Skip to content
Closed
Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
121 commits
Select commit Hold shift + click to select a range
0fbf606
fix: use bf16-mixed precision and gradient accumulation to avoid OOM
hannesill Mar 2, 2026
f73e4b6
fix: cast vis_proj dtype in MAE decoder scatter to avoid bf16-mixed m…
hannesill Mar 2, 2026
4b44760
refactor: switch SSL objectives from observation-level to timestep-le…
hannesill Mar 2, 2026
a2d47e6
refactor: extract data and training utilities into focused modules
hannesill Mar 3, 2026
22a93b3
fix: cast vis_proj dtype in JEPA predictor scatter to avoid bf16-mixe…
hannesill Mar 3, 2026
64dce23
perf: reduce memory footprint and revert to fp32 precision
hannesill Mar 3, 2026
5f071d0
feat: add temporal NT-Xent mode to contrastive objective
hannesill Mar 3, 2026
dee1e70
fix: tensor cache never loads due to `or` on multi-element tensors
hannesill Mar 3, 2026
6c60e29
fix: handle PyTorch 2.6 weights_only default in checkpoint loading
hannesill Mar 3, 2026
387d9d9
fix: prevent JEPA representational collapse
hannesill Mar 3, 2026
5fee7f1
fix: match checkpoint filename placeholders to logged metric names
hannesill Mar 3, 2026
7978678
fix: address JEPA collapse with higher momentum, block masking, and m…
hannesill Mar 3, 2026
e06a523
fix: shorten training schedule and prevent early stopping during warmup
hannesill Mar 3, 2026
da1520f
feat: add sprint 1 launch script and VM auto-shutdown
hannesill Mar 4, 2026
f630e28
fix: use fixed-schedule training and save both last/best encoders
hannesill Mar 4, 2026
e91a6bb
refactor: unify encoder architecture across all paradigms (174K params)
hannesill Mar 4, 2026
497916b
feat: switch contrastive to instance mode with collapse monitoring
hannesill Mar 4, 2026
df8b612
fix: add dropout to obs_proj and increase supervised regularization
hannesill Mar 4, 2026
8099dd7
feat: add parallel experiment runner for all 9 sprints
hannesill Mar 4, 2026
38ac46c
fix: revert supervised-specific regularization, enable balanced class…
hannesill Mar 4, 2026
0fc6821
fix: AKI KDIGO label builder detection window and class weight guard
hannesill Mar 5, 2026
3eaa283
fix: add downstream regularization to combat overfitting
hannesill Mar 5, 2026
174a5f8
feat: add warmup subcommand to pre-build tensor caches before paralle…
hannesill Mar 5, 2026
9efcafd
feat: add lr=2e-4 to LR ablation sweep for sprint 1b
hannesill Mar 12, 2026
506b957
feat: add complementary masks for contrastive SSL objective
hannesill Mar 12, 2026
011d1f9
feat: add Sprint 1c mask ratio sensitivity study and fix run counts
hannesill Mar 12, 2026
ffdee89
feat: add forward-looking prediction window for AKI labeling
hannesill Mar 12, 2026
f5c30aa
feat: add --revision and --reason flags to experiment runner
hannesill Mar 12, 2026
82604fc
feat: improve wandb tagging and add run query script
hannesill Mar 12, 2026
726d74a
feat: import only essential RICU tables to speed up extraction
hannesill Mar 13, 2026
f5528cd
feat: add sprint baseline tagging and update experiment plan with res…
hannesill Mar 13, 2026
db005fa
fix: truncate long rerun-reason wandb tags to 64 characters
hannesill Mar 13, 2026
942c45d
fix: include run_experiments and prepare_dataset in auto-shutdown act…
hannesill Mar 18, 2026
32d4266
fix: reduce peak memory in tensor preprocessing pipeline
hannesill Mar 18, 2026
3da6cfc
fix: avoid OOM in warmup and tensor extraction for large datasets
hannesill Mar 18, 2026
ad63aa3
fix: derive eICU date_of_death from hospitaldischargeoffset
hannesill Mar 18, 2026
c2572ed
fix: detect interactive sessions in auto-shutdown activity check
hannesill Mar 18, 2026
bd8119e
fix: improve eICU extraction robustness and compatibility
hannesill Mar 17, 2026
4a179ef
fix: harmonize column dtypes before concatenating datasets
hannesill Mar 17, 2026
318f883
fix: derive eICU date_of_death from hospitaldischargeoffset for morta…
hannesill Mar 19, 2026
50d0863
feat: chunk RICU timeseries output by stay to reduce memory usage
hannesill Mar 21, 2026
47d1c7a
feat: add Sprint 7p model capacity pilot for SSL label-efficiency abl…
hannesill Mar 23, 2026
76d3541
fix: remove extraction-only fields from training task config
hannesill Mar 22, 2026
65dbf61
feat: add Sprints 10-12 for 5-seed expansion, classical baselines, an…
hannesill Mar 24, 2026
8ad2651
feat: add GRU-D encoder and XGBoost baseline
hannesill Mar 24, 2026
328a731
fix: cache raw tensors per dataset to eliminate 140GB+ cache bloat
hannesill Mar 26, 2026
582e0e4
feat: add canonical results export pipeline from W&B
hannesill Mar 27, 2026
d4b585b
feat: add post-run fairness evaluation script with W&B write-back
hannesill Mar 27, 2026
c5667a5
fix: handle Lightning checkpoint dir issue in fairness eval
hannesill Mar 27, 2026
6242014
fix: replace '/' with '_' in checkpoint filenames to prevent silent s…
hannesill Mar 27, 2026
7c5c742
fix: use summary_metrics to avoid N+1 W&B API calls
hannesill Mar 28, 2026
bb6f016
fix: align SMART implementation with original paper code
hannesill Mar 30, 2026
23e2dd8
feat: add TS2Vec temporal contrastive SSL objective
hannesill Mar 30, 2026
ded458b
feat: add offline FLOPs measurement script for SSL paradigm comparison
hannesill Mar 30, 2026
b67a82a
fix: export pipeline handles sprint 10/11 correctly
hannesill Apr 1, 2026
cf103d7
fix: implement FIXES.md Issues 2-9 for experiment integrity recovery
hannesill Apr 2, 2026
c05c28e
Add agent files to gitignore
hannesill Apr 7, 2026
930ce5e
fix: harden preprocessing cache freshness and experiment cleanup
hannesill Apr 7, 2026
2f52316
feat: add thesis rerun launcher and wandb targeting
hannesill Apr 7, 2026
4a329c7
feat: expand thesis sprint coverage
hannesill Apr 7, 2026
3cb462f
docs: align thesis plan with formal sprint scope
hannesill Apr 7, 2026
97f30e7
Fix audit issues across training, extraction, and export
hannesill Apr 10, 2026
c4c372e
Canonicalize thesis baseline matrix
hannesill Apr 11, 2026
e10aeec
Use 24h observation window defaults
hannesill Apr 12, 2026
d5f754b
Align 24h benchmark defaults and add regressions
hannesill Apr 12, 2026
6148462
test: derive conftest fixture seq_length from benchmark constant
hannesill Apr 12, 2026
8b86687
Surface AKI label quality and low-positive train subsets
hannesill Apr 12, 2026
d72e796
Validate raw label horizon for 24h benchmark tasks
hannesill Apr 13, 2026
ee4e9d8
Align configs and utilities with 24h benchmark
hannesill Apr 14, 2026
400f6ee
Persist AKI label quality stats and clarify raw export horizon
hannesill Apr 14, 2026
09a0176
Unblock combined prep, SMART teachers, and fairness revisions
hannesill Apr 14, 2026
bb018dd
Finalize experiment readiness fixes across fairness, export, and SSL
hannesill Apr 14, 2026
ab696e8
Align medium-priority experiment fixes
hannesill Apr 14, 2026
7bd520a
Refresh tracked documentation for the current thesis workflow
hannesill Apr 14, 2026
323a1fa
Guard sparse contrastive views and clarify fairness thresholds
hannesill Apr 14, 2026
c2ef1c6
Normalize RICU mortality timestamps before label comparisons
hannesill Apr 16, 2026
f4afef8
Mask fully unobserved timesteps in JEPA predictor
hannesill Apr 16, 2026
1b14b88
Fix experiment audit gaps in export and fairness flows
hannesill Apr 16, 2026
ebf9a09
Harden label freshness and extraction safety checks
hannesill Apr 16, 2026
6b9d4eb
Fix evaluation defaults and downstream audit regressions
hannesill Apr 16, 2026
726f884
Validate cached splits against task cohorts
hannesill Apr 20, 2026
a3cda16
Preserve experiment provenance and scheduler safeguards
hannesill Apr 20, 2026
6717f29
Fix SSL timestep reassembly for sparse batches
hannesill Apr 21, 2026
8e74403
Prepare repository for public benchmark release
hannesill Apr 21, 2026
dbf9aa9
Document AKI label filtering in README
hannesill Apr 21, 2026
866cd65
Tighten thesis experiment evaluation scope
hannesill Apr 21, 2026
cd3a34d
Harden benchmark validation gates
hannesill Apr 21, 2026
87eff39
Add processed data rebuild wrapper
hannesill Apr 21, 2026
2120ca4
Filter invalid RICU stays before cohort validation
hannesill Apr 21, 2026
94bb249
Fix thesis preflight training blockers
hannesill Apr 21, 2026
f28f098
Fix thesis export statistical safeguards
hannesill Apr 21, 2026
ea58cb3
Add operational experiment guardrails
hannesill Apr 21, 2026
cdf032f
Honor inherited capacity baselines during export
hannesill Apr 22, 2026
0f02f79
Fix benchmark leakage and experiment safeguards
hannesill Apr 22, 2026
ef18cac
Exclude derived RICU features from extraction
hannesill Apr 22, 2026
e0a887e
Align downstream normalization with SSL cohorts
hannesill Apr 22, 2026
c725bfe
Harden export and fairness integrity
hannesill Apr 22, 2026
be9d6ef
Harden baseline and protocol runner semantics
hannesill Apr 22, 2026
3303d26
Migrate thesis runs to experiment classes
hannesill Apr 22, 2026
f8c8a75
Harden thesis launch and export validation
hannesill Apr 22, 2026
7f1c076
Harden experiment rerun validation
hannesill Apr 22, 2026
6f0ceaf
Fix launch identity and checkpoint safeguards
hannesill Apr 22, 2026
00bbcca
Require scoped thesis experiment launches
hannesill Apr 22, 2026
fba0e7b
Clarify controlled SSL masking semantics
hannesill Apr 22, 2026
32b406e
Harden thesis export and masking checks
hannesill Apr 22, 2026
02d6c9d
Prepare benchmark release exports
hannesill Apr 23, 2026
9f2f79f
Fix mortality label handling for unknown eICU outcomes
hannesill Apr 23, 2026
9f8d714
Tighten final launch provenance
hannesill Apr 23, 2026
89e757f
Align publication fairness and XGBoost metrics
hannesill Apr 23, 2026
eee1155
Fix final experiment matrix safeguards
hannesill Apr 23, 2026
2cc5fc3
Harden benchmark launch and artifact provenance
hannesill Apr 23, 2026
ffb2679
Harden thesis benchmark guardrails
hannesill Apr 23, 2026
d40e12d
Rename downstream protocols for public benchmark
hannesill Apr 23, 2026
fe99fb7
Track codex sessions in auto shutdown
hannesill Apr 23, 2026
ffd6dc5
Handle invalid W&B entity errors cleanly
hannesill Apr 23, 2026
d47fe18
Harden final experiment launch failures
hannesill Apr 23, 2026
600d6a6
Fix revision retry dependent recovery
hannesill Apr 23, 2026
926acf4
Harden final experiment launch guardrails
hannesill Apr 23, 2026
cde200b
Tighten SSL and training configuration contracts
hannesill Apr 23, 2026
24cbc24
Harden final experiment guardrails
hannesill Apr 24, 2026
4d64025
Clarify SSL objective scope
hannesill May 16, 2026
File filter

Filter by extension

Filter by extension


Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
36 changes: 36 additions & 0 deletions .github/workflows/ci.yml
Original file line number Diff line number Diff line change
@@ -0,0 +1,36 @@
name: CI

on:
push:
branches: [main, sprints]
pull_request:

jobs:
checks:
runs-on: ubuntu-latest

steps:
- name: Check out repository
uses: actions/checkout@v4

- name: Set up Python
uses: actions/setup-python@v5
with:
python-version: "3.12"

- name: Set up uv
uses: astral-sh/setup-uv@v5
with:
enable-cache: true

- name: Install dependencies
run: uv sync --dev --frozen

- name: Lint
run: uv run ruff check src scripts tests

- name: Type check
run: uv run mypy src

- name: Test
run: uv run pytest -q
8 changes: 7 additions & 1 deletion .gitignore
Original file line number Diff line number Diff line change
Expand Up @@ -195,11 +195,14 @@ cython_debug/
.pypirc

# AI
.agents/
.codex/
.cursorignore
.cursorindexingignore
.cursorrules
.claude/
.cursor/
.rulesync/

# Marimo
marimo/_static/
Expand All @@ -208,13 +211,16 @@ __marimo__/

# Notes
notes/
docs/
docs/internal/

# SLICES-specific
/data/
outputs/
/results/*
!/results/.gitkeep
wandb/
.tensor_cache/
logs/

# PyTorch Lightning
lightning_logs/
Expand Down
2 changes: 1 addition & 1 deletion .pre-commit-config.yaml
Original file line number Diff line number Diff line change
Expand Up @@ -14,7 +14,7 @@ repos:
args: [--line-length=100, --target-version=py312]

- repo: https://github.com/astral-sh/ruff-pre-commit
rev: v0.3.4
rev: v0.14.8
hooks:
- id: ruff
args: [--fix]
Expand Down
6 changes: 6 additions & 0 deletions EXPERIMENTS_PLAN.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,6 @@
# Experiment Plan Pointer

The active SLICES experiment plan lives at `docs/internal/EXPERIMENT_PLAN.md`.

This file exists so tools or reviews that look for `EXPERIMENTS_PLAN.md` land on
the canonical plan rather than treating the plan as missing.
Loading
Loading