaffold

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# Model-Scaffold Skill
2
3	## Purpose
4
5	This skill stamps out a runnable PyTorch training repo for a medical-imaging task — `--task`
6	segmentation (U-Net), classification (CNN / `timm` backbone), detection (torchvision Faster
7	R-CNN / FPN), synthesis (Pix2Pix generator + PatchGAN), ssl (SimCLR encoder), or finetune
8	(transfer-learning a pretrained backbone with a frozen→unfrozen schedule + a provenance record) —
9	with the reproducibility guarantees baked in by construction — so the build is leakage-safe and
10	reproducible before a single epoch runs. It is the imaging analogue of how `/analyze-stats` generates
11	runnable statistical code: the generator produces the repo, you run the training on your GPU / Colab,
12	and the lane's deterministic gates verify the network-free parts.
13
14	It is the missing middle link in the lane: `/architecture-zoo` (choose) → model-scaffold (build)
15	→ `/model-validation` (validate the split / design) → `/model-evaluation` + `/analyze-stats` (metrics)
16	→ `/write-paper` + `/check-reporting` (publish). It integrates MONAI / nnU-Net / TorchIO (referenced
17	in the generated `requirements.txt`); it does not reimplement them.
18
19	## When to use
20	- You have a data manifest (one row per image, with a patient/subject ID) and want a reproducible,
21	leakage-safe starting repo for a segmentation model.
22	- You want to fine-tune a pretrained backbone (transfer learning — the common clinician workflow:
23	a `timm` / MONAI / MedSAM checkpoint adapted to your collected clinical data) with the freeze schedule,
24	discriminative learning rates, and pretrained-weight provenance recorded (`--task finetune`).
25
26	## When NOT to use
27	- Auditing an already-trained model's validation design → `/model-validation`.
28	- Held-out metrics / calibration / bootstrap CIs → `/model-evaluation` then `/analyze-stats`.
29	- Choosing the architecture for the research question → `/architecture-zoo` (when available).
30	- Reimplementing MONAI / nnU-Net → out of scope (the scaffold integrates them).
31	- LLM / MLLM evaluation → `/mllm-eval`.
32
33	## Workflow
34
35	### Phase 1 — Prepare the manifest
36	A CSV with one row per image and a patient/subject ID column (`patient_id` / `subject_id` /
37	`case_id`), plus image and label path columns. The ID column is load-bearing: the split is done at the
38	patient level off this column.
39
40	### Phase 2 — Generate the repo
41	```bash
42	python3 ${CLAUDE_SKILL_DIR}/scripts/scaffold.py \
43	--manifest <manifest.csv> --task segmentation --out model_repo --seed 42 \
44	--in-channels 1 --out-channels 1
45	# --task = segmentation \| classification \| detection \| synthesis \| ssl \| finetune
46	# (out-channels = num classes for classification/finetune, target channels for synthesis)
47	# fine-tuning a pretrained backbone (transfer learning) on collected clinical data:
48	python3 ${CLAUDE_SKILL_DIR}/scripts/scaffold.py \
49	--manifest <manifest.csv> --task finetune --out model_repo --seed 42 \
50	--out-channels <num_classes> --from-pretrained timm:resnet50.a1_in1k
51	# emits PRETRAINED.md (provenance) + a frozen→unfrozen train.py with discriminative LRs;
52	# record the exact pretrained source so the fine-tune is reproducible.
53	```
54	This writes `model_repo/` with `config.yaml`, `model.py` (the task's model — U-Net / CNN / Faster R-CNN
55	/ Pix2Pix / SimCLR encoder), `dataset.py` (reads the frozen split), `losses.py` (task-appropriate),
56	`train.py`, `evaluate.py`, `requirements.txt`,
57	`REPRODUCIBILITY.md`, `methods_st

People who liked this also viewed

6 picks

Type