MyoSeg / self_train.py
skarugu's picture
Update self_train.py
f469e6c verified
"""
self_train.py
=============
Autonomous continual-learning pipeline for MyoSeg.
Place at the ROOT of your Hugging Face Space repo (same level as Dockerfile).
IMPORTANT: This file is completely self-contained.
It does NOT import from train_myotube_nuclei_unet.py.
The train script is a separate PyCharm tool.
Trigger conditions (any one fires a retrain):
1. User submitted corrected label pairs via the app β†’ corrections/ folder
2. N unlabelled images accumulated in queue β†’ retrain_queue/
3. K consecutive low-confidence images β†’ retrain_queue/ (reason=low_confidence)
4. Nightly scheduled run β†’ APScheduler cron 02:00 UTC
5. User-optimized parameters submitted β†’ corrections/ (reason=user_optimized_params)
These submissions include the image, postprocessed masks from the user's
tuned parameter set, and a full snapshot of the sidebar settings. The
self-training pipeline uses these as additional supervised training pairs
and can aggregate parameter statistics to learn optimal defaults.
Privacy note:
Images processed in Private Mode (toggle in the Streamlit sidebar) are
NEVER queued for retraining. Only images explicitly submitted by the
user via "Submit for training" or "Submit corrections" are used.
After each retrain:
β€’ Fine-tunes from current HF Hub weights
β€’ Validates on held-out 20% split
β€’ Only pushes to Hub if new Dice > previous best
β€’ Archives queue β†’ runs/<run_id>/processed_queue/
β€’ Appends entry to manifest.json
β€’ Aggregates user-submitted parameter snapshots β†’ optimal_params.json
Usage:
python self_train.py # check triggers once
python self_train.py --manual # force retrain now
python self_train.py --scheduler # blocking APScheduler loop (for Docker)
Environment variables / HF Secrets:
HF_TOKEN write-access Hugging Face token
HF_REPO_ID model repo, e.g. "skarugu/myotube-unet"
HF_FILENAME model filename, e.g. "model_final.pt"
DATA_ROOT path to base training data/ folder
BATCH_TRIGGER_N images before batch trigger (default 20)
CONF_DROP_K consecutive low-conf before trigger (default 5)
FT_EPOCHS fine-tuning epochs per run (default 10)
FT_LR fine-tuning learning rate (default 5e-4)
SCHEDULE_HOUR nightly retrain UTC hour (default 2)
"""
import argparse
import json
import logging
import os
import random
import shutil
import tempfile
from datetime import datetime
from pathlib import Path
from typing import Optional
import numpy as np
import scipy.ndimage as ndi
import torch
import torch.nn as nn
from PIL import Image
from huggingface_hub import HfApi, hf_hub_download
from skimage import measure
from skimage.feature import peak_local_max
from skimage.morphology import disk, opening, remove_small_objects
from skimage.segmentation import watershed
from torch.utils.data import DataLoader, Dataset, random_split
try:
from apscheduler.schedulers.blocking import BlockingScheduler
HAS_SCHEDULER = True
except ImportError:
HAS_SCHEDULER = False
# ─────────────────────────────────────────────────────────────────────────────
# Configuration
# ─────────────────────────────────────────────────────────────────────────────
ROOT = Path(__file__).parent
HF_REPO_ID = os.environ.get("HF_REPO_ID", "skarugu/myotube-unet")
HF_FILENAME = os.environ.get("HF_FILENAME", "model_final.pt")
HF_TOKEN = os.environ.get("HF_TOKEN", None)
DATA_ROOT = os.environ.get("DATA_ROOT", str(ROOT / "data"))
BATCH_TRIGGER_N = int(os.environ.get("BATCH_TRIGGER_N", 20))
CONF_DROP_K = int(os.environ.get("CONF_DROP_K", 5))
CONF_FLAG_THR = float(os.environ.get("CONF_FLAG_THR", 0.60))
SCHEDULE_HOUR = int(os.environ.get("SCHEDULE_HOUR", 2))
FT_EPOCHS = int(os.environ.get("FT_EPOCHS", 10))
FT_LR = float(os.environ.get("FT_LR", 5e-4))
FT_BATCH_SIZE = int(os.environ.get("FT_BATCH_SIZE", 4))
IMAGE_SIZE = int(os.environ.get("IMAGE_SIZE", 512))
QUEUE_DIR = ROOT / "retrain_queue"
CORRECTIONS_DIR = ROOT / "corrections"
RUNS_DIR = ROOT / "runs"
STATE_PATH = ROOT / "self_train_state.json"
MANIFEST_PATH = ROOT / "manifest.json"
logging.basicConfig(
level=logging.INFO,
format="%(asctime)s [%(levelname)s] %(message)s",
datefmt="%Y-%m-%d %H:%M:%S",
)
log = logging.getLogger("self_train")
# ─────────────────────────────────────────────────────────────────────────────
# State helpers
# ─────────────────────────────────────────────────────────────────────────────
def _load_state() -> dict:
if STATE_PATH.exists():
return json.loads(STATE_PATH.read_text())
return {"best_dice": 0.0, "last_retrain_ts": None, "current_hf_sha": None}
def _save_state(s: dict): STATE_PATH.write_text(json.dumps(s, indent=2))
def _load_manifest() -> list:
return json.loads(MANIFEST_PATH.read_text()) if MANIFEST_PATH.exists() else []
def _save_manifest(m: list): MANIFEST_PATH.write_text(json.dumps(m, indent=2, default=str))
def _aggregate_user_params(corrections_dir: Path, run_dir: Path):
"""
Scan corrections for user_optimized_params submissions and aggregate
their parameter settings. Writes optimal_params.json to run_dir with
median values β€” useful for tuning defaults.
"""
all_params = []
if not corrections_dir.exists():
return
for meta_p in corrections_dir.glob("*/meta.json"):
try:
meta = json.loads(meta_p.read_text())
if meta.get("reason") == "user_optimized_params" and "parameters" in meta:
all_params.append(meta["parameters"])
except Exception:
continue
if not all_params:
return
# Compute median for each numeric parameter
aggregated = {}
for key in all_params[0]:
vals = [p[key] for p in all_params if key in p and isinstance(p[key], (int, float))]
if vals:
vals.sort()
mid = len(vals) // 2
aggregated[key] = vals[mid] if len(vals) % 2 else (vals[mid-1] + vals[mid]) / 2
result = {
"n_submissions": len(all_params),
"aggregated_params": aggregated,
"all_submissions": all_params,
}
out = run_dir / "optimal_params.json"
out.write_text(json.dumps(result, indent=2))
log.info("Aggregated %d user param submissions β†’ %s", len(all_params), out)
# ─────────────────────────────────────────────────────────────────────────────
# Trigger checks
# ─────────────────────────────────────────────────────────────────────────────
def should_retrain(force=False):
if force:
return True, "manual"
corrections = list(CORRECTIONS_DIR.glob("*/meta.json")) if CORRECTIONS_DIR.exists() else []
if corrections:
return True, f"user_correction ({len(corrections)} pairs)"
q_jsons = list(QUEUE_DIR.glob("*.json")) if QUEUE_DIR.exists() else []
if len(q_jsons) >= BATCH_TRIGGER_N:
return True, f"batch_trigger ({len(q_jsons)} images)"
low_conf = sum(
1 for jf in q_jsons
if json.loads(jf.read_text()).get("reason") == "low_confidence"
) if q_jsons else 0
if low_conf >= CONF_DROP_K:
return True, f"confidence_drop ({low_conf} low-conf images)"
return False, "none"
# ─────────────────────────────────────────────────────────────────────────────
# Model definition (must be identical to the training script)
# ─────────────────────────────────────────────────────────────────────────────
class DoubleConv(nn.Module):
def __init__(self, in_ch, out_ch):
super().__init__()
self.net = nn.Sequential(
nn.Conv2d(in_ch, out_ch, 3, padding=1), nn.BatchNorm2d(out_ch), nn.ReLU(True),
nn.Conv2d(out_ch, out_ch, 3, padding=1), nn.BatchNorm2d(out_ch), nn.ReLU(True),
)
def forward(self, x): return self.net(x)
class UNet(nn.Module):
def __init__(self, in_ch=2, out_ch=2, base=32):
super().__init__()
self.d1=DoubleConv(in_ch,base); self.p1=nn.MaxPool2d(2)
self.d2=DoubleConv(base,base*2); self.p2=nn.MaxPool2d(2)
self.d3=DoubleConv(base*2,base*4); self.p3=nn.MaxPool2d(2)
self.d4=DoubleConv(base*4,base*8); self.p4=nn.MaxPool2d(2)
self.bn=DoubleConv(base*8,base*16)
self.u4=nn.ConvTranspose2d(base*16,base*8,2,2); self.du4=DoubleConv(base*16,base*8)
self.u3=nn.ConvTranspose2d(base*8,base*4,2,2); self.du3=DoubleConv(base*8,base*4)
self.u2=nn.ConvTranspose2d(base*4,base*2,2,2); self.du2=DoubleConv(base*4,base*2)
self.u1=nn.ConvTranspose2d(base*2,base,2,2); self.du1=DoubleConv(base*2,base)
self.out=nn.Conv2d(base,out_ch,1)
def forward(self, x):
d1=self.d1(x); p1=self.p1(d1)
d2=self.d2(p1); p2=self.p2(d2)
d3=self.d3(p2); p3=self.p3(d3)
d4=self.d4(p3); p4=self.p4(d4)
b=self.bn(p4)
x=self.u4(b); x=torch.cat([x,d4],1); x=self.du4(x)
x=self.u3(x); x=torch.cat([x,d3],1); x=self.du3(x)
x=self.u2(x); x=torch.cat([x,d2],1); x=self.du2(x)
x=self.u1(x); x=torch.cat([x,d1],1); x=self.du1(x)
return self.out(x)
# ─────────────────────────────────────────────────────────────────────────────
# Minimal Dataset for fine-tuning
# ─────────────────────────────────────────────────────────────────────────────
class _FTDataset(Dataset):
IMG_EXTS = {".jpg", ".jpeg", ".png", ".tif", ".tiff"}
def __init__(self, root, size=512, augment=True):
root = Path(root)
img_dir = root / "images"
nuc_dir = root / "masks" / "Nuclei_m"
myo_dir = root / "masks" / "Myotubes_m"
imgs = sorted([p for p in img_dir.glob("*") if p.suffix.lower() in self.IMG_EXTS])
self.samples = []
for p in imgs:
nuc = self._mp(nuc_dir, p.stem)
myo = self._mp(myo_dir, p.stem)
if nuc and myo:
self.samples.append((p, nuc, myo))
if not self.samples:
raise FileNotFoundError(f"No labelled samples found under {root}")
self.size = size
self.augment = augment
@staticmethod
def _mp(d, stem):
for ext in (".tif", ".tiff", ".png"):
p = d / f"{stem}{ext}"
if p.exists(): return p
return None
def __len__(self): return len(self.samples)
def __getitem__(self, idx):
ip, nuc_path, mp = self.samples[idx]
rgb = np.array(Image.open(ip).convert("RGB"), dtype=np.uint8)
H = W = self.size
def _ch(arr): return np.array(Image.fromarray(arr, "L").resize((W, H), Image.BILINEAR), dtype=np.float32) / 255.0
def _mk(p): return (np.array(Image.open(p).convert("L").resize((W, H), Image.NEAREST)) > 0).astype(np.uint8)
red = _ch(rgb[..., 0])
blue = _ch(rgb[..., 2])
yn = _mk(nuc_path)
ym = _mk(mp)
if self.augment:
f = np.stack([red, blue, np.zeros_like(red)], -1).astype(np.float32)
for ax in [1, 0]:
if random.random() < 0.5:
f = np.flip(f, ax); yn = np.flip(yn, ax); ym = np.flip(ym, ax)
k = random.randint(0, 3)
if k: f = np.rot90(f, k); yn = np.rot90(yn, k); ym = np.rot90(ym, k)
red, blue = f[..., 0], f[..., 1]
x = np.stack([red, blue], 0).astype(np.float32)
y = np.stack([yn, ym], 0).astype(np.float32)
return torch.from_numpy(x.copy()), torch.from_numpy(y.copy()), ip.stem
# ─────────────────────────────────────────────────────────────────────────────
# Loss + Dice
# ─────────────────────────────────────────────────────────────────────────────
class _BCEDice(nn.Module):
def __init__(self):
super().__init__()
self.bce = nn.BCEWithLogitsLoss()
def forward(self, logits, target):
bce = self.bce(logits, target)
p = torch.sigmoid(logits)
inter = (p * target).sum(dim=(2,3))
union = p.sum(dim=(2,3)) + target.sum(dim=(2,3))
dice = 1 - (2*inter+1e-6)/(union+1e-6)
return 0.5*bce + 0.5*dice.mean()
@torch.no_grad()
def _dice(probs, target, thr=0.5):
pred = (probs > thr).float()
inter = (pred * target).sum(dim=(2,3))
union = pred.sum(dim=(2,3)) + target.sum(dim=(2,3))
return ((2*inter+1e-6)/(union+1e-6)).mean(dim=0)
# ─────────────────────────────────────────────────────────────────────────────
# Prepare fine-tune data (base + corrections merged into a temp folder)
# ─────────────────────────────────────────────────────────────────────────────
def _prepare_data(base: str) -> str:
tmp = Path(tempfile.mkdtemp()) / "ft"
orig = Path(base)
if (orig / "images").exists():
shutil.copytree(str(orig), str(tmp), dirs_exist_ok=True)
else:
for sub in ("images", "masks/Nuclei_m", "masks/Myotubes_m"):
(tmp / sub).mkdir(parents=True, exist_ok=True)
log.warning("DATA_ROOT %s has no images/ β€” training on corrections only.", orig)
injected = 0
if CORRECTIONS_DIR.exists():
for meta_p in CORRECTIONS_DIR.glob("*/meta.json"):
folder = meta_p.parent
img, nuc, myo = folder/"image.png", folder/"nuclei_mask.png", folder/"myotube_mask.png"
if not (img.exists() and nuc.exists() and myo.exists()):
continue
stem = folder.name
shutil.copy(img, tmp/"images"/f"{stem}.png")
shutil.copy(nuc, tmp/"masks"/"Nuclei_m"/f"{stem}.png")
shutil.copy(myo, tmp/"masks"/"Myotubes_m"/f"{stem}.png")
injected += 1
log.info("Fine-tune data ready: %d correction(s) injected β†’ %s", injected, tmp)
return str(tmp)
# ─────────────────────────────────────────────────────────────────────────────
# HF Hub helpers
# ─────────────────────────────────────────────────────────────────────────────
def _load_from_hub():
path = hf_hub_download(repo_id=HF_REPO_ID, filename=HF_FILENAME,
token=HF_TOKEN, force_download=True)
ckpt = torch.load(path, map_location="cpu")
state = ckpt["model"] if isinstance(ckpt, dict) and "model" in ckpt else ckpt
model = UNet(in_ch=2, out_ch=2, base=32)
model.load_state_dict(state)
log.info("Loaded model from Hub (repo=%s, file=%s)", HF_REPO_ID, HF_FILENAME)
return model
def _push_to_hub(model_path: Path, metrics: dict, run_id: str) -> bool:
if not HF_TOKEN:
log.warning("HF_TOKEN not set β€” skipping Hub push.")
return False
api = HfApi(token=HF_TOKEN)
api.upload_file(
path_or_fileobj=str(model_path),
path_in_repo=HF_FILENAME,
repo_id=HF_REPO_ID,
repo_type="model",
commit_message=(f"Auto-retrain {run_id} | "
f"dice_nuc={metrics['dice_nuc']:.3f} "
f"dice_myo={metrics['dice_myo']:.3f}"),
)
api.upload_file(
path_or_fileobj=json.dumps({**metrics, "run_id": run_id,
"timestamp": datetime.now().isoformat()},
indent=2).encode(),
path_in_repo="auto_retrain_metrics.json",
repo_id=HF_REPO_ID,
repo_type="model",
commit_message=f"Metrics for auto-retrain {run_id}",
)
log.info("βœ… Pushed new weights to %s/%s", HF_REPO_ID, HF_FILENAME)
return True
# ─────────────────────────────────────────────────────────────────────────────
# Core retrain
# ─────────────────────────────────────────────────────────────────────────────
def run_retrain(reason: str = "scheduled"):
random.seed(42); np.random.seed(42); torch.manual_seed(42)
device = "cuda" if torch.cuda.is_available() else "cpu"
run_id = datetime.now().strftime("%Y%m%d_%H%M%S")
run_dir = RUNS_DIR / run_id
run_dir.mkdir(parents=True, exist_ok=True)
log.info("══ Retrain run %s | reason=%s | device=%s ══", run_id, reason, device)
ft_data = _prepare_data(DATA_ROOT)
try:
ds = _FTDataset(ft_data, size=IMAGE_SIZE, augment=True)
except FileNotFoundError as e:
log.error("No data: %s β€” aborting.", e)
return None
n_val = max(1, int(len(ds) * 0.2))
n_train = len(ds) - n_val
if n_train < 1:
log.warning("Only %d samples β€” need β‰₯2. Aborting.", len(ds))
return None
train_ds, val_ds = random_split(
ds, [n_train, n_val], generator=torch.Generator().manual_seed(42)
)
val_ds.dataset.augment = False
train_dl = DataLoader(train_ds, batch_size=FT_BATCH_SIZE, shuffle=True, num_workers=0)
val_dl = DataLoader(val_ds, batch_size=1, shuffle=False, num_workers=0)
model = _load_from_hub().to(device)
loss_fn = _BCEDice()
opt = torch.optim.Adam(model.parameters(), lr=FT_LR)
sched = torch.optim.lr_scheduler.CosineAnnealingLR(opt, T_max=FT_EPOCHS, eta_min=1e-5)
state = _load_state()
prev_best = state.get("best_dice", 0.0)
best_run_dice = -1.0
best_path = run_dir / "model_best.pt"
for ep in range(1, FT_EPOCHS + 1):
model.train()
for x, y, _ in train_dl:
x, y = x.to(device), y.to(device)
opt.zero_grad(); loss_fn(model(x), y).backward(); opt.step()
sched.step()
model.eval()
dices = []
with torch.no_grad():
for x, y, _ in val_dl:
probs = torch.sigmoid(model(x.to(device))).cpu()
dices.append(_dice(probs, y).numpy())
d = np.array(dices)
d_nuc, d_myo = float(d[:,0].mean()), float(d[:,1].mean())
score = (d_nuc + d_myo) / 2.0
log.info(" Ep %02d | dice_nuc=%.3f | dice_myo=%.3f | mean=%.3f", ep, d_nuc, d_myo, score)
if score > best_run_dice:
best_run_dice = score
torch.save({"model": model.state_dict()}, best_path)
metrics = {
"dice_nuc": round(d_nuc, 4),
"dice_myo": round(d_myo, 4),
"mean_dice": round(best_run_dice, 4),
"reason": reason,
"n_train": n_train,
"n_val": n_val,
}
pushed = False
log.info("Best this run: %.4f | Previous best: %.4f", best_run_dice, prev_best)
if best_run_dice > prev_best:
pushed = _push_to_hub(best_path, metrics, run_id)
state["best_dice"] = best_run_dice
state["current_hf_sha"] = str(best_path)
else:
log.info("New model did not beat previous best β€” NOT pushing.")
# Archive queue
archive = run_dir / "processed_queue"
archive.mkdir(parents=True, exist_ok=True)
# Before archiving, collect user-submitted parameter snapshots
_aggregate_user_params(CORRECTIONS_DIR, run_dir)
for p in list(QUEUE_DIR.glob("*")) if QUEUE_DIR.exists() else []:
shutil.move(str(p), str(archive / p.name))
for folder in list(CORRECTIONS_DIR.glob("*")) if CORRECTIONS_DIR.exists() else []:
if folder.is_dir():
shutil.move(str(folder), str(archive / folder.name))
state["last_retrain_ts"] = datetime.now().isoformat()
_save_state(state)
manifest = _load_manifest()
manifest.append({"run_id": run_id, "timestamp": state["last_retrain_ts"],
"reason": reason, "metrics": metrics, "pushed": pushed})
_save_manifest(manifest)
log.info("══ Run %s complete | pushed=%s ══", run_id, pushed)
return metrics
# ─────────────────────────────────────────────────────────────────────────────
# Trigger check entry point
# ─────────────────────────────────────────────────────────────────────────────
def check_and_retrain(force=False):
ok, reason = should_retrain(force=force)
if ok:
log.info("Trigger met: %s β†’ retraining…", reason)
run_retrain(reason=reason)
else:
log.info("No trigger met β€” skipping.")
# ─────────────────────────────────────────────────────────────────────────────
# Scheduler
# ─────────────────────────────────────────────────────────────────────────────
def start_scheduler():
if not HAS_SCHEDULER:
log.error("APScheduler not installed. pip install apscheduler")
return
s = BlockingScheduler(timezone="UTC")
s.add_job(lambda: check_and_retrain(force=True),
"cron", hour=SCHEDULE_HOUR, minute=0, id="nightly")
s.add_job(check_and_retrain, "interval", minutes=30, id="poll")
log.info("Scheduler running. Nightly at %02d:00 UTC. Polling every 30 min.", SCHEDULE_HOUR)
try:
s.start()
except (KeyboardInterrupt, SystemExit):
log.info("Scheduler stopped.")
# ─────────────────────────────────────────────────────────────────────────────
# CLI
# ─────────────────────────────────────────────────────────────────────────────
if __name__ == "__main__":
ap = argparse.ArgumentParser()
ap.add_argument("--manual", action="store_true", help="Force retrain now")
ap.add_argument("--scheduler", action="store_true", help="Start blocking scheduler")
ap.add_argument("--data_root", default=None, help="Override DATA_ROOT env var")
a = ap.parse_args()
if a.data_root:
DATA_ROOT = a.data_root
if a.scheduler:
start_scheduler()
else:
check_and_retrain(force=a.manual)