ryan/hypertower
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

update 3-19

Browse Source
This commit is contained in:
rpotter6298
2026-03-19 11:18:58 +01:00
parent 7ea85d5426
commit 786457b30d
35 changed files with 4019 additions and 258 deletions
Download Patch File Download Diff File Expand all files Collapse all files
scripts/main/pipeline.ipynb
+28 -83
View File
@@ -56,6 +56,7 @@
{
"cell_type": "code",
"execution_count": 4,
"id": "6a600aed",
"metadata": {},
"outputs": [
{
@@ -84,6 +85,7 @@
},
{
"cell_type": "markdown",
"id": "d1ea8b19",
"metadata": {},
"source": [
"## 1) Build UNet Manifest"
@@ -92,6 +94,7 @@
{
"cell_type": "code",
"execution_count": 5,
"id": "b07b3e69",
"metadata": {},
"outputs": [
{
@@ -108,21 +111,28 @@
},
{
"cell_type": "markdown",
"id": "dgglybgo5wg",
"source": "## 2) Build Refugelike Backbone\n\nThe `refugelike` backbone is a ResNet-50 pre-trained on REFUGE as an optic disc/cup classifier, then stripped of its classification head and used as a frozen or partially-frozen feature extractor in the HyperTower image tower.\n\n**Steps:**\n1. Train the REFUGE classifier (`--train-clf`)\n2. Export its backbone weights to `models/v2/refuge/refugelike_backbone.pt` (`--export-backbone`)\n\nThe classifier checkpoint is saved to `models/v2/refuge/classifier/resnet50/refuge_classifier_best.pt` by default. \nThe exported backbone is what `--backbone refugelike` loads at runtime (see `classes/v2/backbones.py`).",
"metadata": {}
},
{
"cell_type": "code",
"id": "b7jt033ul4v",
"source": "BACKBONE_PATH = \"models/v2/refuge/refugelike_backbone.pt\"\n\n# Step 1: train the REFUGE classifier (ResNet-50, 30 epochs by default)\n!python3 scripts/main/refuge/refuge_build.py \\\n --train-clf \\\n --manifest manifest.csv \\\n --device cuda\n\n# Step 2: strip the head and export backbone weights\n!python3 scripts/main/refuge/refuge_build.py \\\n --export-backbone {BACKBONE_PATH} \\\n --manifest manifest.csv \\\n --device cuda",
"metadata": {},
"source": [
"## 2) Train UNet Segmenter (per-image normalization)\n",
"\n",
"Current tuned baseline:\n",
"- `--device cuda`\n",
"- `--batch-size 8`\n",
"- `--loader-workers 14`\n",
"- `--in-memory-cache`\n",
"- `--cache-workers 4`"
]
"execution_count": null,
"outputs": []
},
{
"cell_type": "markdown",
"id": "8df524a2",
"metadata": {},
"source": "## 2b) Train UNet Segmenter (per-image normalization)\n\nCurrent tuned baseline:\n- `--device cuda`\n- `--batch-size 8`\n- `--loader-workers 14`\n- `--in-memory-cache`\n- `--cache-workers 4`"
},
{
"cell_type": "code",
"execution_count": 7,
"id": "be2b499a",
"metadata": {},
"outputs": [
{
@@ -1270,6 +1280,7 @@
},
{
"cell_type": "markdown",
"id": "a2a3bd04",
"metadata": {},
"source": [
"## 5) Run Pipeline Experiments\n",
@@ -1283,6 +1294,7 @@
{
"cell_type": "code",
"execution_count": null,
"id": "e395f268",
"metadata": {},
"outputs": [],
"source": [
@@ -1301,41 +1313,18 @@
{
"cell_type": "code",
"execution_count": null,
"id": "125f35c6",
"metadata": {},
"outputs": [],
"source": [
"# 3b) GT crop — expert segmentation masks crop the optic disc region\n",
"!python scripts/main/v2/multirun_hypertower.py \\\n",
" --tower-modes single ensemble \\\n",
" --eval-modes binary multiclass \\\n",
" --epochs 40 --n-splits 5 \\\n",
" --backbone refugelike \\\n",
" --img-crop-gt \\\n",
" --warmup-md-epochs 50 \\\n",
" --fused-head \\\n",
" --run-name pipeline_gt"
]
"source": "# 3b) GT crop — expert segmentation masks crop the optic disc region\n!python scripts/main/v2/multirun_hypertower.py \\\n --tower-modes single ensemble \\\n --eval-modes binary multiclass \\\n --epochs 40 --n-splits 5 \\\n --backbone refugelike \\\n --img-crop-manifest manifest.csv \\\n --img-crop-gt \\\n --warmup-md-epochs 50 \\\n --fused-head \\\n --run-name pipeline_gt"
},
{
"cell_type": "code",
"execution_count": null,
"id": "83df73ac",
"metadata": {},
"outputs": [],
"source": [
"# 3c) UNet crop — trained segmenter crops the optic disc region\n",
"!python scripts/main/v2/multirun_hypertower.py \\\n",
" --tower-modes single ensemble \\\n",
" --eval-modes binary multiclass \\\n",
" --epochs 40 --n-splits 5 \\\n",
" --backbone refugelike \\\n",
" --img-crop-manifest analysis_data/unet_manifest.csv \\\n",
" --img-crop-weights models/v2/refuge/segmentation/per_image/best.pt \\\n",
" --img-crop-normalize per_image \\\n",
" --img-crop-cache analysis_data/v2_crops_unet_refuge \\\n",
" --warmup-md-epochs 50 \\\n",
" --fused-head \\\n",
" --run-name pipeline_unet"
]
"source": "# 3c) UNet crop — trained segmenter crops the optic disc region\n!python scripts/main/v2/multirun_hypertower.py \\\n --tower-modes single ensemble \\\n --eval-modes binary multiclass \\\n --epochs 40 --n-splits 5 \\\n --backbone refugelike \\\n --img-crop-manifest manifest.csv \\\n --img-crop-weights models/v2/refuge/segmentation/per_image/best.pt \\\n --img-crop-normalize per_image \\\n --img-crop-cache analysis_data/v2_crops_unet_refuge \\\n --warmup-md-epochs 50 \\\n --fused-head \\\n --run-name pipeline_unet"
},
{
"cell_type": "markdown",
@@ -1353,51 +1342,7 @@
"id": "7yrcfu0bv1w",
"metadata": {},
"outputs": [],
"source": [
"import subprocess\n",
"from pathlib import Path\n",
"\n",
"RUN_DIRS = {\n",
" \"nocrop\": Path(\"analysis_data/pipeline_nocrop\"),\n",
" \"gt\": Path(\"analysis_data/pipeline_gt\"),\n",
" \"unet\": Path(\"analysis_data/pipeline_unet\"),\n",
"}\n",
"EVAL_MODES = [\"binary\", \"multiclass\"]\n",
"TOWER_MODES = [\"single\", \"ensemble\"]\n",
"\n",
"for run_name, run_dir in RUN_DIRS.items():\n",
" for eval_mode in EVAL_MODES:\n",
" for tower_mode in TOWER_MODES:\n",
" mode_dir = run_dir / eval_mode / tower_mode\n",
" if not mode_dir.exists():\n",
" print(f\" skip (not found): {mode_dir}\")\n",
" continue\n",
" print(f\"--- {run_name} / {eval_mode} / {tower_mode} ---\")\n",
"\n",
" # ROC curves\n",
" subprocess.run([\n",
" \"python\", \"scripts/output_analysis/visualizations/plot_run_roc_v2.py\",\n",
" \"--run-dir\", str(run_dir),\n",
" \"--eval-mode\", eval_mode,\n",
" \"--tower-mode\", tower_mode,\n",
" ], check=True)\n",
"\n",
" # Probability strips — binary only\n",
" if eval_mode == \"binary\":\n",
" subprocess.run([\n",
" \"python\", \"scripts/output_analysis/visualizations/plot_prob_strips.py\",\n",
" \"--run-dir\", str(mode_dir),\n",
" \"--head\", \"fused\", \"--style\", \"strips\",\n",
" ], check=True)\n",
"\n",
" # Probability triangle-3D — multiclass only\n",
" if eval_mode == \"multiclass\":\n",
" subprocess.run([\n",
" \"python\", \"scripts/output_analysis/visualizations/plot_prob_strips.py\",\n",
" \"--run-dir\", str(mode_dir),\n",
" \"--head\", \"fused\", \"--style\", \"triangle3d\",\n",
" ], check=True)"
]
"source": "import subprocess\nfrom pathlib import Path\n\nRUN_DIRS = {\n \"nocrop\": (Path(\"analysis_data/pipeline_nocrop\"), [\"single\", \"ensemble\"]),\n \"gt\": (Path(\"analysis_data/pipeline_gt\"), [\"single\", \"ensemble\"]),\n \"unet\": (Path(\"analysis_data/pipeline_unet\"), [\"single\", \"ensemble\"]),\n \"imgonly_nocrop\":(Path(\"analysis_data/pipeline_imgonly_nocrop\"),[\"single\"]),\n \"imgonly_gt\": (Path(\"analysis_data/pipeline_imgonly_gt\"), [\"single\"]),\n \"imgonly_unet\": (Path(\"analysis_data/pipeline_imgonly_unet\"), [\"single\"]),\n}\nEVAL_MODES = [\"binary\", \"multiclass\"]\n\nfor run_name, (run_dir, tower_modes) in RUN_DIRS.items():\n for eval_mode in EVAL_MODES:\n for tower_mode in tower_modes:\n mode_dir = run_dir / eval_mode / tower_mode\n if not mode_dir.exists():\n print(f\" skip (not found): {mode_dir}\")\n continue\n print(f\"--- {run_name} / {eval_mode} / {tower_mode} ---\")\n\n # ROC curves (auto-detects all available probs stems)\n subprocess.run([\n \"python\", \"scripts/output_analysis/visualizations/plot_run_roc_v2.py\",\n \"--run-dir\", str(run_dir),\n \"--eval-mode\", eval_mode,\n \"--tower-mode\", tower_mode,\n ], check=True)\n\n # Probability strips — binary only (auto-detects all heads)\n if eval_mode == \"binary\":\n subprocess.run([\n \"python\", \"scripts/output_analysis/visualizations/plot_prob_strips.py\",\n \"--run-dir\", str(mode_dir),\n \"--style\", \"strips\",\n ], check=True)\n\n # Probability triangle-3D — multiclass only (auto-detects all heads)\n if eval_mode == \"multiclass\":\n subprocess.run([\n \"python\", \"scripts/output_analysis/visualizations/plot_prob_strips.py\",\n \"--run-dir\", str(mode_dir),\n \"--style\", \"triangle3d\",\n ], check=True)"
},
{
"cell_type": "markdown",
@@ -1458,4 +1403,4 @@
},
"nbformat": 4,
"nbformat_minor": 5
}
}
scripts/main/refuge/refuge_build.py
+3 -3
View File
@@ -6,7 +6,7 @@ Usage examples (after activating .venv_refuge):
python refuge_build.py --eval --with-ttt
The script expects the REFUGE folder and writes checkpoints under
models/refuge/segmentation and models/refuge/classifier.
models/v2/refuge/segmentation and models/v2/refuge/classifier.
"""
from __future__ import annotations
@@ -46,7 +46,7 @@ from classes.papila_builders import build_papila_clinical
REFUGE_ROOT = Path("REFUGE")
SEG_CKPT = Path("models/refuge/segmentation/refuge_segmentation_best.pt")
CLF_DIR = Path("models/refuge/classifier")
CLF_DIR = Path("models/v2/refuge/classifier")
UNET_WEIGHT_CANDIDATES = (
Path("models/v2/refuge/segmentation/per_image/best.pt"),
Path("models/v2/refuge/segmentation/best.pt"),
@@ -820,7 +820,7 @@ def parse_args() -> argparse.Namespace:
"--clf-checkpoint-path",
type=Path,
default=None,
help="Optional explicit path for the classifier checkpoint (defaults to models/refuge/classifier/<backbone>/refuge_classifier_best.pt)",
help="Optional explicit path for the classifier checkpoint (defaults to models/v2/refuge/classifier/<backbone>/refuge_classifier_best.pt)",
)
parser.add_argument(
"--eval-datasets",
scripts/main/v2/multirun_hypertower.py
+1 -1
View File
@@ -7,7 +7,7 @@ from pathlib import Path
import sys
import argparse
REPO_ROOT = Path(__file__).resolve().parents[2]
REPO_ROOT = Path(__file__).resolve().parents[3]
if str(REPO_ROOT) not in sys.path:
sys.path.insert(0, str(REPO_ROOT))
scripts/main/v2/run_10x5cv.py
+107
View File
@@ -0,0 +1,107 @@
#!/usr/bin/env python3
"""
10× repeated 5-fold CV runner for the best hypertower configuration
(nocrop, ensemble mode, both binary and multiclass).
Each repetition uses a different fold-seed so the 5 folds are split
differently, giving 50 folds per eval-mode total. Holdout composition
is kept identical across repetitions (same --holdout-seed).
Results land under:
{output-root}/rep{N:02d}/{eval_mode}/ensemble/fold{K}/
Usage
-----
python scripts/main/v2/run_10x5cv.py \
--n-reps 10 \
--eval-modes binary multiclass \
--output-root analysis_data/pipeline_10x5 \
--epochs 40 --fused-head \
--backbone refugelike
Any extra flags are forwarded directly to V2HyperTower.
"""
from __future__ import annotations
import argparse
import sys
from pathlib import Path
REPO_ROOT = Path(__file__).resolve().parents[3]
if str(REPO_ROOT) not in sys.path:
sys.path.insert(0, str(REPO_ROOT))
from classes.v2.v2_hypertower import V2HyperTower
# Base fold seed for rep 0; rep N uses BASE_SEED + N * SEED_STRIDE
_BASE_SEED = 100
_SEED_STRIDE = 100
def _parse_own(argv=None):
ap = argparse.ArgumentParser(
description=__doc__,
formatter_class=argparse.RawDescriptionHelpFormatter,
add_help=False,
)
ap.add_argument("--n-reps", type=int, default=10,
help="Number of repetitions (default: 10).")
ap.add_argument("--eval-modes", nargs="+",
choices=["binary", "multiclass"],
default=["binary", "multiclass"])
ap.add_argument("--output-root", default="analysis_data/pipeline_10x5",
help="Parent directory for all rep sub-runs.")
ap.add_argument("-h", "--help", action="store_true")
return ap.parse_known_args(argv)
def main(argv=None):
own, remaining = _parse_own(argv)
if own.help:
print(__doc__)
base_parser = V2HyperTower.build_parser()
base_parser.print_help()
return
base_parser = V2HyperTower.build_parser()
output_root = Path(own.output_root)
first_run = True
for rep in range(own.n_reps):
fold_seed = _BASE_SEED + rep * _SEED_STRIDE
rep_label = f"rep{rep:02d}"
for eval_mode in own.eval_modes:
tower_mode = "ensemble"
# Skip if already fully complete
tm_dir = output_root / rep_label / eval_mode / tower_mode
if (tm_dir / "summary.json").exists():
print(f"[10x5cv] {rep_label} {eval_mode}:{tower_mode} — already done, skipping.")
first_run = False
continue
cli = list(remaining) + [
"--eval-mode", eval_mode,
"--tower-mode", tower_mode,
"--fold-seed", str(fold_seed),
"--run-name", rep_label,
"--output-root", str(output_root),
]
# Reuse crop cache across runs after the first
if not first_run:
cli.append("--persist-img-crop-cache")
print(f"\n[10x5cv] Starting {rep_label} {eval_mode}:{tower_mode} "
f"(fold_seed={fold_seed})")
args = base_parser.parse_args(cli)
V2HyperTower(args).run()
first_run = False
print(f"\n[10x5cv] All done. Results in: {output_root}")
if __name__ == "__main__":
main()
scripts/main/v2/run_iop_corr_comparison.py
+90
View File
@@ -0,0 +1,90 @@
#!/usr/bin/env python3
"""
Run single-mode binary + multiclass for each IOP correction method and
collect all results under one output root for easy comparison.
Output layout:
analysis_data/iop_corr_comparison/
ratio/binary/single/ ratio/multiclass/single/
ols/binary/single/ ols/multiclass/single/
lad/binary/single/ lad/multiclass/single/
multi/binary/single/ multi/multiclass/single/
Usage
-----
python scripts/main/v2/run_iop_corr_comparison.py [V2HyperTower args...]
Any extra args (backbone, epochs, img-crop-*, etc.) are forwarded to every run.
"""
from __future__ import annotations
import sys
from pathlib import Path
REPO_ROOT = Path(__file__).resolve().parents[3]
if str(REPO_ROOT) not in sys.path:
sys.path.insert(0, str(REPO_ROOT))
from classes.v2.v2_hypertower import V2HyperTower
IOP_METHODS = ["ratio", "ols", "lad", "multi"]
EVAL_MODES = ["binary", "multiclass"]
OUTPUT_ROOT = "analysis_data/iop_corr_comparison"
def main() -> None:
base_parser = V2HyperTower.build_parser()
# Consume only the remaining (forwarded) args — iop-corr-method and
# run-name are set by this script; eval-mode and tower-mode likewise.
_, remaining = base_parser.parse_known_args()
first_run = True
for method in IOP_METHODS:
for eval_mode in EVAL_MODES:
run_name = method # one sub-folder per method
tm_dir = (Path(OUTPUT_ROOT) / run_name / eval_mode / "single")
if (tm_dir / "summary.json").exists():
print(f"[iop_corr] {method}/{eval_mode}/single — already done, skipping.")
first_run = False
continue
cli = list(remaining) + [
"--eval-mode", eval_mode,
"--tower-mode", "single",
"--iop-corr-method", method,
"--output-root", OUTPUT_ROOT,
"--run-name", run_name,
]
if not first_run:
cli.append("--persist-img-crop-cache")
print(f"\n[iop_corr] Starting {method}/{eval_mode}/single ...")
args = base_parser.parse_args(cli)
V2HyperTower(args).run()
first_run = False
# ── summary table ──────────────────────────────────────────────────────
import json
print("\n" + "=" * 60)
print("IOP correction method comparison — single mode")
print("=" * 60)
header = f"{'Method':<8} {'Mode':<12} {'Val AUC':>10} {'Hld AUC':>10}"
print(header)
print("-" * len(header))
for method in IOP_METHODS:
for eval_mode in EVAL_MODES:
p = Path(OUTPUT_ROOT) / method / eval_mode / "single" / "summary.json"
if not p.exists():
print(f"{method:<8} {eval_mode:<12} {'missing':>10} {'missing':>10}")
continue
ms = json.loads(p.read_text()).get("mode_summary", {})
val = ms.get("classic_best_val", {})
hld = ms.get("classic_holdout", {})
val_s = f"{val['auc_mean']:.3f}±{val['auc_std']:.3f}" if val.get("auc_mean") else "—"
hld_s = f"{hld['auc_mean']:.3f}±{hld['auc_std']:.3f}" if hld.get("auc_mean") else "—"
print(f"{method:<8} {eval_mode:<12} {val_s:>10} {hld_s:>10}")
print("=" * 60)
if __name__ == "__main__":
main()
scripts/main/v2/run_md_mlp.py
+23
View File
@@ -323,6 +323,7 @@ def main() -> None:
best_epoch = 0
best_phase = ""
best_state = None
epoch_log_rows = []
print(
f"\n[fold {fold_idx+1}/{args.n_splits}] "
@@ -350,6 +351,14 @@ def main() -> None:
aggregate_patient=aggregate_patient,
)
hld_auc_ep = float("nan")
hld_acc_ep = float("nan")
if holdout_loader is not None:
_, _, hld_auc_ep, hld_acc_ep = _evaluate_single(
model, holdout_loader, device, num_classes,
aggregate_patient=aggregate_patient,
)
is_main = phase == "main"
if is_main and (not np.isnan(val_auc)) and val_auc > best_auc:
best_auc = float(val_auc)
@@ -357,6 +366,17 @@ def main() -> None:
best_epoch = ep + 1
best_phase = phase
epoch_log_rows.append({
"epoch": ep + 1,
"phase": phase,
"train_loss": float(tr_loss),
"train_acc": float(tr_acc),
"val_auc": float(val_auc),
"val_acc": float(val_acc),
"hld_auc": float(hld_auc_ep),
"hld_acc": float(hld_acc_ep),
})
if ep == 0 or (ep + 1) % 10 == 0 or (ep + 1) == total_epochs:
print(
f" ep {ep+1:>3}/{total_epochs} [{phase}:{main_ep}/{args.epochs}] "
@@ -366,6 +386,9 @@ def main() -> None:
flush=True,
)
import pandas as _pd
_pd.DataFrame(epoch_log_rows).to_csv(fold_dir / "epoch_log.csv", index=False)
if best_state is not None:
model.load_state_dict(best_state)
scripts/main/v2/run_multifold_v2_modes.py
+1 -1
View File
@@ -6,7 +6,7 @@ from __future__ import annotations
from pathlib import Path
import sys
REPO_ROOT = Path(__file__).resolve().parents[1]
REPO_ROOT = Path(__file__).resolve().parents[3]
if str(REPO_ROOT) not in sys.path:
sys.path.insert(0, str(REPO_ROOT))