reworked iop_corr, added explainability tools and plotting tools, cleanup codebase
This commit is contained in:
rpotter6298
@@ -1,105 +0,0 @@
|
||||
# bridge.py
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
from classes.SE_attention import SEBlock, SEGateLogger
|
||||
|
||||
# class SEBlock(nn.Module):
|
||||
# def __init__(self, dim: int, reduction: int = 16):
|
||||
# super().__init__()
|
||||
# hidden = max(1, dim // max(1, reduction))
|
||||
# self.net = nn.Sequential(
|
||||
# nn.Linear(dim, hidden, bias=True),
|
||||
# nn.ReLU(inplace=True),
|
||||
# nn.Linear(hidden, dim, bias=True),
|
||||
# nn.Sigmoid(),
|
||||
# )
|
||||
|
||||
# def forward(self, x):
|
||||
# return self.net(x)
|
||||
|
||||
class Bridge(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
img_dim,
|
||||
meta_dim,
|
||||
num_classes,
|
||||
fusion_dim=256,
|
||||
mode="fused",
|
||||
use_se: bool = True,
|
||||
se_reduction: int = 16,
|
||||
se_pre_norm: bool = True,
|
||||
):
|
||||
|
||||
super().__init__()
|
||||
self.mode = mode
|
||||
self.use_se = use_se
|
||||
# self.se_reduction = se_reduction
|
||||
# self.se_pre_norm = se_pre_norm
|
||||
|
||||
#project towers to equal width
|
||||
self.W_img = nn.Linear(img_dim, fusion_dim)
|
||||
self.W_md = nn.Linear(meta_dim, fusion_dim)
|
||||
|
||||
#(optional) : set layernorm for se so one tower doesn't dominate the other
|
||||
self.ln_img = nn.LayerNorm(fusion_dim) if se_pre_norm else nn.Identity()
|
||||
self.ln_md = nn.LayerNorm(fusion_dim) if se_pre_norm else nn.Identity()
|
||||
|
||||
#SE gate on the fused vector
|
||||
self.se = SEBlock(fusion_dim, reduction=se_reduction, residual=True) if use_se else None
|
||||
self.se_log = SEGateLogger(enabled=use_se, track_channels=False, dim=fusion_dim)
|
||||
|
||||
|
||||
#heads
|
||||
self.classifier_fused = nn.Sequential(
|
||||
nn.ReLU(), nn.Dropout(0.5), nn.Linear(fusion_dim, num_classes)
|
||||
)
|
||||
self.classifier_img = nn.Linear(img_dim, num_classes)
|
||||
self.classifier_md = nn.Linear(meta_dim, num_classes)
|
||||
def reset_se_stats(self):
|
||||
"""Call at epoch start."""
|
||||
if getattr(self, "se_log", None):
|
||||
self.se_log.reset()
|
||||
|
||||
def get_se_stats(self, reset: bool = True):
|
||||
"""Call after eval. Returns dict or None."""
|
||||
if getattr(self, "se_log", None) and self.se_log.enabled:
|
||||
return self.se_log.get(reset=reset)
|
||||
return None
|
||||
|
||||
def forward(self, img_feats, md_feats):
|
||||
out_img = None if self.mode == "metadata_only" else self.classifier_img(img_feats)
|
||||
out_md = None if self.mode == "image_only" else self.classifier_md(md_feats)
|
||||
|
||||
if self.mode == "fused":
|
||||
hi = self.ln_img(self.W_img(img_feats)) #image features
|
||||
hm = self.ln_md(self.W_md(md_feats)) #metadata features
|
||||
fused = hi * hm #elementwise product
|
||||
#apply SE gates
|
||||
if self.se is not None:
|
||||
fused, gates = self.se(fused)
|
||||
if self.se_log.enabled:
|
||||
self.se_log.accumulate(gates)
|
||||
|
||||
if self.se is not None and self.training and self.se_log.enabled:
|
||||
if not hasattr(self, "_dbg_seen"):
|
||||
self._dbg_seen = 0
|
||||
if self._dbg_seen < 3: # print only a few times
|
||||
print("[SE] gate mean this batch:", gates.mean().item())
|
||||
self._dbg_seen += 1
|
||||
out_f = self.classifier_fused(fused)
|
||||
return out_f, out_img, out_md
|
||||
# if ablation modes:
|
||||
if self.mode == "image_only":
|
||||
return out_img, out_img, None
|
||||
if self.mode == "metadata_only":
|
||||
return out_md, None, out_md
|
||||
|
||||
|
||||
class VoteBridge(nn.Module):
|
||||
def __init__(self, num_classes):
|
||||
super().__init__()
|
||||
self.vote_combiner = nn.Linear(num_classes * 2, num_classes) # two sets of logits
|
||||
|
||||
def forward(self, out_img, out_md):
|
||||
votes = torch.cat([out_img, out_md], dim=1)
|
||||
return self.vote_combiner(votes)
|
||||
@@ -1,264 +0,0 @@
|
||||
# clinical_data.py
|
||||
from __future__ import annotations
|
||||
from pathlib import Path
|
||||
from typing import Iterable, Optional, Dict, List, Tuple
|
||||
import numpy as np
|
||||
import pandas as pd
|
||||
|
||||
class ClinicalData:
|
||||
"""
|
||||
Torch-free container for clinical metadata and file/label bookkeeping.
|
||||
- Holds one or more dataframes (via add_df) and harmonizes columns
|
||||
- Canonical IDs: 'Patient ID' must exist (or be specified and will be renamed)
|
||||
- Canonical eye column: 'eyeID' recoded to 'OS'/'OD' if present; if absent, set to 0
|
||||
- Feature typing (if cat_cols not provided):
|
||||
* Categorical if (a) <= max_unique categorical threshold (default 4), or
|
||||
(b) values cannot be coerced to float; otherwise numeric (scalar)
|
||||
- Scaling/imputation:
|
||||
* Numeric: min–max to [0,1], median imputation; + one missing flag per numeric feature
|
||||
* Categorical: one-hot with '<UNK>' bucket at index 0
|
||||
- Patient-level K-fold indices stored as dict: folds[k] -> {'train_ids': [...], 'test_ids': [...]}
|
||||
"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
image_dir: str,
|
||||
clinical_dir: Optional[str],
|
||||
label_col: str,
|
||||
# typing / detection
|
||||
cat_cols: Optional[Iterable[str]] = None,
|
||||
max_unique_for_cat: int = 4,
|
||||
# splitting
|
||||
n_splits: int = 5,
|
||||
random_seed: int = 42,
|
||||
):
|
||||
self.image_dir = Path(image_dir)
|
||||
self.clinical_dir = Path(clinical_dir) if clinical_dir else None
|
||||
self.label_col = label_col
|
||||
self.max_unique_for_cat = max_unique_for_cat
|
||||
self.n_splits = n_splits
|
||||
|
||||
# Internal state
|
||||
self.frames: List[pd.DataFrame] = [] # raw frames as added
|
||||
self.df: pd.DataFrame = pd.DataFrame() # concatenated
|
||||
self.scalar_cols: List[str] = []
|
||||
self.cat_cols: List[str] = list(cat_cols) if cat_cols is not None else []
|
||||
self.scalar_stats: Dict[str, Dict[str, float]] = {}
|
||||
self.cat_maps: Dict[str, Dict[object, int]] = {}
|
||||
self.feature_dim: int = 0
|
||||
self.folds: Dict[int, Dict[str, List[object]]] = {} # fold -> {'train_ids': [], 'test_ids': []}
|
||||
self.random_seed = int(random_seed)
|
||||
|
||||
# ------------------- Public API -------------------
|
||||
def add_df(
|
||||
self,
|
||||
df: pd.DataFrame,
|
||||
id_column: Optional[str] = None,
|
||||
eye_column: Optional[str] = None,
|
||||
exclude_cols: Optional[Iterable[str]] = None,
|
||||
) -> None:
|
||||
"""
|
||||
Add a dataframe and re-run harmonization, typing, stats, and K-fold indices.
|
||||
QC rules:
|
||||
- Must have patient ID column; if not provided under that name, specify id_column.
|
||||
- eyeID, if present, must be binary; recoded to 'OS'/'OD'. If absent, create and set to 0.
|
||||
"""
|
||||
df = df.copy()
|
||||
# --- QC: Patient ID ---
|
||||
pid_col = self._ensure_patient_id(df, id_column)
|
||||
# --- QC: eyeID ---
|
||||
self._canonicalize_eye_column(df, eye_column)
|
||||
# --- Normalize label presence ---
|
||||
if self.label_col not in df.columns:
|
||||
raise ValueError(f"label_col '{self.label_col}' not found in added dataframe")
|
||||
|
||||
# append & refresh
|
||||
self.frames.append(df)
|
||||
self._refresh_master_df(exclude_cols=exclude_cols)
|
||||
self._infer_or_validate_feature_types(exclude_cols=exclude_cols)
|
||||
self._compute_numeric_stats()
|
||||
self._build_cat_maps()
|
||||
self._compute_feature_dim()
|
||||
self._build_kfold_indices()
|
||||
|
||||
def get_split_ids(self, fold: int) -> Tuple[List[object], List[object]]:
|
||||
rec = self.folds.get(fold)
|
||||
if not rec: raise KeyError(f"Fold {fold} not available. Built folds: {sorted(self.folds.keys())}")
|
||||
return rec['train_ids'], rec['test_ids']
|
||||
|
||||
def get_split_dfs(self, fold: int) -> Tuple[pd.DataFrame, pd.DataFrame]:
|
||||
train_ids, test_ids = self.get_split_ids(fold)
|
||||
train_df = self.df[self.df['Patient ID'].isin(train_ids)].reset_index(drop=True)
|
||||
test_df = self.df[self.df['Patient ID'].isin(test_ids)].reset_index(drop=True)
|
||||
return train_df, test_df
|
||||
|
||||
def vectorize_row(self, row: pd.Series) -> np.ndarray:
|
||||
"""Return a numpy feature vector (torch-free)."""
|
||||
feats: List[float] = []
|
||||
miss: List[float] = []
|
||||
# numeric
|
||||
for col in self.scalar_cols:
|
||||
v = pd.to_numeric(row.get(col), errors='coerce')
|
||||
if pd.isna(v):
|
||||
miss.append(1.0)
|
||||
v = self.scalar_stats[col]['median']
|
||||
else:
|
||||
miss.append(0.0)
|
||||
lo = self.scalar_stats[col]['min']; hi = self.scalar_stats[col]['max']
|
||||
feats.append((float(v) - lo) / (hi - lo) if hi > lo else 0.0)
|
||||
# categorical
|
||||
for col in self.cat_cols:
|
||||
mapping = self.cat_maps[col]
|
||||
one = [0.0] * len(mapping)
|
||||
key = row.get(col)
|
||||
one[mapping.get(key, 0)] = 1.0 # 0 is <UNK>
|
||||
feats.extend(one)
|
||||
# numeric missing flags
|
||||
feats.extend(miss)
|
||||
return np.asarray(feats, dtype=np.float32)
|
||||
|
||||
def get_image_path(self, row: pd.Series, filename_template: str = "RET{pid:03d}{eye}.jpg") -> Path:
|
||||
pid = int(row['Patient ID']); eye = row.get('eyeID', 0)
|
||||
if eye in ("OS", "OD"):
|
||||
eye_str = eye
|
||||
else:
|
||||
eye_str = str(eye)
|
||||
return self.image_dir / filename_template.format(pid=pid, eye=eye_str)
|
||||
|
||||
# ------------------- Internal helpers -------------------
|
||||
def _ensure_patient_id(self, df: pd.DataFrame, id_column: Optional[str]) -> str:
|
||||
if 'Patient ID' in df.columns:
|
||||
return 'Patient ID'
|
||||
if id_column and id_column in df.columns:
|
||||
df.rename(columns={id_column: 'Patient ID'}, inplace=True)
|
||||
return 'Patient ID'
|
||||
# try auto-detect common variants
|
||||
candidates = [c for c in df.columns if c.lower().replace(" ", "") in {"patientid","patient","pid"}]
|
||||
if len(candidates) == 1:
|
||||
df.rename(columns={candidates[0]: 'Patient ID'}, inplace=True)
|
||||
return 'Patient ID'
|
||||
raise ValueError("A 'Patient ID' column is required; provide id_column=... if it has a different name.")
|
||||
|
||||
def _canonicalize_eye_column(self, df: pd.DataFrame, eye_column: Optional[str]) -> None:
|
||||
# Find source
|
||||
src = None
|
||||
if 'eyeID' in df.columns: src = 'eyeID'
|
||||
elif eye_column and eye_column in df.columns: src = eye_column
|
||||
else:
|
||||
# try auto detect
|
||||
for c in df.columns:
|
||||
if 'eye' in c.lower():
|
||||
src = c; break
|
||||
if src is None:
|
||||
df['eyeID'] = 0
|
||||
return
|
||||
# Map to OS/OD
|
||||
s = df[src]
|
||||
def norm(v):
|
||||
if pd.isna(v): return None
|
||||
x = str(v).strip().upper()
|
||||
if x in {"OS","L","LEFT","0"}: return "OS"
|
||||
if x in {"OD","R","RIGHT","1"}: return "OD"
|
||||
# numbers like 2? fall back by parity
|
||||
try:
|
||||
num = int(float(x))
|
||||
return "OD" if num % 2 == 1 else "OS"
|
||||
except Exception:
|
||||
return None
|
||||
mapped = s.map(norm)
|
||||
uniq = {u for u in mapped.dropna().unique().tolist()}
|
||||
if not uniq.issubset({"OS","OD"}):
|
||||
raise ValueError(f"eyeID must be binary; found values {sorted(uniq)}")
|
||||
df['eyeID'] = mapped.fillna("OS")
|
||||
if src != 'eyeID':
|
||||
# keep original too if you want, but we standardize on 'eyeID'
|
||||
pass
|
||||
|
||||
def _refresh_master_df(self, exclude_cols: Optional[Iterable[str]] = None) -> None:
|
||||
self.df = pd.concat(self.frames, axis=0, ignore_index=True)
|
||||
# drop columns explicitly excluded
|
||||
if exclude_cols:
|
||||
self.df = self.df.drop(columns=[c for c in exclude_cols if c in self.df.columns])
|
||||
|
||||
def _infer_or_validate_feature_types(self, exclude_cols: Optional[Iterable[str]] = None) -> None:
|
||||
excluded = set(exclude_cols or []) | {self.label_col, 'Patient ID'}
|
||||
# we keep canonical 'eyeID' as categorical if present
|
||||
feature_candidates = [c for c in self.df.columns if c not in excluded]
|
||||
# If user pre-specified cat_cols in __init__, respect them and infer the rest
|
||||
cats = set(self.cat_cols) if self.cat_cols else set()
|
||||
scalars = set()
|
||||
for c in feature_candidates:
|
||||
if c == 'eyeID':
|
||||
cats.add('eyeID'); continue
|
||||
if c in cats: continue
|
||||
s = self.df[c]
|
||||
# try numeric coercion
|
||||
as_num = pd.to_numeric(s, errors='coerce')
|
||||
num_missing = as_num.isna().mean()
|
||||
num_unique = s.dropna().nunique()
|
||||
if as_num.notna().any() and num_missing < 1.0 and num_unique > self.max_unique_for_cat:
|
||||
scalars.add(c)
|
||||
else:
|
||||
# categorical if few uniques OR non-numeric
|
||||
if num_unique <= self.max_unique_for_cat or as_num.isna().mean() > 0.0:
|
||||
cats.add(c)
|
||||
else:
|
||||
scalars.add(c)
|
||||
self.cat_cols = sorted(cats)
|
||||
self.scalar_cols = sorted(scalars)
|
||||
|
||||
def _compute_numeric_stats(self) -> None:
|
||||
self.scalar_stats.clear()
|
||||
for col in self.scalar_cols:
|
||||
s = pd.to_numeric(self.df[col], errors='coerce')
|
||||
vals = s.dropna().astype(float).values
|
||||
if vals.size == 0:
|
||||
lo, hi, med = 0.0, 1.0, 0.0
|
||||
else:
|
||||
lo, hi = float(np.min(vals)), float(np.max(vals))
|
||||
med = float(np.median(vals))
|
||||
if hi <= lo: hi = lo + 1.0
|
||||
self.scalar_stats[col] = {"min": lo, "max": hi, "median": med}
|
||||
|
||||
def _build_cat_maps(self) -> None:
|
||||
self.cat_maps.clear()
|
||||
for col in self.cat_cols:
|
||||
cats = [v for v in self.df[col].dropna().unique().tolist()]
|
||||
try: cats = sorted(cats)
|
||||
except Exception: pass
|
||||
mapping = {"<UNK>": 0}
|
||||
for i, v in enumerate(cats, start=1): mapping[v] = i
|
||||
self.cat_maps[col] = mapping
|
||||
|
||||
def _compute_feature_dim(self) -> None:
|
||||
self.feature_dim = len(self.scalar_cols) + sum(len(m) for m in self.cat_maps.values()) + len(self.scalar_cols)
|
||||
|
||||
# ------------------- K-fold on unique patients -------------------
|
||||
def _build_kfold_indices(self) -> None:
|
||||
# unique patients and a per-patient label for stratification if possible
|
||||
pats = self.df['Patient ID'].unique().tolist()
|
||||
# Derive a patient label as the mode of their rows (fallback to first valid)
|
||||
labels_by_pat = {}
|
||||
for pid, grp in self.df.groupby('Patient ID'):
|
||||
lab = grp[self.label_col].dropna()
|
||||
if len(lab) == 0:
|
||||
labels_by_pat[pid] = 0
|
||||
else:
|
||||
labels_by_pat[pid] = lab.mode().iloc[0]
|
||||
y_pat = np.array([labels_by_pat[p] for p in pats])
|
||||
|
||||
# Try to use StratifiedGroupKFold if available, else fall back to StratifiedKFold on patient labels
|
||||
try:
|
||||
from sklearn.model_selection import StratifiedGroupKFold
|
||||
sgkf = StratifiedGroupKFold(n_splits=self.n_splits, shuffle=True, random_state=self.random_seed)
|
||||
split_iter = sgkf.split(X=pats, y=y_pat, groups=pats)
|
||||
except Exception:
|
||||
from sklearn.model_selection import StratifiedKFold
|
||||
skf = StratifiedKFold(n_splits=self.n_splits, shuffle=True, random_state=self.random_seed)
|
||||
split_iter = skf.split(X=np.zeros(len(pats)), y=y_pat)
|
||||
|
||||
self.folds.clear()
|
||||
for i, (train_idx, test_idx) in enumerate(split_iter):
|
||||
train_ids = [pats[j] for j in train_idx]
|
||||
test_ids = [pats[j] for j in test_idx]
|
||||
self.folds[i] = {"train_ids": train_ids, "test_ids": test_ids}
|
||||
@@ -1,58 +0,0 @@
|
||||
# dataset.py
|
||||
from torch.utils.data import Dataset
|
||||
from PIL import Image
|
||||
import numpy as np
|
||||
import torch
|
||||
|
||||
|
||||
class ClinicalDataset(Dataset):
|
||||
"""Generic dataset wrapping a ClinicalData instance.
|
||||
Returns (img_tensor, meta_tensor, label)."""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
clinical_data,
|
||||
img_transform,
|
||||
meta_transform=None,
|
||||
image_preprocessor=None,
|
||||
geometry_provider=None,
|
||||
geometry_dim: int = 0,
|
||||
):
|
||||
self.clinical = clinical_data
|
||||
self.transform_image = img_transform
|
||||
self.meta_transform = meta_transform or (lambda x: x)
|
||||
self.image_preprocessor = image_preprocessor
|
||||
self.geometry_provider = geometry_provider
|
||||
self.geometry_dim = geometry_dim if geometry_provider is not None else 0
|
||||
|
||||
def __len__(self):
|
||||
return len(self.clinical.df)
|
||||
|
||||
def __getitem__(self, idx: int):
|
||||
row = self.clinical.df.iloc[idx]
|
||||
# load & transform image
|
||||
img_path = self.clinical.get_image_path(row)
|
||||
orig_img = Image.open(img_path).convert("RGB")
|
||||
img = orig_img
|
||||
if self.image_preprocessor is not None:
|
||||
img = self.image_preprocessor(img, img_path)
|
||||
img_t = self.transform_image(img)
|
||||
# encode & transform metadata
|
||||
meta = self.clinical.encode_metadata(row)
|
||||
meta_t = self.meta_transform(meta)
|
||||
# label
|
||||
label = self.clinical.get_label(row)
|
||||
if self.geometry_dim > 0:
|
||||
features = None
|
||||
if self.geometry_provider is not None and hasattr(self.geometry_provider, "geometry_features"):
|
||||
features = self.geometry_provider.geometry_features(orig_img, img_path)
|
||||
if features is None:
|
||||
geom_vec = torch.zeros(self.geometry_dim, dtype=torch.float32)
|
||||
else:
|
||||
features = np.asarray(features, dtype=np.float32)
|
||||
if features.shape[0] != self.geometry_dim:
|
||||
geom_vec = torch.zeros(self.geometry_dim, dtype=torch.float32)
|
||||
else:
|
||||
geom_vec = torch.from_numpy(features)
|
||||
return img_t, meta_t, geom_vec, label
|
||||
return img_t, meta_t, label
|
||||
@@ -1,89 +0,0 @@
|
||||
import math, copy, torch
|
||||
|
||||
class EarlyStopper:
|
||||
def __init__(self, monitor: str, mode: str = "auto",
|
||||
patience: int = 5, min_delta: float = 0.0,
|
||||
save_path: str | None = None, restore_best: bool = True):
|
||||
"""
|
||||
monitor: key in your epoch row, e.g. 'eval_loss', 'auc_fused', 'acc_fused'
|
||||
mode: 'max' (higher is better), 'min', or 'auto' (min for '*loss*', else max)
|
||||
patience: epochs without improvement before stopping
|
||||
min_delta: required improvement magnitude
|
||||
save_path: optional .pth file to save best weights each time it improves
|
||||
restore_best: if True, load best weights back at the end
|
||||
"""
|
||||
self.monitor = monitor
|
||||
if mode == "auto":
|
||||
mode = "min" if "loss" in monitor.lower() else "max"
|
||||
self.mode = mode
|
||||
self.patience = int(patience)
|
||||
self.min_delta = float(min_delta)
|
||||
self.save_path = save_path
|
||||
self.restore_best = restore_best
|
||||
|
||||
self.best = -math.inf if mode == "max" else math.inf
|
||||
self.bad_epochs = 0
|
||||
self.best_state = None
|
||||
self.best_epoch = -1
|
||||
self.last_improved = False
|
||||
|
||||
def _is_better(self, val):
|
||||
if val is None or (isinstance(val, float) and math.isnan(val)):
|
||||
return False
|
||||
if self.mode == "max":
|
||||
return val > (self.best + self.min_delta)
|
||||
else:
|
||||
return val < (self.best - self.min_delta)
|
||||
|
||||
def step(self, metrics: dict, trainer, epoch: int) -> bool:
|
||||
val = metrics.get(self.monitor, None)
|
||||
improved = self._is_better(val)
|
||||
self.last_improved = improved
|
||||
|
||||
if improved:
|
||||
self.best = val
|
||||
self.best_epoch = epoch
|
||||
self.bad_epochs = 0
|
||||
# snapshot + optional save
|
||||
state = {
|
||||
"img_tower": trainer.img_tower.state_dict(),
|
||||
"md_tower": trainer.md_tower.state_dict(),
|
||||
"optimizer": trainer.optimizer.state_dict(),
|
||||
}
|
||||
if hasattr(trainer, "bridge"): state["bridge"] = trainer.bridge.state_dict()
|
||||
if hasattr(trainer, "head_img"): state["head_img"] = trainer.head_img.state_dict()
|
||||
if hasattr(trainer, "head_md"): state["head_md"] = trainer.head_md.state_dict()
|
||||
# keep an in-memory copy for restore(); file save is optional
|
||||
self.best_state = copy.deepcopy(state)
|
||||
if self.save_path: torch.save(state, self.save_path)
|
||||
print(f"[early] ↑ new best {self.monitor}={val:.5f} at epoch {epoch+1}")
|
||||
else:
|
||||
self.bad_epochs += 1
|
||||
|
||||
stop = self.bad_epochs >= self.patience
|
||||
if stop:
|
||||
print(f"[early] stopping: no improvement in {self.patience} epochs "
|
||||
f"(best {self.monitor}={self.best:.5f} @ epoch {self.best_epoch+1})")
|
||||
return stop
|
||||
|
||||
def restore(self, trainer):
|
||||
if not self.restore_best:
|
||||
return
|
||||
# Prefer in-memory best state; otherwise try loading from save_path
|
||||
st = self.best_state
|
||||
if st is None and self.save_path:
|
||||
try:
|
||||
st = torch.load(self.save_path, map_location="cpu")
|
||||
except Exception:
|
||||
st = None
|
||||
if st is None:
|
||||
return
|
||||
trainer.img_tower.load_state_dict(st["img_tower"])
|
||||
trainer.md_tower.load_state_dict(st["md_tower"])
|
||||
if "bridge" in st and hasattr(trainer, "bridge"):
|
||||
trainer.bridge.load_state_dict(st["bridge"])
|
||||
if "head_img" in st and hasattr(trainer, "head_img"):
|
||||
trainer.head_img.load_state_dict(st["head_img"])
|
||||
if "head_md" in st and hasattr(trainer, "head_md"):
|
||||
trainer.head_md.load_state_dict(st["head_md"])
|
||||
trainer.optimizer.load_state_dict(st["optimizer"])
|
||||
-1402
File diff suppressed because it is too large
Load Diff
File diff suppressed because it is too large
Load Diff
@@ -1,125 +0,0 @@
|
||||
# classes/image_tower.py
|
||||
from __future__ import annotations
|
||||
import math
|
||||
from typing import Optional
|
||||
|
||||
import torch
|
||||
from torch import nn
|
||||
from torchvision import transforms
|
||||
from classes.backbones import BACKBONES, list_names, load_backbone_weights
|
||||
from classes.SE_attention import SEBlock
|
||||
|
||||
def build_backbone(name: str, freeze_ratio: float = 0.0, augment: bool = True):
|
||||
"""
|
||||
Operational builder:
|
||||
- instantiate with DEFAULT weights
|
||||
- strip classifier → features
|
||||
- apply ratio-based freezing over coarse blocks
|
||||
- return (model, out_dim, transform)
|
||||
"""
|
||||
key = (name or "").lower()
|
||||
if key not in BACKBONES:
|
||||
raise ValueError(f"Unsupported backbone '{name}'. Valid options: {list_names()}")
|
||||
|
||||
spec = BACKBONES[key]
|
||||
m = spec.ctor(weights=spec.weights_default)
|
||||
out_dim, m = spec.strip(m)
|
||||
load_backbone_weights(key, m)
|
||||
|
||||
# transforms: use the weights’ mean/std, but keep your augmentation pipeline
|
||||
mean = getattr(spec.weights_default, "meta", {}).get("mean", (0.485, 0.456, 0.406))
|
||||
std = getattr(spec.weights_default, "meta", {}).get("std", (0.229, 0.224, 0.225))
|
||||
crop = 299 if key == "inception_v3" else 224
|
||||
|
||||
if augment:
|
||||
transform = transforms.Compose([
|
||||
transforms.Resize(256),
|
||||
transforms.CenterCrop(crop),
|
||||
transforms.RandomHorizontalFlip(),
|
||||
transforms.RandomVerticalFlip(),
|
||||
transforms.RandomRotation(15),
|
||||
transforms.ColorJitter(0.1, 0.1, 0.1, 0.05),
|
||||
transforms.ToTensor(),
|
||||
transforms.Normalize(mean=mean, std=std),
|
||||
])
|
||||
else:
|
||||
transform = transforms.Compose([
|
||||
transforms.Resize(256),
|
||||
transforms.CenterCrop(crop),
|
||||
transforms.ToTensor(),
|
||||
transforms.Normalize(mean=mean, std=std),
|
||||
])
|
||||
|
||||
# ratio-based freezing: freeze earliest floor(N * freeze_ratio) blocks
|
||||
fr = max(0.0, min(1.0, float(freeze_ratio)))
|
||||
blocks = spec.blocks(m)
|
||||
n = len(blocks)
|
||||
freeze_n = int(math.floor(n * fr))
|
||||
for b in blocks[:freeze_n]:
|
||||
for p in b.parameters():
|
||||
p.requires_grad = False
|
||||
|
||||
return m, out_dim, transform
|
||||
|
||||
class ImageTower(nn.Module):
|
||||
"""
|
||||
Vision backbone → pooled features.
|
||||
- backbone: one of list_names() (default 'efficientnet_b0')
|
||||
- always DEFAULT torchvision weights
|
||||
- freeze_ratio ∈ [0,1] freezes earliest floor(N*freeze_ratio) blocks
|
||||
- returns [N, out_dim] features from backbone forward
|
||||
"""
|
||||
def __init__(self, backbone: str = "efficientnet_b0", freeze_ratio: float = 0.0,
|
||||
use_se: bool = False, se_reduction: int = 16, se_pre_norm: bool = True,
|
||||
augment: bool = True, geometry_dim: int = 0):
|
||||
super().__init__()
|
||||
self.backbone, base_dim, self.transform = build_backbone(backbone, freeze_ratio, augment=augment)
|
||||
self._name = backbone
|
||||
# Keep ordered blocks for dynamic freezing/thawing
|
||||
key = (self._name or "").lower()
|
||||
self._spec = BACKBONES[key]
|
||||
self._blocks = self._spec.blocks(self.backbone)
|
||||
# Optional tower-level SE over the final feature vector
|
||||
self.base_dim = base_dim
|
||||
self.geometry_dim = max(0, int(geometry_dim))
|
||||
self.out_dim = self.base_dim + self.geometry_dim
|
||||
self.tower_ln = nn.LayerNorm(self.base_dim) if se_pre_norm else nn.Identity()
|
||||
self.tower_se = SEBlock(self.base_dim, reduction=se_reduction, residual=True) if use_se else None
|
||||
|
||||
def forward(self, x: torch.Tensor, geometry: Optional[torch.Tensor] = None) -> torch.Tensor:
|
||||
y = self.backbone(x)
|
||||
# sanity: pooled features, not logits
|
||||
assert y.dim() == 2 and y.size(1) == self.base_dim, \
|
||||
f"Expected features [N,{self.base_dim}], got {tuple(y.shape)}"
|
||||
if self.tower_se is not None:
|
||||
y, _ = self.tower_se(self.tower_ln(y))
|
||||
if self.geometry_dim > 0:
|
||||
if geometry is None or geometry.numel() == 0:
|
||||
geom = torch.zeros(y.size(0), self.geometry_dim, device=y.device, dtype=y.dtype)
|
||||
else:
|
||||
if geometry.dim() == 1:
|
||||
geom = geometry.unsqueeze(0)
|
||||
else:
|
||||
geom = geometry
|
||||
geom = geom.to(device=y.device, dtype=y.dtype)
|
||||
if geom.size(0) != y.size(0):
|
||||
raise ValueError(f"Geometry batch size mismatch: {geom.size(0)} vs {y.size(0)}")
|
||||
if geom.size(1) != self.geometry_dim:
|
||||
raise ValueError(f"Expected geometry dim {self.geometry_dim}, got {geom.size(1)}")
|
||||
y = torch.cat([y, geom], dim=1)
|
||||
return y
|
||||
|
||||
def set_freeze_ratio(self, ratio: float):
|
||||
"""Dynamically freeze earliest floor(N*ratio) backbone blocks.
|
||||
ratio in [0,1]."""
|
||||
r = max(0.0, min(1.0, float(ratio)))
|
||||
n = len(self._blocks)
|
||||
freeze_n = int(math.floor(n * r))
|
||||
# Unfreeze all first
|
||||
for b in self._blocks:
|
||||
for p in b.parameters():
|
||||
p.requires_grad = True
|
||||
# Freeze earliest blocks
|
||||
for b in self._blocks[:freeze_n]:
|
||||
for p in b.parameters():
|
||||
p.requires_grad = False
|
||||
@@ -1,54 +0,0 @@
|
||||
# md_tower.py
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
from classes import ClinicalData
|
||||
from classes.SE_attention import SEBlock
|
||||
|
||||
class MDTower(nn.Module):
|
||||
"""MLP over ClinicalData.vectorize_row outputs (convert to torch inside tower)."""
|
||||
def __init__(self, clinical_data: ClinicalData, hidden_dim: int = 128, dropout: float = 0.1,
|
||||
use_se: bool = False, se_reduction: int = 16, se_pre_norm: bool = True):
|
||||
super().__init__()
|
||||
self.feature_dim = clinical_data.feature_dim
|
||||
self.out_dim = hidden_dim
|
||||
# two-block MLP so we can optionally freeze/thaw per block
|
||||
self.block0 = nn.Sequential(
|
||||
nn.Linear(self.feature_dim, hidden_dim),
|
||||
nn.LayerNorm(hidden_dim),
|
||||
nn.ReLU(inplace=True),
|
||||
nn.Dropout(dropout),
|
||||
)
|
||||
self.block1 = nn.Sequential(
|
||||
nn.Linear(hidden_dim, hidden_dim),
|
||||
nn.ReLU(inplace=True),
|
||||
)
|
||||
self.net = nn.Sequential(self.block0, self.block1)
|
||||
self.tower_ln = nn.LayerNorm(hidden_dim) if se_pre_norm else nn.Identity()
|
||||
self.tower_se = SEBlock(hidden_dim, reduction=se_reduction, residual=True) if use_se else None
|
||||
|
||||
def forward(self, meta_np_or_torch) -> torch.Tensor:
|
||||
if isinstance(meta_np_or_torch, torch.Tensor):
|
||||
x = meta_np_or_torch
|
||||
else:
|
||||
x = torch.as_tensor(meta_np_or_torch, dtype=torch.float32)
|
||||
h = self.net(x)
|
||||
if self.tower_se is not None:
|
||||
h, _ = self.tower_se(self.tower_ln(h))
|
||||
return h
|
||||
|
||||
def set_freeze_ratio(self, ratio: float):
|
||||
"""Optionally freeze earliest blocks of the MLP.
|
||||
With two blocks, ratio≥0.5 freezes block0; ratio≥1.0 freezes both."""
|
||||
r = max(0.0, min(1.0, float(ratio)))
|
||||
# Unfreeze all
|
||||
for p in self.block0.parameters():
|
||||
p.requires_grad = True
|
||||
for p in self.block1.parameters():
|
||||
p.requires_grad = True
|
||||
# Freeze earliest blocks based on ratio threshold
|
||||
if r >= 0.5:
|
||||
for p in self.block0.parameters():
|
||||
p.requires_grad = False
|
||||
if r >= 1.0:
|
||||
for p in self.block1.parameters():
|
||||
p.requires_grad = False
|
||||
@@ -1,99 +0,0 @@
|
||||
# papila_builders.py
|
||||
from typing import List, Dict
|
||||
import numpy as np
|
||||
import pandas as pd
|
||||
|
||||
from classes import ClinicalData # adjust import path if needed
|
||||
|
||||
# ---- Pachymetry → IOP correction (per PAPILA Table 3) ----
|
||||
_PACHY_TABLE: Dict[int, int] = {
|
||||
475:+5, 485:+4, 495:+4, 505:+3, 515:+2, 525:+1, 535:+1,
|
||||
545: 0, 555:-1, 565:-1, 575:-2, 585:-3, 595:-4, 605:-4, 615:-5,
|
||||
}
|
||||
_PACHY_KEYS = np.array(sorted(_PACHY_TABLE.keys()))
|
||||
|
||||
def _nearest_pachy_key(x: float) -> int:
|
||||
idx = int(np.argmin(np.abs(_PACHY_KEYS - float(x))))
|
||||
return int(_PACHY_KEYS[idx])
|
||||
|
||||
def _pick_iop(row: pd.Series) -> float:
|
||||
"""Prefer Pneumatic, else Perkins; may return NaN."""
|
||||
raw = row["Pneumatic"] if not pd.isna(row.get("Pneumatic", np.nan)) else row.get("Perkins", np.nan)
|
||||
return float(raw) if not pd.isna(raw) else np.nan
|
||||
|
||||
def _correct_iop(raw_iop: float, pachy: float) -> float:
|
||||
"""Return corrected IOP using nearest pachymetry bin; if pachy missing, return raw."""
|
||||
if pd.isna(raw_iop):
|
||||
return np.nan
|
||||
if pd.isna(pachy):
|
||||
return float(raw_iop)
|
||||
key = _nearest_pachy_key(float(pachy))
|
||||
return float(raw_iop) + float(_PACHY_TABLE[key])
|
||||
|
||||
def _apply_iop_and_drop_md(df: pd.DataFrame) -> pd.DataFrame:
|
||||
"""Add IOP_raw/IOP_corr and drop VF_MD if present (in-place safe)."""
|
||||
# IOP_raw
|
||||
df["IOP_raw"] = df.apply(_pick_iop, axis=1)
|
||||
|
||||
# IOP_corr
|
||||
pachy = df.get("Pachymetry", pd.Series(np.nan, index=df.index))
|
||||
df["IOP_corr"] = [
|
||||
_correct_iop(r, p) for r, p in zip(df["IOP_raw"].values, pachy.values)
|
||||
]
|
||||
|
||||
# Drop VF_MD if present
|
||||
if "VF_MD" in df.columns:
|
||||
df.drop(columns=["VF_MD"], inplace=True)
|
||||
return df
|
||||
|
||||
|
||||
def build_papila_clinical(
|
||||
image_dir: str,
|
||||
clinical_dir: str,
|
||||
label_col: str,
|
||||
cat_cols: List[str],
|
||||
n_splits: int = 5,
|
||||
random_seed: int = 42,
|
||||
) -> ClinicalData:
|
||||
"""
|
||||
Build ClinicalData exactly like the user's original build_clinical:
|
||||
- add_df(OD), set eyeID='OD'
|
||||
- add_df(OS), set eyeID='OS'
|
||||
- normalize 'Patient ID' on frames
|
||||
THEN:
|
||||
- compute IOP_raw / IOP_corr on each frame
|
||||
- drop VF_MD
|
||||
- refresh master df + kfold indices
|
||||
"""
|
||||
clinical = ClinicalData(
|
||||
image_dir=image_dir,
|
||||
clinical_dir=clinical_dir,
|
||||
label_col=label_col,
|
||||
cat_cols=cat_cols,
|
||||
n_splits=n_splits,
|
||||
random_seed=random_seed,
|
||||
)
|
||||
|
||||
# --- Load exactly like original build_clinical ---
|
||||
clinical.add_df(pd.read_excel(f"{clinical_dir}/patient_data_od.xlsx", header=1), id_column="ID")
|
||||
clinical.frames[0]["eyeID"] = "OD"
|
||||
|
||||
clinical.add_df(pd.read_excel(f"{clinical_dir}/patient_data_os.xlsx", header=1), id_column="ID")
|
||||
clinical.frames[1]["eyeID"] = "OS"
|
||||
|
||||
# Normalize 'Patient ID' on the per-eye frames (string → int)
|
||||
for frame in clinical.frames:
|
||||
frame["Patient ID"] = frame["Patient ID"].astype(str).str.extract(r"(\d+)")[0].astype(int)
|
||||
|
||||
# Build initial master as in original
|
||||
clinical._refresh_master_df()
|
||||
|
||||
# --- Post-processing ON THE FRAMES (so everything stays consistent) ---
|
||||
for i in range(len(clinical.frames)):
|
||||
clinical.frames[i] = _apply_iop_and_drop_md(clinical.frames[i])
|
||||
|
||||
# Refresh master again so IOP_raw/IOP_corr & MD removal propagate
|
||||
clinical._refresh_master_df()
|
||||
clinical._build_kfold_indices()
|
||||
|
||||
return clinical
|
||||
@@ -1,849 +0,0 @@
|
||||
"""REFUGE glaucoma classification with rotation-based TTT."""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import math
|
||||
from dataclasses import dataclass
|
||||
from pathlib import Path
|
||||
from typing import Callable, Dict, Iterable, List, Optional, Sequence, Tuple
|
||||
import random
|
||||
|
||||
import numpy as np
|
||||
from PIL import Image
|
||||
import torch
|
||||
from torch import nn
|
||||
from torch.utils.data import DataLoader, Dataset
|
||||
from torchvision import models, transforms
|
||||
from torchvision.transforms import functional as TF
|
||||
import torch.nn.functional as F
|
||||
from sklearn.metrics import roc_auc_score
|
||||
from skimage.transform import warp_polar
|
||||
from tqdm import tqdm
|
||||
|
||||
from classes.geometry_features import (
|
||||
FEATURE_DIM,
|
||||
EPS,
|
||||
compute_geometry_features,
|
||||
disc_cup_from_mask_image,
|
||||
)
|
||||
from classes.refuge_preprocessing import RefugePreprocessing, RefugeSample
|
||||
from classes.refuge_segmentation import RefugeSegmentation
|
||||
from classes.unet_segmenter import UNetSegmenter
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Dataset utilities
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
|
||||
def _default_image_transform(size: int = 256) -> transforms.Compose:
|
||||
return transforms.Compose(
|
||||
[
|
||||
transforms.Resize((size, size)),
|
||||
transforms.ToTensor(),
|
||||
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
|
||||
]
|
||||
)
|
||||
|
||||
|
||||
def _augment_image_transform(size: int = 256) -> transforms.Compose:
|
||||
return transforms.Compose(
|
||||
[
|
||||
transforms.Resize((size, size)),
|
||||
transforms.RandomHorizontalFlip(),
|
||||
transforms.RandomRotation(10),
|
||||
transforms.ColorJitter(0.1, 0.1, 0.1, 0.05),
|
||||
transforms.ToTensor(),
|
||||
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
|
||||
]
|
||||
)
|
||||
|
||||
|
||||
def _crop_from_geometry(image: Image.Image, geometry: Dict[str, float], size: int = 256) -> Image.Image:
|
||||
cx, cy = geometry["centre_x"], geometry["centre_y"]
|
||||
r = geometry["crop_radius"]
|
||||
left = max(0.0, cx - r)
|
||||
upper = max(0.0, cy - r)
|
||||
right = min(image.width, cx + r)
|
||||
lower = min(image.height, cy + r)
|
||||
crop = image.crop((left, upper, right, lower))
|
||||
return crop.resize((size, size), Image.BILINEAR)
|
||||
|
||||
|
||||
def _geometry_from_mask(mask: np.ndarray, scale: float) -> Dict[str, float]:
|
||||
mask = np.asarray(mask) > 0
|
||||
coords = np.argwhere(mask)
|
||||
if coords.size == 0:
|
||||
raise RuntimeError("Empty mask; cannot derive geometry")
|
||||
ys, xs = coords[:, 0], coords[:, 1]
|
||||
centre_x = float(xs.mean())
|
||||
centre_y = float(ys.mean())
|
||||
width = float(xs.max() - xs.min())
|
||||
height = float(ys.max() - ys.min())
|
||||
diameter = max(width, height)
|
||||
radius = diameter / 2.0
|
||||
crop_radius = radius * scale
|
||||
return {
|
||||
"centre_x": centre_x,
|
||||
"centre_y": centre_y,
|
||||
"radius": radius,
|
||||
"crop_radius": crop_radius,
|
||||
"crop_size": crop_radius * 2.0,
|
||||
}
|
||||
|
||||
|
||||
def _compute_feature_vector(disc_mask: np.ndarray, cup_mask: np.ndarray) -> np.ndarray:
|
||||
return compute_geometry_features(disc_mask, cup_mask)
|
||||
|
||||
|
||||
def _compute_polar_image(crop: Image.Image, size: int) -> Image.Image:
|
||||
arr = np.asarray(crop).astype(np.float32) / 255.0
|
||||
radius = min(arr.shape[0], arr.shape[1]) / 2.0
|
||||
polar = warp_polar(
|
||||
arr,
|
||||
radius=radius,
|
||||
scaling="linear",
|
||||
channel_axis=-1,
|
||||
)
|
||||
polar = np.clip(polar, 0.0, 1.0)
|
||||
polar_img = Image.fromarray((polar * 255).astype(np.uint8))
|
||||
return polar_img.resize((size, size), Image.BILINEAR)
|
||||
|
||||
|
||||
def _crop_mask_from_geometry(mask: np.ndarray, geometry: Dict[str, float], size: int) -> np.ndarray:
|
||||
mask_img = Image.fromarray((mask > 0).astype(np.uint8) * 255)
|
||||
cx, cy = geometry["centre_x"], geometry["centre_y"]
|
||||
r = geometry["crop_radius"]
|
||||
left = max(0.0, cx - r)
|
||||
upper = max(0.0, cy - r)
|
||||
right = min(mask_img.width, cx + r)
|
||||
lower = min(mask_img.height, cy + r)
|
||||
crop = mask_img.crop((left, upper, right, lower)).resize((size, size), Image.NEAREST)
|
||||
return (np.asarray(crop) > 0).astype(np.uint8)
|
||||
|
||||
|
||||
@dataclass
|
||||
class RefugeClassificationRecord:
|
||||
sample: RefugeSample
|
||||
geometry: Dict[str, float]
|
||||
disc_mask: Optional[np.ndarray] = None
|
||||
cup_mask: Optional[np.ndarray] = None
|
||||
|
||||
|
||||
class RefugeClassificationDataset(Dataset):
|
||||
def __init__(
|
||||
self,
|
||||
records: Sequence[RefugeClassificationRecord],
|
||||
transform: transforms.Compose,
|
||||
polar_transform: transforms.Compose,
|
||||
size: int = 256,
|
||||
) -> None:
|
||||
self.records = list(records)
|
||||
self.transform = transform
|
||||
self.polar_transform = polar_transform
|
||||
self.size = size
|
||||
|
||||
def __len__(self) -> int:
|
||||
return len(self.records)
|
||||
|
||||
def __getitem__(self, idx: int) -> Dict[str, torch.Tensor]:
|
||||
rec = self.records[idx]
|
||||
image = Image.open(rec.sample.image_path).convert("RGB")
|
||||
crop = _crop_from_geometry(image, rec.geometry, size=self.size)
|
||||
polar_image = _compute_polar_image(crop, size=self.size)
|
||||
tensor = self.transform(crop)
|
||||
polar_tensor = self.polar_transform(polar_image)
|
||||
|
||||
features = np.zeros((FEATURE_DIM,), dtype=np.float32)
|
||||
if rec.disc_mask is not None and rec.cup_mask is not None:
|
||||
disc_crop = _crop_mask_from_geometry(rec.disc_mask, rec.geometry, self.size)
|
||||
cup_crop = _crop_mask_from_geometry(rec.cup_mask, rec.geometry, self.size)
|
||||
features = _compute_feature_vector(disc_crop, cup_crop)
|
||||
|
||||
feature_tensor = torch.from_numpy(features).float()
|
||||
label = rec.sample.label
|
||||
if label is None:
|
||||
raise ValueError(f"Sample {rec.sample.sample_id} is missing glaucoma label")
|
||||
return {
|
||||
"image": tensor,
|
||||
"polar": polar_tensor,
|
||||
"features": feature_tensor,
|
||||
"label": torch.tensor(label, dtype=torch.long),
|
||||
"sample_id": rec.sample.sample_id,
|
||||
}
|
||||
|
||||
|
||||
class RefugeTTTDataset(Dataset):
|
||||
"""Dataset providing unlabeled crops for test-time training."""
|
||||
|
||||
def __init__(self, records: Sequence[RefugeClassificationRecord], transform: transforms.Compose, size: int = 256) -> None:
|
||||
self.records = list(records)
|
||||
self.transform = transform
|
||||
self.size = size
|
||||
|
||||
def __len__(self) -> int:
|
||||
return len(self.records)
|
||||
|
||||
def __getitem__(self, idx: int) -> torch.Tensor:
|
||||
rec = self.records[idx]
|
||||
image = Image.open(rec.sample.image_path).convert("RGB")
|
||||
crop = _crop_from_geometry(image, rec.geometry, size=self.size)
|
||||
return self.transform(crop)
|
||||
|
||||
|
||||
class UNetGeometryProvider:
|
||||
"""Callable wrapper that derives disc geometry using a trained UNetSegmenter."""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
segmenter: UNetSegmenter,
|
||||
threshold: float = 0.5,
|
||||
tta: bool = False,
|
||||
) -> None:
|
||||
self.segmenter = segmenter
|
||||
self.threshold = threshold
|
||||
self.tta = tta
|
||||
self.segmenter.model.eval()
|
||||
|
||||
def __call__(self, sample: RefugeSample, scale: float) -> Tuple[Dict[str, float], np.ndarray, np.ndarray]:
|
||||
image = Image.open(sample.image_path).convert("RGB")
|
||||
resized = self.segmenter.preprocess_image(image)
|
||||
tensor = transforms.ToTensor()(resized)
|
||||
tensor = self.segmenter._normalize_tensor(tensor)
|
||||
tensor = tensor.unsqueeze(0).to(self.segmenter.device)
|
||||
with torch.no_grad():
|
||||
logits = self.segmenter.model(tensor)
|
||||
if self.tta:
|
||||
t_h = torch.flip(tensor, dims=[3])
|
||||
log_h = self.segmenter.model(t_h)
|
||||
log_h = torch.flip(log_h, dims=[3])
|
||||
t_v = torch.flip(tensor, dims=[2])
|
||||
log_v = self.segmenter.model(t_v)
|
||||
log_v = torch.flip(log_v, dims=[2])
|
||||
logits = (logits + log_h + log_v) / 3.0
|
||||
probs = torch.sigmoid(logits)[0].cpu().numpy()
|
||||
|
||||
disc_pred = (probs[0] > self.threshold).astype(np.uint8) * 255
|
||||
cup_pred = (probs[1] > self.threshold).astype(np.uint8) * 255
|
||||
disc_img = Image.fromarray(disc_pred, mode="L").resize(image.size, Image.NEAREST)
|
||||
cup_img = Image.fromarray(cup_pred, mode="L").resize(image.size, Image.NEAREST)
|
||||
disc_mask = (np.array(disc_img, dtype=np.uint8) > 0).astype(np.uint8)
|
||||
cup_mask = (np.array(cup_img, dtype=np.uint8) > 0).astype(np.uint8)
|
||||
cup_mask = (cup_mask > 0) & (disc_mask > 0)
|
||||
cup_mask = cup_mask.astype(np.uint8)
|
||||
geom = _geometry_from_mask(disc_mask, scale)
|
||||
return geom, disc_mask, cup_mask
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Classification module
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
|
||||
class ArcMarginProduct(nn.Module):
|
||||
"""Additive angular margin (ArcFace) head."""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
in_features: int,
|
||||
out_features: int,
|
||||
s: float = 30.0,
|
||||
m: float = 0.5,
|
||||
easy_margin: bool = False,
|
||||
) -> None:
|
||||
super().__init__()
|
||||
self.in_features = in_features
|
||||
self.out_features = out_features
|
||||
self.s = float(s)
|
||||
self.m = float(m)
|
||||
self.easy_margin = easy_margin
|
||||
self.weight = nn.Parameter(torch.empty(out_features, in_features))
|
||||
nn.init.xavier_uniform_(self.weight)
|
||||
|
||||
self.cos_m = math.cos(m)
|
||||
self.sin_m = math.sin(m)
|
||||
self.th = math.cos(math.pi - m)
|
||||
self.mm = math.sin(math.pi - m) * m
|
||||
|
||||
def forward(self, input: torch.Tensor, label: Optional[torch.Tensor] = None) -> torch.Tensor:
|
||||
cosine = F.linear(F.normalize(input), F.normalize(self.weight))
|
||||
if label is None:
|
||||
return cosine * self.s
|
||||
|
||||
sine = torch.sqrt(torch.clamp(1.0 - cosine.pow(2), min=0.0))
|
||||
phi = cosine * self.cos_m - sine * self.sin_m
|
||||
if self.easy_margin:
|
||||
phi = torch.where(cosine > 0, phi, cosine)
|
||||
else:
|
||||
phi = torch.where(cosine > self.th, phi, cosine - self.mm)
|
||||
|
||||
one_hot = torch.zeros_like(cosine)
|
||||
one_hot.scatter_(1, label.view(-1, 1), 1.0)
|
||||
logits = (one_hot * phi) + ((1.0 - one_hot) * cosine)
|
||||
logits *= self.s
|
||||
return logits
|
||||
|
||||
|
||||
class RefugeClassification:
|
||||
"""Train and evaluate REFUGE glaucoma classifiers with TTT support."""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
preprocessing: RefugePreprocessing,
|
||||
segmentation: RefugeSegmentation,
|
||||
backbone: Optional[nn.Module] = None,
|
||||
geometry_fn: Optional[
|
||||
Callable[
|
||||
[RefugeSample, float],
|
||||
Tuple[Dict[str, float], Optional[np.ndarray], Optional[np.ndarray]],
|
||||
]
|
||||
] = None,
|
||||
cache_dir: Optional[Path] = None,
|
||||
use_all_labeled: bool = False,
|
||||
auto_val_ratio: float = 0.1,
|
||||
use_margin: bool = False,
|
||||
margin_s: float = 30.0,
|
||||
margin_m: float = 0.5,
|
||||
) -> None:
|
||||
self.preprocessing = preprocessing
|
||||
self.segmentation = segmentation
|
||||
if backbone is not None:
|
||||
self.backbone = backbone
|
||||
in_features = getattr(self.backbone, "_feature_dim", None)
|
||||
if in_features is None:
|
||||
if hasattr(self.backbone, "fc") and hasattr(self.backbone.fc, "in_features"):
|
||||
in_features = self.backbone.fc.in_features # type: ignore[attr-defined]
|
||||
self.backbone.fc = nn.Identity() # type: ignore[attr-defined]
|
||||
else:
|
||||
raise ValueError(
|
||||
"Provided backbone must have '_feature_dim' or expose fc.in_features"
|
||||
)
|
||||
else:
|
||||
self.backbone = self._default_backbone()
|
||||
in_features = getattr(self.backbone, "_feature_dim", None)
|
||||
if in_features is None:
|
||||
in_features = self.backbone.fc.in_features # type: ignore[attr-defined]
|
||||
self.backbone.fc = nn.Identity() # type: ignore[attr-defined]
|
||||
self.feature_dim = in_features
|
||||
self.use_polar = True
|
||||
self.extra_feature_dim = FEATURE_DIM
|
||||
combined_dim = self.feature_dim * (1 + int(self.use_polar)) + self.extra_feature_dim
|
||||
self.margin_s = float(margin_s)
|
||||
self.margin_m = float(margin_m)
|
||||
self.use_margin = bool(use_margin)
|
||||
if self.use_margin:
|
||||
self.classifier_head = ArcMarginProduct(
|
||||
combined_dim, 2, s=self.margin_s, m=self.margin_m
|
||||
)
|
||||
else:
|
||||
self.classifier_head = nn.Linear(combined_dim, 2)
|
||||
self.rotation_head = nn.Linear(self.feature_dim, 4)
|
||||
|
||||
self.train_dataset: Optional[RefugeClassificationDataset] = None
|
||||
self.val_dataset: Optional[RefugeClassificationDataset] = None
|
||||
self.train_loader: Optional[DataLoader] = None
|
||||
self.val_loader: Optional[DataLoader] = None
|
||||
self.ttt_transform = _default_image_transform()
|
||||
self.train_transform = _augment_image_transform()
|
||||
self.eval_transform = _default_image_transform()
|
||||
self.polar_transform = _default_image_transform()
|
||||
self.crop_scale = 2.5
|
||||
self.crop_size = 256
|
||||
self.geometry_cache: Dict[
|
||||
str, Tuple[Dict[str, float], Optional[np.ndarray], Optional[np.ndarray]]
|
||||
] = {}
|
||||
self.train_records: List[RefugeClassificationRecord] = []
|
||||
self.val_records: List[RefugeClassificationRecord] = []
|
||||
self._geometry_fn = geometry_fn
|
||||
self.cache_dir = cache_dir
|
||||
if self.cache_dir is not None:
|
||||
self.cache_dir.mkdir(parents=True, exist_ok=True)
|
||||
self.use_all_labeled = use_all_labeled
|
||||
self.auto_val_ratio = auto_val_ratio
|
||||
|
||||
# ------------------------------------------------------------------
|
||||
@staticmethod
|
||||
def _default_backbone() -> nn.Module:
|
||||
weights = models.ResNet50_Weights.IMAGENET1K_V2
|
||||
model = models.resnet50(weights=weights)
|
||||
in_features = model.fc.in_features
|
||||
model.fc = nn.Identity()
|
||||
setattr(model, "_feature_dim", in_features)
|
||||
return model
|
||||
|
||||
# ------------------------------------------------------------------
|
||||
def build_datasets(
|
||||
self,
|
||||
crop_scale: float = 2.5,
|
||||
crop_size: int = 256,
|
||||
batch_size: int = 16,
|
||||
num_workers: int = 4,
|
||||
) -> None:
|
||||
self.crop_scale = crop_scale
|
||||
self.crop_size = crop_size
|
||||
self.train_transform = _augment_image_transform(crop_size)
|
||||
self.eval_transform = _default_image_transform(crop_size)
|
||||
self.ttt_transform = _default_image_transform(crop_size)
|
||||
self.polar_transform = _default_image_transform(crop_size)
|
||||
|
||||
manifest = list(self.preprocessing.build_manifest())
|
||||
train_records: List[RefugeClassificationRecord] = []
|
||||
val_records: List[RefugeClassificationRecord] = []
|
||||
|
||||
allowed_splits = {"train", "val"}
|
||||
candidates = [
|
||||
sample
|
||||
for sample in manifest
|
||||
if sample.label is not None and sample.split in allowed_splits
|
||||
]
|
||||
|
||||
print(
|
||||
f"[classifier] Building datasets from {len(candidates)} labelled samples (train/val)"
|
||||
)
|
||||
|
||||
skipped: List[str] = []
|
||||
for sample in tqdm(
|
||||
candidates,
|
||||
desc="Preparing records",
|
||||
unit="sample",
|
||||
leave=False,
|
||||
):
|
||||
try:
|
||||
geom, disc_mask, cup_mask = self._resolve_geometry(sample, crop_scale)
|
||||
except RuntimeError:
|
||||
skipped.append(sample.sample_id)
|
||||
continue
|
||||
record = RefugeClassificationRecord(
|
||||
sample=sample,
|
||||
geometry=geom,
|
||||
disc_mask=disc_mask,
|
||||
cup_mask=cup_mask,
|
||||
)
|
||||
if sample.split == "train" or (
|
||||
self.use_all_labeled and sample.split == "val"
|
||||
):
|
||||
train_records.append(record)
|
||||
else:
|
||||
val_records.append(record)
|
||||
|
||||
if skipped:
|
||||
print(
|
||||
f"[classifier] WARNING: {len(skipped)}/{len(candidates)} samples skipped "
|
||||
f"due to empty segmentation mask: {skipped}"
|
||||
)
|
||||
|
||||
if (not val_records or self.use_all_labeled) and train_records and self.auto_val_ratio > 0.0:
|
||||
rng = random.Random(42)
|
||||
label_groups: Dict[int, List[RefugeClassificationRecord]] = {}
|
||||
for rec in train_records:
|
||||
label = int(rec.sample.label or 0)
|
||||
label_groups.setdefault(label, []).append(rec)
|
||||
|
||||
new_train: List[RefugeClassificationRecord] = []
|
||||
new_val: List[RefugeClassificationRecord] = []
|
||||
for recs in label_groups.values():
|
||||
rng.shuffle(recs)
|
||||
if len(recs) <= 1:
|
||||
new_train.extend(recs)
|
||||
continue
|
||||
val_count = max(1, int(round(len(recs) * self.auto_val_ratio)))
|
||||
if val_count >= len(recs):
|
||||
val_count = len(recs) - 1
|
||||
new_val.extend(recs[:val_count])
|
||||
new_train.extend(recs[val_count:])
|
||||
|
||||
if not new_val:
|
||||
# Fallback: ensure at least one validation sample if possible
|
||||
if len(new_train) > 1:
|
||||
new_val.append(new_train.pop())
|
||||
|
||||
if new_val:
|
||||
val_records = new_val
|
||||
train_records = new_train
|
||||
|
||||
self.train_records = train_records
|
||||
self.val_records = val_records
|
||||
|
||||
print(
|
||||
f"[classifier] Records ready → train: {len(train_records)}, val: {len(val_records)}"
|
||||
)
|
||||
|
||||
self.train_dataset = RefugeClassificationDataset(
|
||||
train_records,
|
||||
transform=self.train_transform,
|
||||
polar_transform=self.polar_transform,
|
||||
size=crop_size,
|
||||
)
|
||||
self.val_dataset = RefugeClassificationDataset(
|
||||
val_records,
|
||||
transform=self.eval_transform,
|
||||
polar_transform=self.polar_transform,
|
||||
size=crop_size,
|
||||
)
|
||||
|
||||
self.train_loader = DataLoader(
|
||||
self.train_dataset,
|
||||
batch_size=batch_size,
|
||||
shuffle=True,
|
||||
num_workers=num_workers,
|
||||
pin_memory=True,
|
||||
)
|
||||
self.val_loader = DataLoader(
|
||||
self.val_dataset,
|
||||
batch_size=batch_size,
|
||||
shuffle=False,
|
||||
num_workers=num_workers,
|
||||
pin_memory=True,
|
||||
)
|
||||
|
||||
print(
|
||||
"[classifier] DataLoaders prepared — training batches will start shortly"
|
||||
)
|
||||
|
||||
# ------------------------------------------------------------------
|
||||
def _resolve_geometry(
|
||||
self, sample: RefugeSample, scale: float
|
||||
) -> Tuple[Dict[str, float], Optional[np.ndarray], Optional[np.ndarray]]:
|
||||
key = self._cache_key(sample.sample_id, scale)
|
||||
cached = self.geometry_cache.get(key)
|
||||
if cached is not None:
|
||||
return cached
|
||||
|
||||
cache_path = self._cache_path(sample.sample_id, scale)
|
||||
if cache_path is not None and cache_path.exists():
|
||||
data = np.load(cache_path, allow_pickle=False)
|
||||
geom = {
|
||||
"centre_x": float(data["centre_x"]),
|
||||
"centre_y": float(data["centre_y"]),
|
||||
"radius": float(data["radius"]),
|
||||
"crop_radius": float(data["crop_radius"]),
|
||||
"crop_size": float(data["crop_size"]),
|
||||
}
|
||||
disc_mask = None
|
||||
cup_mask = None
|
||||
if int(data["has_disc"]):
|
||||
disc_mask = data["disc_mask"].astype(np.uint8)
|
||||
if int(data["has_cup"]):
|
||||
cup_mask = data["cup_mask"].astype(np.uint8)
|
||||
self.geometry_cache[key] = (geom, disc_mask, cup_mask)
|
||||
return geom, disc_mask, cup_mask
|
||||
|
||||
disc_mask: Optional[np.ndarray] = None
|
||||
cup_mask: Optional[np.ndarray] = None
|
||||
|
||||
if sample.mask_path and sample.mask_path.exists():
|
||||
mask_img = Image.open(sample.mask_path).convert("RGB")
|
||||
disc_mask, cup_mask = disc_cup_from_mask_image(mask_img)
|
||||
geom = _geometry_from_mask(disc_mask, scale)
|
||||
elif self._geometry_fn is not None:
|
||||
geom, disc_mask, cup_mask = self._geometry_fn(sample, scale)
|
||||
else:
|
||||
geom = self.segmentation.infer_disc_geometry(sample, scale=scale)
|
||||
try:
|
||||
pred_mask = self.segmentation.predict_mask(sample).numpy()
|
||||
disc_mask = pred_mask.astype(np.uint8)
|
||||
except Exception:
|
||||
disc_mask = None
|
||||
cup_mask = None
|
||||
|
||||
if cache_path is not None:
|
||||
try:
|
||||
np.savez_compressed(
|
||||
cache_path,
|
||||
centre_x=geom["centre_x"],
|
||||
centre_y=geom["centre_y"],
|
||||
radius=geom["radius"],
|
||||
crop_radius=geom["crop_radius"],
|
||||
crop_size=geom.get("crop_size", geom["crop_radius"] * 2.0),
|
||||
disc_mask=disc_mask if disc_mask is not None else np.array([], dtype=np.uint8),
|
||||
cup_mask=cup_mask if cup_mask is not None else np.array([], dtype=np.uint8),
|
||||
has_disc=int(disc_mask is not None),
|
||||
has_cup=int(cup_mask is not None),
|
||||
)
|
||||
except Exception:
|
||||
pass
|
||||
|
||||
self.geometry_cache[key] = (geom, disc_mask, cup_mask)
|
||||
return geom, disc_mask, cup_mask
|
||||
|
||||
def set_geometry_fn(
|
||||
self,
|
||||
geometry_fn: Optional[
|
||||
Callable[
|
||||
[RefugeSample, float],
|
||||
Tuple[Dict[str, float], Optional[np.ndarray], Optional[np.ndarray]],
|
||||
]
|
||||
],
|
||||
) -> None:
|
||||
self._geometry_fn = geometry_fn
|
||||
self.geometry_cache.clear()
|
||||
|
||||
def build_records_for_samples(
|
||||
self,
|
||||
samples: Sequence[RefugeSample],
|
||||
crop_scale: Optional[float] = None,
|
||||
progress_prefix: Optional[str] = None,
|
||||
) -> List[RefugeClassificationRecord]:
|
||||
scale = crop_scale if crop_scale is not None else self.crop_scale
|
||||
records: List[RefugeClassificationRecord] = []
|
||||
skipped: List[str] = []
|
||||
iterator: Iterable[RefugeSample]
|
||||
if progress_prefix is not None:
|
||||
iterator = tqdm(samples, desc=progress_prefix, unit="sample", leave=False)
|
||||
else:
|
||||
iterator = samples
|
||||
labeled = [s for s in samples if s.label is not None]
|
||||
for sample in iterator:
|
||||
if sample.label is None:
|
||||
continue
|
||||
try:
|
||||
geom, disc_mask, cup_mask = self._resolve_geometry(sample, scale)
|
||||
except RuntimeError:
|
||||
skipped.append(sample.sample_id)
|
||||
continue
|
||||
records.append(
|
||||
RefugeClassificationRecord(
|
||||
sample=sample,
|
||||
geometry=geom,
|
||||
disc_mask=disc_mask,
|
||||
cup_mask=cup_mask,
|
||||
)
|
||||
)
|
||||
prefix = f"[{progress_prefix}]" if progress_prefix else "[classifier]"
|
||||
if skipped:
|
||||
print(
|
||||
f"{prefix} WARNING: {len(skipped)}/{len(labeled)} samples skipped "
|
||||
f"due to empty segmentation mask: {skipped}"
|
||||
)
|
||||
else:
|
||||
print(f"{prefix} All {len(labeled)} samples processed successfully.")
|
||||
return records
|
||||
|
||||
def clear_disk_cache(self) -> None:
|
||||
"""Delete all cached geometry/mask .npz files in cache_dir."""
|
||||
if self.cache_dir is None or not self.cache_dir.exists():
|
||||
return
|
||||
removed = 0
|
||||
for f in self.cache_dir.glob("*.npz"):
|
||||
f.unlink()
|
||||
removed += 1
|
||||
self.geometry_cache.clear()
|
||||
print(f"[classifier] Cleared {removed} cached geometry files from {self.cache_dir}")
|
||||
|
||||
def _cache_key(self, sample_id: str, scale: float) -> str:
|
||||
scale_tag = int(round(scale * 100))
|
||||
return f"{sample_id}_s{scale_tag}"
|
||||
|
||||
def _cache_path(self, sample_id: str, scale: float) -> Optional[Path]:
|
||||
if self.cache_dir is None:
|
||||
return None
|
||||
return self.cache_dir / f"{self._cache_key(sample_id, scale)}.npz"
|
||||
|
||||
# ------------------------------------------------------------------
|
||||
def train(
|
||||
self,
|
||||
epochs: int = 30,
|
||||
lr: float = 1e-4,
|
||||
weight_decay: float = 1e-4,
|
||||
device: Optional[str] = None,
|
||||
rotation_weight: float = 0.5,
|
||||
checkpoint_dir: Optional[Path] = None,
|
||||
) -> Dict[str, float]:
|
||||
if self.train_loader is None or self.val_loader is None:
|
||||
self.build_datasets()
|
||||
|
||||
device = device or ("cuda" if torch.cuda.is_available() else "cpu")
|
||||
self.backbone.to(device)
|
||||
self.classifier_head.to(device)
|
||||
self.rotation_head.to(device)
|
||||
|
||||
params = list(self.backbone.parameters()) + list(self.classifier_head.parameters()) + list(self.rotation_head.parameters())
|
||||
optimizer = torch.optim.Adam(params, lr=lr, weight_decay=weight_decay)
|
||||
clf_loss = nn.CrossEntropyLoss()
|
||||
rot_loss = nn.CrossEntropyLoss()
|
||||
|
||||
best_auc = 0.0
|
||||
history: Dict[str, float] = {}
|
||||
|
||||
epoch_iter = tqdm(range(1, epochs + 1), desc="Epochs", unit="epoch")
|
||||
|
||||
print(
|
||||
f"[classifier] Starting training for {epochs} epochs with batch size {self.train_loader.batch_size}"
|
||||
)
|
||||
|
||||
for epoch in epoch_iter:
|
||||
self.backbone.train()
|
||||
self.classifier_head.train()
|
||||
self.rotation_head.train()
|
||||
running_loss = 0.0
|
||||
|
||||
batch_iter = tqdm(
|
||||
self.train_loader, # type: ignore[arg-type]
|
||||
desc=f"Train {epoch}/{epochs}",
|
||||
leave=False,
|
||||
unit="batch",
|
||||
)
|
||||
|
||||
for batch in batch_iter:
|
||||
images = batch["image"].to(device)
|
||||
polars = batch["polar"].to(device)
|
||||
extra_feats = batch["features"].to(device)
|
||||
labels = batch["label"].to(device)
|
||||
optimizer.zero_grad()
|
||||
|
||||
feats_img = self.backbone(images)
|
||||
feats = feats_img
|
||||
if self.use_polar:
|
||||
feats_polar = self.backbone(polars)
|
||||
feats = torch.cat([feats, feats_polar], dim=1)
|
||||
if self.extra_feature_dim > 0:
|
||||
feats = torch.cat([feats, extra_feats], dim=1)
|
||||
if self.use_margin:
|
||||
logits = self.classifier_head(feats, labels)
|
||||
else:
|
||||
logits = self.classifier_head(feats)
|
||||
loss_cls = clf_loss(logits, labels)
|
||||
|
||||
rot_imgs, rot_labels = self._build_rotation_batch(images)
|
||||
feats_rot = self.backbone(rot_imgs)
|
||||
logits_rot = self.rotation_head(feats_rot)
|
||||
loss_rot = rot_loss(logits_rot, rot_labels)
|
||||
|
||||
loss = loss_cls + rotation_weight * loss_rot
|
||||
loss.backward()
|
||||
optimizer.step()
|
||||
running_loss += loss.item() * images.size(0)
|
||||
|
||||
train_loss = running_loss / len(self.train_loader.dataset) # type: ignore[arg-type]
|
||||
metrics = self.evaluate(device=device)
|
||||
history[f"epoch_{epoch}_loss"] = train_loss
|
||||
history[f"epoch_{epoch}_auc"] = metrics.get("auc", float("nan"))
|
||||
|
||||
auc_val = metrics.get("auc", 0.0)
|
||||
epoch_iter.set_postfix(loss=f"{train_loss:.4f}", auc=f"{auc_val:.4f}")
|
||||
|
||||
if auc_val > best_auc:
|
||||
best_auc = metrics["auc"]
|
||||
if checkpoint_dir is not None:
|
||||
checkpoint_dir.mkdir(parents=True, exist_ok=True)
|
||||
torch.save({
|
||||
"backbone": self.backbone.state_dict(),
|
||||
"classifier": self.classifier_head.state_dict(),
|
||||
"rotation": self.rotation_head.state_dict(),
|
||||
}, checkpoint_dir / "refuge_classifier_best.pt")
|
||||
|
||||
return {"best_auc": best_auc, **history}
|
||||
|
||||
# ------------------------------------------------------------------
|
||||
def evaluate(
|
||||
self,
|
||||
split: str = "val",
|
||||
apply_ttt: bool = False,
|
||||
device: Optional[str] = None,
|
||||
) -> Dict[str, float]:
|
||||
if split != "val":
|
||||
raise ValueError("Only validation split supported currently")
|
||||
if self.val_loader is None:
|
||||
self.build_datasets()
|
||||
|
||||
device = device or ("cuda" if torch.cuda.is_available() else "cpu")
|
||||
self.backbone.to(device)
|
||||
self.classifier_head.to(device)
|
||||
self.rotation_head.to(device)
|
||||
|
||||
if apply_ttt:
|
||||
ttt_ds = RefugeTTTDataset(self.val_records, transform=self.ttt_transform, size=self.crop_size)
|
||||
ttt_loader = DataLoader(ttt_ds, batch_size=32, shuffle=False)
|
||||
self.apply_ttt(ttt_loader, device=device)
|
||||
|
||||
self.backbone.eval()
|
||||
self.classifier_head.eval()
|
||||
preds: List[float] = []
|
||||
targets: List[int] = []
|
||||
|
||||
with torch.no_grad():
|
||||
val_iter = tqdm(self.val_loader, desc="Validate", leave=False, unit="batch")
|
||||
for batch in val_iter: # type: ignore[arg-type]
|
||||
images = batch["image"].to(device)
|
||||
labels = batch["label"].to(device)
|
||||
polars = batch["polar"].to(device)
|
||||
extra_feats = batch["features"].to(device)
|
||||
feats_img = self.backbone(images)
|
||||
feats = feats_img
|
||||
if self.use_polar:
|
||||
feats_polar = self.backbone(polars)
|
||||
feats = torch.cat([feats, feats_polar], dim=1)
|
||||
if self.extra_feature_dim > 0:
|
||||
feats = torch.cat([feats, extra_feats], dim=1)
|
||||
if self.use_margin:
|
||||
logits = self.classifier_head(feats)
|
||||
else:
|
||||
logits = self.classifier_head(feats)
|
||||
probs = torch.softmax(logits, dim=1)[:, 1]
|
||||
preds.extend(probs.cpu().numpy().tolist())
|
||||
targets.extend(labels.cpu().numpy().tolist())
|
||||
|
||||
auc = 0.0
|
||||
try:
|
||||
if len(set(targets)) > 1:
|
||||
auc = float(roc_auc_score(targets, preds))
|
||||
except ValueError:
|
||||
auc = 0.0
|
||||
|
||||
return {"auc": auc}
|
||||
|
||||
# ------------------------------------------------------------------
|
||||
def apply_ttt(self, loader: DataLoader, device: Optional[str] = None, steps: int = 1, lr: float = 1e-5) -> None:
|
||||
device = device or ("cuda" if torch.cuda.is_available() else "cpu")
|
||||
self.backbone.to(device)
|
||||
self.rotation_head.to(device)
|
||||
self.backbone.train()
|
||||
self.rotation_head.train()
|
||||
|
||||
optimizer = torch.optim.Adam(list(self.backbone.parameters()) + list(self.rotation_head.parameters()), lr=lr)
|
||||
criterion = nn.CrossEntropyLoss()
|
||||
|
||||
for _ in range(steps):
|
||||
for batch in tqdm(loader, desc="TTT adapt", leave=False, unit="batch"):
|
||||
if isinstance(batch, dict):
|
||||
images = batch["image"].to(device)
|
||||
else:
|
||||
images = batch.to(device)
|
||||
optimizer.zero_grad()
|
||||
rot_imgs, rot_labels = self._build_rotation_batch(images)
|
||||
feats = self.backbone(rot_imgs)
|
||||
logits = self.rotation_head(feats)
|
||||
loss = criterion(logits, rot_labels)
|
||||
loss.backward()
|
||||
optimizer.step()
|
||||
|
||||
# ------------------------------------------------------------------
|
||||
def extract_backbone(self) -> nn.Module:
|
||||
return self.backbone
|
||||
|
||||
def save_checkpoint(self, output_dir: Path) -> None:
|
||||
output_dir.mkdir(parents=True, exist_ok=True)
|
||||
torch.save({
|
||||
"backbone": self.backbone.state_dict(),
|
||||
"classifier": self.classifier_head.state_dict(),
|
||||
"rotation": self.rotation_head.state_dict(),
|
||||
}, output_dir / "refuge_classifier.pt")
|
||||
|
||||
def load_checkpoint(self, checkpoint_path: Path) -> None:
|
||||
payload = torch.load(checkpoint_path, map_location="cpu")
|
||||
self.backbone.load_state_dict(payload["backbone"])
|
||||
self.classifier_head.load_state_dict(payload["classifier"])
|
||||
self.rotation_head.load_state_dict(payload["rotation"])
|
||||
|
||||
# ------------------------------------------------------------------
|
||||
def _build_rotation_batch(self, images: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
|
||||
rotations = [0, 90, 180, 270]
|
||||
rotated = []
|
||||
labels = []
|
||||
for idx, angle in enumerate(rotations):
|
||||
rot = TF.rotate(images, angle)
|
||||
rotated.append(rot)
|
||||
labels.append(torch.full((images.size(0),), idx, dtype=torch.long, device=images.device))
|
||||
batch = torch.cat(rotated, dim=0)
|
||||
batch_labels = torch.cat(labels, dim=0)
|
||||
return batch, batch_labels
|
||||
@@ -1,306 +0,0 @@
|
||||
"""Utilities for preparing REFUGE (REFUGE1/REFUGE2) datasets.
|
||||
|
||||
Builds a unified manifest across all provided splits (REFUGE1 train/val/test
|
||||
and REFUGE2 validation/test), exposing image paths, glaucoma labels, disc/cup
|
||||
masks, and fovea coordinates so downstream segmentation/classification modules
|
||||
can operate without additional bookkeeping.
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
from dataclasses import dataclass
|
||||
from pathlib import Path
|
||||
from typing import Dict, Iterable, List, Optional, Tuple
|
||||
|
||||
import pandas as pd
|
||||
|
||||
|
||||
@dataclass
|
||||
class RefugeSample:
|
||||
"""Lightweight container describing a REFUGE sample."""
|
||||
|
||||
sample_id: str
|
||||
dataset: str
|
||||
split: str
|
||||
image_path: Path
|
||||
label: Optional[int]
|
||||
device: Optional[str]
|
||||
mask_path: Optional[Path]
|
||||
fovea_coord: Optional[Tuple[float, float]]
|
||||
|
||||
|
||||
class RefugePreprocessing:
|
||||
"""Builds manifests and provides shared helpers for REFUGE workflows.
|
||||
|
||||
Responsibilities:
|
||||
* scan the REFUGE directory structure and build a consistent manifest
|
||||
(train/val/test, device vendor, ground-truth labels)
|
||||
* expose convenience loaders for raw RGB frames, OD/OC masks, and
|
||||
optional fovea landmarks
|
||||
* compute geometric metadata (disc centres, diameters) so downstream
|
||||
stages can crop ROIs lazily instead of storing pre-rendered tiles
|
||||
"""
|
||||
|
||||
def __init__(self, root_dir: Path | str) -> None:
|
||||
self.root_dir = Path(root_dir)
|
||||
self._manifest = None # populated by build_manifest()
|
||||
|
||||
# ------------------------------------------------------------------
|
||||
# Manifest handling
|
||||
# ------------------------------------------------------------------
|
||||
def build_manifest(self, refresh: bool = False) -> Iterable[RefugeSample]:
|
||||
"""Return an iterable of :class:`RefugeSample` records.
|
||||
|
||||
Parameters
|
||||
----------
|
||||
refresh:
|
||||
when True, force a rescan of the filesystem instead of reusing the
|
||||
cached manifest.
|
||||
|
||||
Returns
|
||||
-------
|
||||
Iterable[RefugeSample]
|
||||
A sequence containing one entry per sample in the REFUGE datasets.
|
||||
|
||||
Notes
|
||||
-----
|
||||
The actual manifest-building logic will live here: parsing the
|
||||
directory structure, reading any provided CSV/Excel metadata, and
|
||||
aligning masks/labels. For now, this method raises ``NotImplementedError``
|
||||
so callers are reminded to hook it up before use.
|
||||
"""
|
||||
|
||||
if self._manifest is not None and not refresh:
|
||||
return self._manifest
|
||||
|
||||
manifest: List[RefugeSample] = []
|
||||
|
||||
manifest.extend(self._collect_refuge1_train())
|
||||
manifest.extend(self._collect_refuge1_val())
|
||||
manifest.extend(self._collect_refuge1_test())
|
||||
manifest.extend(self._collect_refuge2_val())
|
||||
manifest.extend(self._collect_refuge2_test())
|
||||
|
||||
self._manifest = manifest
|
||||
return self._manifest
|
||||
|
||||
# ------------------------------------------------------------------
|
||||
# Accessors for downstream modules
|
||||
# ------------------------------------------------------------------
|
||||
def load_image(self, sample: RefugeSample):
|
||||
"""Return the RGB fundus image for ``sample``.
|
||||
|
||||
Implementors should handle color-space consistency (e.g., ensure RGB vs
|
||||
BGR) and any global normalisation desired across devices.
|
||||
"""
|
||||
|
||||
raise NotImplementedError("Image loading to be implemented")
|
||||
|
||||
def load_mask(self, sample: RefugeSample):
|
||||
"""Return the optic disc/cup mask for ``sample`` if available."""
|
||||
|
||||
raise NotImplementedError("Mask loading to be implemented")
|
||||
|
||||
def disc_geometry(self, sample: RefugeSample) -> Dict[str, float]:
|
||||
"""Compute disc centre and diameter from the mask.
|
||||
|
||||
The segmentation module will rely on this to crop 2.5–3× disc-diameter
|
||||
ROIs at training time.
|
||||
"""
|
||||
|
||||
raise NotImplementedError("Disc geometry helper to be implemented")
|
||||
|
||||
# ------------------------------------------------------------------
|
||||
# Internal helpers
|
||||
# ------------------------------------------------------------------
|
||||
def _collect_refuge1_train(self) -> List[RefugeSample]:
|
||||
base = self.root_dir / "Train" / "REFUGE1-train"
|
||||
if not base.exists():
|
||||
return []
|
||||
|
||||
fovea_path = base / "Fovea_location.xlsx"
|
||||
fovea_map = self._read_fovea_table(fovea_path, img_col="ImgName")
|
||||
|
||||
samples: List[RefugeSample] = []
|
||||
image_root = base / "Training400"
|
||||
mask_root = base / "Disc_Cup_Masks"
|
||||
|
||||
for label_name, label_val in ("Glaucoma", 1), ("Non-Glaucoma", 0):
|
||||
img_dir = image_root / label_name
|
||||
mask_dir = mask_root / label_name
|
||||
if not img_dir.exists():
|
||||
continue
|
||||
for image_path in sorted(img_dir.glob("*.jpg")):
|
||||
img_name = image_path.name
|
||||
mask_path = (mask_dir / image_path.with_suffix(".bmp").name)
|
||||
fovea = fovea_map.get(img_name)
|
||||
sample_id = f"refuge1_train_{image_path.stem}"
|
||||
samples.append(
|
||||
RefugeSample(
|
||||
sample_id=sample_id,
|
||||
dataset="refuge1",
|
||||
split="train",
|
||||
image_path=image_path,
|
||||
label=label_val,
|
||||
device=None,
|
||||
mask_path=mask_path if mask_path.exists() else None,
|
||||
fovea_coord=fovea,
|
||||
)
|
||||
)
|
||||
return samples
|
||||
|
||||
def _collect_refuge1_val(self) -> List[RefugeSample]:
|
||||
base = self.root_dir / "Train" / "REFUGE1-val"
|
||||
if not base.exists():
|
||||
return []
|
||||
|
||||
fovea_path = base / "Fovea_locations.xlsx"
|
||||
df = pd.read_excel(fovea_path)
|
||||
samples: List[RefugeSample] = []
|
||||
image_root = base / "REFUGE-Validation400"
|
||||
mask_root = base / "Disc_Cup_Masks"
|
||||
|
||||
for _, row in df.iterrows():
|
||||
img_name = row["ImgName"]
|
||||
image_path = image_root / img_name
|
||||
mask_path = mask_root / Path(img_name).with_suffix(".bmp").name
|
||||
fovea = self._extract_fovea(row, x_key="Fovea_X", y_key="Fovea_Y")
|
||||
label = int(row.get("Glaucoma Label", 0)) if not pd.isna(row.get("Glaucoma Label", 0)) else None
|
||||
sample_id = f"refuge1_val_{Path(img_name).stem}"
|
||||
samples.append(
|
||||
RefugeSample(
|
||||
sample_id=sample_id,
|
||||
dataset="refuge1",
|
||||
split="val",
|
||||
image_path=image_path,
|
||||
label=label,
|
||||
device=None,
|
||||
mask_path=mask_path if mask_path.exists() else None,
|
||||
fovea_coord=fovea,
|
||||
)
|
||||
)
|
||||
return samples
|
||||
|
||||
def _collect_refuge1_test(self) -> List[RefugeSample]:
|
||||
base = self.root_dir / "Train" / "REFUGE1-test"
|
||||
if not base.exists():
|
||||
return []
|
||||
|
||||
df = pd.read_excel(base / "Glaucoma_label_and_Fovea_location.xlsx")
|
||||
image_root = base / "Test400"
|
||||
mask_root = base / "Disc_Cup_Masks"
|
||||
samples: List[RefugeSample] = []
|
||||
|
||||
for _, row in df.iterrows():
|
||||
img_name = row["ImgName"]
|
||||
image_path = image_root / img_name
|
||||
mask_path = mask_root / Path(img_name).with_suffix(".bmp").name
|
||||
fovea = self._extract_fovea(row, x_key="Fovea_X", y_key="Fovea_Y")
|
||||
label = int(row.get("Label(Glaucoma=1)", 0)) if not pd.isna(row.get("Label(Glaucoma=1)", 0)) else None
|
||||
sample_id = f"refuge1_test_{Path(img_name).stem}"
|
||||
samples.append(
|
||||
RefugeSample(
|
||||
sample_id=sample_id,
|
||||
dataset="refuge1",
|
||||
split="test",
|
||||
image_path=image_path,
|
||||
label=label,
|
||||
device=None,
|
||||
mask_path=mask_path if mask_path.exists() else None,
|
||||
fovea_coord=fovea,
|
||||
)
|
||||
)
|
||||
return samples
|
||||
|
||||
def _collect_refuge2_val(self) -> List[RefugeSample]:
|
||||
base = self.root_dir / "Validation"
|
||||
if not base.exists():
|
||||
return []
|
||||
|
||||
label_df = pd.read_csv(base / "glaucoma.csv")
|
||||
fovea_df = pd.read_csv(base / "fovea.csv")
|
||||
fovea_map = {
|
||||
row["ImageName"]: (float(row["Fovea_X"]), float(row["Fovea_Y"]))
|
||||
for _, row in fovea_df.iterrows()
|
||||
}
|
||||
samples: List[RefugeSample] = []
|
||||
image_root = base / "Images"
|
||||
mask_root = base / "Disc_Masks"
|
||||
|
||||
for _, row in label_df.iterrows():
|
||||
img_name = row["FileName"]
|
||||
image_path = image_root / img_name
|
||||
mask_path = mask_root / Path(img_name).with_suffix(".png").name
|
||||
label = row.get("Glaucoma Risk")
|
||||
label = int(label) if label == label else None
|
||||
sample_id = f"refuge2_val_{Path(img_name).stem}"
|
||||
samples.append(
|
||||
RefugeSample(
|
||||
sample_id=sample_id,
|
||||
dataset="refuge2",
|
||||
split="val",
|
||||
image_path=image_path,
|
||||
label=label,
|
||||
device=None,
|
||||
mask_path=mask_path if mask_path.exists() else None,
|
||||
fovea_coord=fovea_map.get(img_name),
|
||||
)
|
||||
)
|
||||
return samples
|
||||
|
||||
def _collect_refuge2_test(self) -> List[RefugeSample]:
|
||||
base = self.root_dir / "Test"
|
||||
if not base.exists():
|
||||
return []
|
||||
|
||||
label_df = pd.read_excel(base / "task1.xls", header=None, names=["ImgName", "Glaucoma"])
|
||||
fovea_df = pd.read_excel(base / "fovea.xlsx")
|
||||
fovea_map = {
|
||||
row["ImageName"]: (float(row["Fovea_X"]), float(row["Fovea_Y"]))
|
||||
for _, row in fovea_df.iterrows()
|
||||
}
|
||||
samples: List[RefugeSample] = []
|
||||
image_root = base / "refuge2-test"
|
||||
mask_root = base / "Disc_Mask"
|
||||
|
||||
for _, row in label_df.iterrows():
|
||||
img_name = row["ImgName"]
|
||||
image_path = image_root / img_name
|
||||
mask_path = mask_root / Path(img_name).with_suffix(".png").name
|
||||
label = row.get("Glaucoma")
|
||||
label = int(label) if label == label else None
|
||||
sample_id = f"refuge2_test_{Path(img_name).stem}"
|
||||
samples.append(
|
||||
RefugeSample(
|
||||
sample_id=sample_id,
|
||||
dataset="refuge2",
|
||||
split="test",
|
||||
image_path=image_path,
|
||||
label=label,
|
||||
device=None,
|
||||
mask_path=mask_path if mask_path.exists() else None,
|
||||
fovea_coord=fovea_map.get(img_name),
|
||||
)
|
||||
)
|
||||
return samples
|
||||
|
||||
@staticmethod
|
||||
def _read_fovea_table(path: Path, img_col: str) -> Dict[str, Tuple[float, float]]:
|
||||
if not path.exists():
|
||||
return {}
|
||||
df = pd.read_excel(path)
|
||||
mapping: Dict[str, Tuple[float, float]] = {}
|
||||
for _, row in df.iterrows():
|
||||
mapping[row[img_col]] = (
|
||||
float(row.get("Fovea_X", float("nan"))),
|
||||
float(row.get("Fovea_Y", float("nan"))),
|
||||
)
|
||||
return mapping
|
||||
|
||||
@staticmethod
|
||||
def _extract_fovea(row: pd.Series, x_key: str, y_key: str) -> Optional[Tuple[float, float]]:
|
||||
x_val = row.get(x_key)
|
||||
y_val = row.get(y_key)
|
||||
if pd.isna(x_val) or pd.isna(y_val):
|
||||
return None
|
||||
return float(x_val), float(y_val)
|
||||
@@ -1,383 +0,0 @@
|
||||
"""REFUGE optic disc / cup segmentation utilities."""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import math
|
||||
from dataclasses import dataclass
|
||||
from pathlib import Path
|
||||
from typing import Dict, Iterable, List, Optional, Sequence, Tuple
|
||||
|
||||
import numpy as np
|
||||
from PIL import Image
|
||||
import torch
|
||||
from torch import nn
|
||||
from torch.utils.data import DataLoader, Dataset
|
||||
from torchvision import transforms
|
||||
|
||||
from classes.refuge_preprocessing import RefugePreprocessing, RefugeSample
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Dataset helpers
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
|
||||
def _load_rgb(path: Path) -> Image.Image:
|
||||
img = Image.open(path)
|
||||
if img.mode != "RGB":
|
||||
img = img.convert("RGB")
|
||||
return img
|
||||
|
||||
|
||||
def _load_mask_array(path: Path) -> np.ndarray:
|
||||
mask_img = Image.open(path).convert("L")
|
||||
mask = np.array(mask_img, dtype=np.float32)
|
||||
# REFUGE masks encode disc/cup with different intensities; treat any
|
||||
# positive value as disc for coarse localisation.
|
||||
mask = np.where(mask > 0, 1.0, 0.0)
|
||||
return mask
|
||||
|
||||
|
||||
@dataclass
|
||||
class RefugeSegmentationSample:
|
||||
sample: RefugeSample
|
||||
image_path: Path
|
||||
mask_path: Path
|
||||
|
||||
|
||||
class RefugeSegmentationDataset(Dataset):
|
||||
"""Simple segmentation dataset returning tensors."""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
samples: Sequence[RefugeSegmentationSample],
|
||||
image_size: int = 512,
|
||||
) -> None:
|
||||
self.samples = list(samples)
|
||||
self.image_size = image_size
|
||||
self.to_tensor = transforms.ToTensor()
|
||||
|
||||
def __len__(self) -> int:
|
||||
return len(self.samples)
|
||||
|
||||
def __getitem__(self, idx: int) -> Dict[str, torch.Tensor]:
|
||||
rec = self.samples[idx]
|
||||
image = _load_rgb(rec.image_path)
|
||||
mask_arr = _load_mask_array(rec.mask_path)
|
||||
|
||||
if self.image_size is not None:
|
||||
image = image.resize((self.image_size, self.image_size), Image.BILINEAR)
|
||||
mask_img = Image.fromarray(mask_arr).resize(
|
||||
(self.image_size, self.image_size), Image.NEAREST
|
||||
)
|
||||
mask_arr = np.array(mask_img, dtype=np.float32)
|
||||
|
||||
image_tensor = self.to_tensor(image)
|
||||
mask_tensor = torch.from_numpy(mask_arr).unsqueeze(0) # [1,H,W]
|
||||
return {
|
||||
"image": image_tensor,
|
||||
"mask": mask_tensor,
|
||||
"sample_id": rec.sample.sample_id,
|
||||
}
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Model definition (lightweight U-Net)
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
|
||||
class DoubleConv(nn.Module):
|
||||
def __init__(self, in_channels: int, out_channels: int):
|
||||
super().__init__()
|
||||
self.net = nn.Sequential(
|
||||
nn.Conv2d(in_channels, out_channels, 3, padding=1, bias=False),
|
||||
nn.BatchNorm2d(out_channels),
|
||||
nn.ReLU(inplace=True),
|
||||
nn.Conv2d(out_channels, out_channels, 3, padding=1, bias=False),
|
||||
nn.BatchNorm2d(out_channels),
|
||||
nn.ReLU(inplace=True),
|
||||
)
|
||||
|
||||
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
||||
return self.net(x)
|
||||
|
||||
|
||||
class UNet(nn.Module):
|
||||
def __init__(self, in_channels: int = 3, base_channels: int = 64):
|
||||
super().__init__()
|
||||
self.enc1 = DoubleConv(in_channels, base_channels)
|
||||
self.enc2 = DoubleConv(base_channels, base_channels * 2)
|
||||
self.enc3 = DoubleConv(base_channels * 2, base_channels * 4)
|
||||
self.enc4 = DoubleConv(base_channels * 4, base_channels * 8)
|
||||
|
||||
self.pool = nn.MaxPool2d(2)
|
||||
self.bottleneck = DoubleConv(base_channels * 8, base_channels * 16)
|
||||
|
||||
self.up4 = nn.ConvTranspose2d(base_channels * 16, base_channels * 8, 2, stride=2)
|
||||
self.dec4 = DoubleConv(base_channels * 16, base_channels * 8)
|
||||
self.up3 = nn.ConvTranspose2d(base_channels * 8, base_channels * 4, 2, stride=2)
|
||||
self.dec3 = DoubleConv(base_channels * 8, base_channels * 4)
|
||||
self.up2 = nn.ConvTranspose2d(base_channels * 4, base_channels * 2, 2, stride=2)
|
||||
self.dec2 = DoubleConv(base_channels * 4, base_channels * 2)
|
||||
self.up1 = nn.ConvTranspose2d(base_channels * 2, base_channels, 2, stride=2)
|
||||
self.dec1 = DoubleConv(base_channels * 2, base_channels)
|
||||
|
||||
self.out = nn.Conv2d(base_channels, 1, 1)
|
||||
|
||||
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
||||
e1 = self.enc1(x)
|
||||
e2 = self.enc2(self.pool(e1))
|
||||
e3 = self.enc3(self.pool(e2))
|
||||
e4 = self.enc4(self.pool(e3))
|
||||
b = self.bottleneck(self.pool(e4))
|
||||
|
||||
d4 = self.up4(b)
|
||||
d4 = torch.cat([d4, e4], dim=1)
|
||||
d4 = self.dec4(d4)
|
||||
d3 = self.up3(d4)
|
||||
d3 = torch.cat([d3, e3], dim=1)
|
||||
d3 = self.dec3(d3)
|
||||
d2 = self.up2(d3)
|
||||
d2 = torch.cat([d2, e2], dim=1)
|
||||
d2 = self.dec2(d2)
|
||||
d1 = self.up1(d2)
|
||||
d1 = torch.cat([d1, e1], dim=1)
|
||||
d1 = self.dec1(d1)
|
||||
return self.out(d1)
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Segmentation manager
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
|
||||
class RefugeSegmentation:
|
||||
"""Train and run coarse-to-fine OD/OC segmentation for REFUGE."""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
preprocessing: RefugePreprocessing,
|
||||
model: Optional[nn.Module] = None,
|
||||
) -> None:
|
||||
self.preprocessing = preprocessing
|
||||
self.model = model or UNet()
|
||||
self.train_dataset: Optional[RefugeSegmentationDataset] = None
|
||||
self.val_dataset: Optional[RefugeSegmentationDataset] = None
|
||||
self.train_loader: Optional[DataLoader] = None
|
||||
self.val_loader: Optional[DataLoader] = None
|
||||
|
||||
# ------------------------------------------------------------------
|
||||
def build_datasets(
|
||||
self,
|
||||
image_size: int = 512,
|
||||
batch_size: int = 8,
|
||||
num_workers: int = 4,
|
||||
) -> None:
|
||||
manifest = self.preprocessing.build_manifest()
|
||||
|
||||
train_samples: List[RefugeSegmentationSample] = []
|
||||
val_samples: List[RefugeSegmentationSample] = []
|
||||
|
||||
for sample in manifest:
|
||||
if not sample.mask_path or not sample.mask_path.exists():
|
||||
continue
|
||||
rec = RefugeSegmentationSample(sample=sample, image_path=sample.image_path, mask_path=sample.mask_path)
|
||||
if sample.split == "train":
|
||||
train_samples.append(rec)
|
||||
elif sample.split in {"val", "validation"}:
|
||||
val_samples.append(rec)
|
||||
|
||||
if not val_samples:
|
||||
# Fall back to using a subset of training data for validation
|
||||
split = max(1, int(0.1 * len(train_samples)))
|
||||
val_samples = train_samples[:split]
|
||||
train_samples = train_samples[split:]
|
||||
|
||||
self.train_dataset = RefugeSegmentationDataset(train_samples, image_size=image_size)
|
||||
self.val_dataset = RefugeSegmentationDataset(val_samples, image_size=image_size)
|
||||
self.train_loader = DataLoader(
|
||||
self.train_dataset,
|
||||
batch_size=batch_size,
|
||||
shuffle=True,
|
||||
num_workers=num_workers,
|
||||
pin_memory=True,
|
||||
)
|
||||
self.val_loader = DataLoader(
|
||||
self.val_dataset,
|
||||
batch_size=batch_size,
|
||||
shuffle=False,
|
||||
num_workers=num_workers,
|
||||
pin_memory=True,
|
||||
)
|
||||
|
||||
# ------------------------------------------------------------------
|
||||
def train(
|
||||
self,
|
||||
epochs: int = 40,
|
||||
lr: float = 1e-3,
|
||||
weight_decay: float = 1e-5,
|
||||
device: Optional[str] = None,
|
||||
checkpoint_dir: Optional[Path] = None,
|
||||
) -> Dict[str, float]:
|
||||
if self.train_loader is None or self.val_loader is None:
|
||||
self.build_datasets()
|
||||
|
||||
device = device or ("cuda" if torch.cuda.is_available() else "cpu")
|
||||
self.model.to(device)
|
||||
criterion = nn.BCEWithLogitsLoss()
|
||||
optimizer = torch.optim.Adam(self.model.parameters(), lr=lr, weight_decay=weight_decay)
|
||||
|
||||
best_dice = 0.0
|
||||
history: Dict[str, float] = {}
|
||||
|
||||
for epoch in range(1, epochs + 1):
|
||||
print(f"[Seg] Processing epoch {epoch}/{epochs}")
|
||||
self.model.train()
|
||||
running_loss = 0.0
|
||||
for batch in self.train_loader: # type: ignore[arg-type]
|
||||
images = batch["image"].to(device)
|
||||
masks = batch["mask"].to(device)
|
||||
optimizer.zero_grad()
|
||||
logits = self.model(images)
|
||||
loss = criterion(logits, masks)
|
||||
loss.backward()
|
||||
optimizer.step()
|
||||
running_loss += loss.item() * images.size(0)
|
||||
|
||||
train_loss = running_loss / len(self.train_loader.dataset) # type: ignore[arg-type]
|
||||
val_metrics = self.evaluate(device=device)
|
||||
history[f"epoch_{epoch}_loss"] = train_loss
|
||||
history[f"epoch_{epoch}_dice"] = val_metrics.get("dice", float("nan"))
|
||||
|
||||
if val_metrics.get("dice", 0.0) > best_dice:
|
||||
best_dice = val_metrics["dice"]
|
||||
if checkpoint_dir is not None:
|
||||
checkpoint_dir.mkdir(parents=True, exist_ok=True)
|
||||
torch.save(self.model.state_dict(), checkpoint_dir / "refuge_segmentation_best.pt")
|
||||
|
||||
return {"best_dice": best_dice, **history}
|
||||
|
||||
# ------------------------------------------------------------------
|
||||
def evaluate(self, split: str = "val", device: Optional[str] = None) -> Dict[str, float]:
|
||||
if split != "val":
|
||||
raise ValueError("Only validation split supported currently")
|
||||
if self.val_loader is None:
|
||||
self.build_datasets()
|
||||
|
||||
device = device or ("cuda" if torch.cuda.is_available() else "cpu")
|
||||
self.model.to(device)
|
||||
self.model.eval()
|
||||
|
||||
dices: List[float] = []
|
||||
criterion = nn.BCEWithLogitsLoss()
|
||||
losses: List[float] = []
|
||||
|
||||
with torch.no_grad():
|
||||
for batch in self.val_loader: # type: ignore[arg-type]
|
||||
images = batch["image"].to(device)
|
||||
masks = batch["mask"].to(device)
|
||||
logits = self.model(images)
|
||||
loss = criterion(logits, masks)
|
||||
losses.append(loss.item() * images.size(0))
|
||||
probs = torch.sigmoid(logits)
|
||||
preds = (probs > 0.5).float()
|
||||
dice = self._dice_coefficient(preds, masks)
|
||||
dices.extend(dice)
|
||||
|
||||
mean_dice = float(np.mean(dices)) if dices else 0.0
|
||||
mean_loss = float(np.sum(losses) / len(self.val_loader.dataset)) # type: ignore[arg-type]
|
||||
return {"dice": mean_dice, "loss": mean_loss}
|
||||
|
||||
# ------------------------------------------------------------------
|
||||
def predict_mask(self, sample: RefugeSample, device: Optional[str] = None) -> torch.Tensor:
|
||||
if self.train_dataset is None:
|
||||
self.build_datasets()
|
||||
device = device or ("cuda" if torch.cuda.is_available() else "cpu")
|
||||
self.model.to(device)
|
||||
self.model.eval()
|
||||
|
||||
image = _load_rgb(sample.image_path)
|
||||
original_size = image.size # (width, height)
|
||||
image_resized = image.resize((self.train_dataset.image_size, self.train_dataset.image_size), Image.BILINEAR) # type: ignore[union-attr]
|
||||
tensor = transforms.ToTensor()(image_resized).unsqueeze(0).to(device)
|
||||
|
||||
with torch.no_grad():
|
||||
logits = self.model(tensor)
|
||||
mask_resized = torch.sigmoid(logits)[0, 0]
|
||||
|
||||
mask_np = mask_resized.cpu().numpy()
|
||||
mask_np = (mask_np > 0.5).astype(np.float32)
|
||||
mask_img = Image.fromarray(mask_np)
|
||||
mask_img = mask_img.resize(original_size, Image.NEAREST)
|
||||
return torch.from_numpy(np.array(mask_img, dtype=np.float32))
|
||||
|
||||
def infer_disc_geometry(
|
||||
self,
|
||||
sample: RefugeSample,
|
||||
scale: float = 2.5,
|
||||
) -> Dict[str, float]:
|
||||
if sample.mask_path and sample.mask_path.exists():
|
||||
mask = _load_mask_array(sample.mask_path)
|
||||
else:
|
||||
mask = self.predict_mask(sample).numpy()
|
||||
|
||||
coords = np.argwhere(mask > 0.5)
|
||||
if coords.size == 0:
|
||||
raise RuntimeError(f"Unable to locate disc for sample {sample.sample_id}")
|
||||
|
||||
ys, xs = coords[:, 0], coords[:, 1]
|
||||
centre_x = float(xs.mean())
|
||||
centre_y = float(ys.mean())
|
||||
width = float(xs.max() - xs.min())
|
||||
height = float(ys.max() - ys.min())
|
||||
diameter = max(width, height)
|
||||
radius = diameter / 2.0
|
||||
crop_radius = radius * scale
|
||||
return {
|
||||
"centre_x": centre_x,
|
||||
"centre_y": centre_y,
|
||||
"radius": radius,
|
||||
"crop_radius": crop_radius,
|
||||
"crop_size": crop_radius * 2.0,
|
||||
}
|
||||
|
||||
def batch_crops(
|
||||
self,
|
||||
samples: Iterable[RefugeSample],
|
||||
scale: float = 2.5,
|
||||
output_dir: Optional[Path] = None,
|
||||
size: int = 256,
|
||||
) -> Dict[str, Path]:
|
||||
output_paths: Dict[str, Path] = {}
|
||||
if output_dir is not None:
|
||||
output_dir.mkdir(parents=True, exist_ok=True)
|
||||
|
||||
for sample in samples:
|
||||
geom = self.infer_disc_geometry(sample, scale=scale)
|
||||
image = _load_rgb(sample.image_path)
|
||||
cx, cy = geom["centre_x"], geom["centre_y"]
|
||||
r = geom["crop_radius"]
|
||||
left = max(0.0, cx - r)
|
||||
upper = max(0.0, cy - r)
|
||||
right = min(image.width, cx + r)
|
||||
lower = min(image.height, cy + r)
|
||||
crop = image.crop((left, upper, right, lower)).resize((size, size), Image.BILINEAR)
|
||||
if output_dir is not None:
|
||||
out_path = output_dir / f"{sample.sample_id}_crop.png"
|
||||
crop.save(out_path)
|
||||
output_paths[sample.sample_id] = out_path
|
||||
return output_paths
|
||||
|
||||
# ------------------------------------------------------------------
|
||||
@staticmethod
|
||||
def _dice_coefficient(preds: torch.Tensor, targets: torch.Tensor) -> List[float]:
|
||||
eps = 1e-6
|
||||
dices = []
|
||||
preds = preds.view(preds.size(0), -1)
|
||||
targets = targets.view(targets.size(0), -1)
|
||||
for p, t in zip(preds, targets):
|
||||
intersection = float((p * t).sum().item())
|
||||
union = float(p.sum().item() + t.sum().item())
|
||||
dice = (2.0 * intersection + eps) / (union + eps)
|
||||
dices.append(dice)
|
||||
return dices
|
||||
@@ -1,149 +0,0 @@
|
||||
# tower_watcher.py
|
||||
import matplotlib.pyplot as plt
|
||||
|
||||
# tower_watcher.py
|
||||
import matplotlib.pyplot as plt
|
||||
|
||||
|
||||
class TowerWatcher:
|
||||
"""
|
||||
Live monitor:
|
||||
- Cumulative batch-level: loss & accuracy per batch across all epochs.
|
||||
- Epoch batch-level: loss & accuracy per batch within the current epoch (resets each epoch).
|
||||
- TP/FP/TN/FN bar charts per tower, one chart each, new group each epoch.
|
||||
"""
|
||||
|
||||
def __init__(self):
|
||||
plt.ion()
|
||||
# 2 line plots (cum loss, cum acc), 2 line plots (epoch loss, epoch acc), 3 bar plots
|
||||
self.fig, self.axs = plt.subplots(7, 1, figsize=(10, 28))
|
||||
self.reset()
|
||||
|
||||
def reset(self):
|
||||
# Cumulative batch-level
|
||||
self.global_batches = []
|
||||
self.loss_cum = {"fusion": [], "image": [], "meta": []}
|
||||
self.acc_cum = {"fusion": [], "image": [], "meta": []}
|
||||
# Epoch batch-level
|
||||
self.epoch_batches = []
|
||||
self.loss_epoch_batch = {"fusion": [], "image": [], "meta": []}
|
||||
self.acc_epoch_batch = {"fusion": [], "image": [], "meta": []}
|
||||
# Epoch markers for cum plots
|
||||
self.epoch_markers = []
|
||||
# Stats per epoch for bars
|
||||
self.epoch_stats = {"fusion": [], "image": [], "meta": []}
|
||||
# Track current epoch
|
||||
self.current_epoch = -1
|
||||
|
||||
def on_epoch_start(self, epoch):
|
||||
# mark epoch boundary in cumulative
|
||||
x = self.global_batches[-1] + 1 if self.global_batches else 0
|
||||
self.epoch_markers.append(x)
|
||||
# reset epoch batch-level data
|
||||
self.epoch_batches = []
|
||||
for d in [self.loss_epoch_batch, self.acc_epoch_batch]:
|
||||
for k in d:
|
||||
d[k].clear()
|
||||
self.current_epoch = epoch
|
||||
|
||||
def on_batch_end(self, idx, stats: dict):
|
||||
# Cumulative
|
||||
self.global_batches.append(len(self.global_batches) + 1)
|
||||
for key, lk, ak in [
|
||||
("fusion", "loss_f", "acc_f"),
|
||||
("image", "loss_i", "acc_i"),
|
||||
("meta", "loss_m", "acc_m"),
|
||||
]:
|
||||
self.loss_cum[key].append(stats.get(lk, 0))
|
||||
self.acc_cum[key].append(stats.get(ak, 0))
|
||||
# Epoch-level
|
||||
self.epoch_batches.append(len(self.epoch_batches) + 1)
|
||||
for key, lk, ak in [
|
||||
("fusion", "loss_f", "acc_f"),
|
||||
("image", "loss_i", "acc_i"),
|
||||
("meta", "loss_m", "acc_m"),
|
||||
]:
|
||||
self.loss_epoch_batch[key].append(stats.get(lk, 0))
|
||||
self.acc_epoch_batch[key].append(stats.get(ak, 0))
|
||||
# redraw
|
||||
self._draw_batch_plots()
|
||||
|
||||
def on_epoch_end(self, epoch, stats: dict):
|
||||
# record per-epoch TP/FP/TN/FN
|
||||
for key in ["fusion", "image", "meta"]:
|
||||
self.epoch_stats[key].append(
|
||||
{
|
||||
"tp": stats.get("tp", 0),
|
||||
"fp": stats.get("fp", 0),
|
||||
"tn": stats.get("tn", 0),
|
||||
"fn": stats.get("fn", 0),
|
||||
}
|
||||
)
|
||||
self._draw_epoch_bars()
|
||||
|
||||
def _draw_batch_plots(self):
|
||||
# Cumulative Loss
|
||||
ax = self.axs[0]
|
||||
ax.clear()
|
||||
ax.plot(self.global_batches, self.loss_cum["fusion"], label="Fusion")
|
||||
ax.plot(self.global_batches, self.loss_cum["image"], label="Image Tower")
|
||||
ax.plot(self.global_batches, self.loss_cum["meta"], label="MD Tower")
|
||||
for x in self.epoch_markers:
|
||||
ax.axvline(x=x, color="gray", linestyle="--")
|
||||
ax.set_ylabel("Cumulative Loss")
|
||||
ax.legend()
|
||||
|
||||
# Epoch Loss
|
||||
ax = self.axs[1]
|
||||
ax.clear()
|
||||
ax.plot(self.epoch_batches, self.loss_epoch_batch["fusion"], label="Fusion")
|
||||
ax.plot(self.epoch_batches, self.loss_epoch_batch["image"], label="Image Tower")
|
||||
ax.plot(self.epoch_batches, self.loss_epoch_batch["meta"], label="MD Tower")
|
||||
ax.set_ylabel(f"Epoch {self.current_epoch+1} Loss")
|
||||
ax.set_xlabel("Batch (Epoch)")
|
||||
ax.legend()
|
||||
|
||||
# Cumulative Accuracy
|
||||
ax = self.axs[2]
|
||||
ax.clear()
|
||||
ax.plot(self.global_batches, self.acc_cum["fusion"], label="Fusion")
|
||||
ax.plot(self.global_batches, self.acc_cum["image"], label="Image Tower")
|
||||
ax.plot(self.global_batches, self.acc_cum["meta"], label="MD Tower")
|
||||
for x in self.epoch_markers:
|
||||
ax.axvline(x=x, color="gray", linestyle="--")
|
||||
ax.set_ylabel("Cumulative Accuracy")
|
||||
ax.legend()
|
||||
|
||||
# Epoch Accuracy
|
||||
ax = self.axs[3]
|
||||
ax.clear()
|
||||
ax.plot(self.epoch_batches, self.acc_epoch_batch["fusion"], label="Fusion")
|
||||
ax.plot(self.epoch_batches, self.acc_epoch_batch["image"], label="Image Tower")
|
||||
ax.plot(self.epoch_batches, self.acc_epoch_batch["meta"], label="MD Tower")
|
||||
ax.set_ylabel(f"Epoch {self.current_epoch+1} Accuracy")
|
||||
ax.set_xlabel("Batch (Epoch)")
|
||||
ax.legend()
|
||||
|
||||
plt.pause(0.01)
|
||||
|
||||
def _draw_epoch_bars(self):
|
||||
# Bar charts per tower
|
||||
for i, key in enumerate(["fusion", "image", "meta"]):
|
||||
ax = self.axs[4 + i]
|
||||
ax.clear()
|
||||
data = self.epoch_stats[key]
|
||||
epochs = list(range(1, len(data) + 1))
|
||||
tp = [d["tp"] for d in data]
|
||||
fp = [d["fp"] for d in data]
|
||||
tn = [d["tn"] for d in data]
|
||||
fn = [d["fn"] for d in data]
|
||||
width = 0.2
|
||||
ax.bar([e - width for e in epochs], tp, width, label="TP")
|
||||
ax.bar(epochs, fp, width, label="FP")
|
||||
ax.bar([e + width for e in epochs], tn, width, label="TN")
|
||||
ax.bar([e + 2 * width for e in epochs], fn, width, label="FN")
|
||||
ax.set_title(f"{key.title()} Tower Stats")
|
||||
ax.set_xlabel("Epoch")
|
||||
ax.set_ylabel("Count")
|
||||
ax.legend()
|
||||
plt.pause(0.01)
|
||||
@@ -3,7 +3,7 @@ from __future__ import annotations
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
|
||||
from classes.SE_attention import SEBlock, SEGateLogger
|
||||
from classes.v2.SE_attention import SEBlock, SEGateLogger
|
||||
|
||||
|
||||
class Bridge(nn.Module):
|
||||
|
||||
+24
-3
@@ -10,9 +10,30 @@ import torch
|
||||
from PIL import Image, ImageDraw
|
||||
from torchvision import transforms
|
||||
|
||||
from classes.geometry_features import compute_geometry_features, disc_cup_from_mask_image
|
||||
from classes.refuge_classification import _geometry_from_mask
|
||||
from classes.unet_segmenter import UNetSegmenter
|
||||
from classes.v2.geometry_features import compute_geometry_features, disc_cup_from_mask_image
|
||||
from classes.v2.unet_segmenter import UNetSegmenter
|
||||
|
||||
|
||||
def _geometry_from_mask(mask: np.ndarray, scale: float) -> Dict:
|
||||
mask = np.asarray(mask) > 0
|
||||
coords = np.argwhere(mask)
|
||||
if coords.size == 0:
|
||||
raise RuntimeError("Empty mask; cannot derive geometry")
|
||||
ys, xs = coords[:, 0], coords[:, 1]
|
||||
centre_x = float(xs.mean())
|
||||
centre_y = float(ys.mean())
|
||||
width = float(xs.max() - xs.min())
|
||||
height = float(ys.max() - ys.min())
|
||||
diameter = max(width, height)
|
||||
radius = diameter / 2.0
|
||||
crop_radius = radius * scale
|
||||
return {
|
||||
"centre_x": centre_x,
|
||||
"centre_y": centre_y,
|
||||
"radius": radius,
|
||||
"crop_radius": crop_radius,
|
||||
"crop_size": crop_radius * 2.0,
|
||||
}
|
||||
|
||||
|
||||
class UNetImageCropper:
|
||||
|
||||
@@ -219,6 +219,15 @@ def _set_single_phase(model: SingleEyeHT, phase: str) -> None:
|
||||
# Ablation modes have no fusion bridge; fused_warmup is meaningless — treat as tower_warmup
|
||||
if bridge_mode in ("image_only", "metadata_only") and phase == "fused_warmup":
|
||||
phase = "tower_warmup"
|
||||
if phase == "md_warmup":
|
||||
_set_requires_grad(model.img_tower, False)
|
||||
_set_requires_grad(model.md_tower, True)
|
||||
_set_requires_grad(model.bridge.classifier_img, False)
|
||||
_set_requires_grad(model.bridge.classifier_md, True)
|
||||
_set_requires_grad(model.bridge.W_img, False)
|
||||
_set_requires_grad(model.bridge.W_md, False)
|
||||
_set_requires_grad(model.bridge.classifier_fused, False)
|
||||
return
|
||||
if phase == "tower_warmup":
|
||||
_set_requires_grad(model.img_tower, bridge_mode != "metadata_only")
|
||||
_set_requires_grad(model.md_tower, bridge_mode != "image_only")
|
||||
@@ -282,12 +291,27 @@ def train_single_epoch(
|
||||
x = batch.get("image_1")
|
||||
m = batch.get("matrix_1")
|
||||
y = batch.get("label_1")
|
||||
if phase == "md_warmup":
|
||||
if not torch.is_tensor(m):
|
||||
continue
|
||||
m = m.to(device)
|
||||
y = _to_label_tensor(y, device)
|
||||
md_feats = model.md_tower(m)
|
||||
logits = model.bridge.classifier_md(md_feats)
|
||||
loss = F.cross_entropy(logits, y)
|
||||
opt.zero_grad(); loss.backward(); opt.step()
|
||||
bs = y.shape[0]
|
||||
total_loss += float(loss.item()) * bs
|
||||
total_correct += int((logits.argmax(1) == y).sum())
|
||||
total_n += bs
|
||||
continue
|
||||
if not torch.is_tensor(x) or not torch.is_tensor(m):
|
||||
continue
|
||||
x = x.to(device)
|
||||
m = m.to(device)
|
||||
y = _to_label_tensor(y, device)
|
||||
bridge_mode = model.bridge.mode
|
||||
|
||||
img_feats = None if bridge_mode == "metadata_only" else model.img_tower(x)
|
||||
md_feats = None if bridge_mode == "image_only" else model.md_tower(m)
|
||||
|
||||
|
||||
@@ -33,10 +33,21 @@ def _nearest_pachy_key(x: float) -> int:
|
||||
return int(_PACHY_KEYS[idx])
|
||||
|
||||
|
||||
# Ratio derived from patients with both Pneumatic and Perkins readings (n=41, OD+OS combined).
|
||||
# Pneumatic / Perkins mean ratio = 1.158; applied to Perkins-only rows to put them on the
|
||||
# Pneumatic scale before IOP_corr is computed.
|
||||
_PERKINS_TO_PNEUMATIC_RATIO: float = 1.158
|
||||
|
||||
|
||||
def _pick_iop(row: pd.Series) -> float:
|
||||
"""Prefer Pneumatic, else Perkins; may return NaN."""
|
||||
raw = row["Pneumatic"] if not pd.isna(row.get("Pneumatic", np.nan)) else row.get("Perkins", np.nan)
|
||||
return float(raw) if not pd.isna(raw) else np.nan
|
||||
"""Prefer Pneumatic; scale Perkins to Pneumatic scale if Pneumatic is absent."""
|
||||
pneumatic = row.get("Pneumatic", np.nan)
|
||||
if not pd.isna(pneumatic):
|
||||
return float(pneumatic)
|
||||
perkins = row.get("Perkins", np.nan)
|
||||
if not pd.isna(perkins):
|
||||
return float(perkins) * _PERKINS_TO_PNEUMATIC_RATIO
|
||||
return np.nan
|
||||
|
||||
|
||||
def _correct_iop(raw_iop: float, pachy: float) -> float:
|
||||
|
||||
@@ -7,8 +7,8 @@ import torch
|
||||
from torch import nn
|
||||
from torchvision import transforms
|
||||
|
||||
from classes.backbones import BACKBONES, list_names, load_backbone_weights
|
||||
from classes.SE_attention import SEBlock
|
||||
from classes.v2.backbones import BACKBONES, list_names, load_backbone_weights
|
||||
from classes.v2.SE_attention import SEBlock
|
||||
from classes.v2.data_bundle import DataBundle
|
||||
|
||||
|
||||
|
||||
@@ -7,7 +7,7 @@ from PIL import Image
|
||||
|
||||
from torchvision import transforms
|
||||
|
||||
from classes.backbones import BACKBONES
|
||||
from classes.v2.backbones import BACKBONES
|
||||
|
||||
|
||||
IMAGENET_MEAN: Tuple[float, float, float] = (0.485, 0.456, 0.406)
|
||||
|
||||
+60
-14
@@ -206,6 +206,9 @@ class V2HyperTower:
|
||||
help="Single-eye model tower warmup (overrides --warmup-tower-epochs).")
|
||||
ap.add_argument("--single-warmup-fused-epochs", type=int, default=None,
|
||||
help="Single-eye model fused warmup (overrides --warmup-fused-epochs).")
|
||||
ap.add_argument("--warmup-md-epochs", type=int, default=0,
|
||||
help="MD-only warmup epochs before tower warmup. Trains only md_tower + "
|
||||
"classifier_md (no CNN forward pass, so 50-100 epochs is cheap).")
|
||||
ap.add_argument("--bilat-warmup-tower-epochs", type=int, default=None,
|
||||
help="Bilateral model tower warmup (overrides --warmup-tower-epochs).")
|
||||
ap.add_argument("--bilat-warmup-fused-epochs", type=int, default=None,
|
||||
@@ -424,7 +427,8 @@ class V2HyperTower:
|
||||
if getattr(args, "single_warmup_fused_epochs", None) is not None
|
||||
else int(_global_warmup_fused) if _global_warmup_fused is not None else 2
|
||||
)
|
||||
_total_epochs = _warmup_tower + _warmup_fused + int(args.epochs) + fusion_epochs
|
||||
_warmup_md = int(getattr(args, "warmup_md_epochs", 0))
|
||||
_total_epochs = _warmup_md + _warmup_tower + _warmup_fused + int(args.epochs) + fusion_epochs
|
||||
|
||||
# sample IDs depend on mode: single uses eye IDs, others use patient IDs
|
||||
if tower_mode in ("single", "classic"):
|
||||
@@ -572,6 +576,7 @@ class V2HyperTower:
|
||||
"run_id": run_name,
|
||||
"backbone": args.backbone,
|
||||
"epochs": args.epochs,
|
||||
"warmup_md_epochs": getattr(args, "warmup_md_epochs", 0),
|
||||
"warmup_tower_epochs": args.warmup_tower_epochs,
|
||||
"warmup_fused_epochs": args.warmup_fused_epochs,
|
||||
"single_warmup_tower_epochs": args.single_warmup_tower_epochs,
|
||||
@@ -659,6 +664,7 @@ class V2HyperTower:
|
||||
if getattr(args, "bilat_warmup_fused_epochs", None) is not None
|
||||
else int(global_warmup_fused) if global_warmup_fused is not None else 3
|
||||
)
|
||||
single_warmup_md = int(getattr(args, "warmup_md_epochs", 0)) if run_single else 0
|
||||
if not run_single:
|
||||
single_warmup_tower = 0
|
||||
single_warmup_fused = 0
|
||||
@@ -666,7 +672,7 @@ class V2HyperTower:
|
||||
bilat_warmup_tower = 0
|
||||
bilat_warmup_fused = 0
|
||||
main_epochs = int(args.epochs)
|
||||
total_single_epochs = (single_warmup_tower + single_warmup_fused + main_epochs) if run_single else 0
|
||||
total_single_epochs = (single_warmup_md + single_warmup_tower + single_warmup_fused + main_epochs) if run_single else 0
|
||||
total_bilat_epochs = (bilat_warmup_tower + bilat_warmup_fused + main_epochs) if run_bilat else 0
|
||||
total_epochs = max(total_single_epochs, total_bilat_epochs)
|
||||
|
||||
@@ -744,6 +750,7 @@ class V2HyperTower:
|
||||
train_single_loader = None
|
||||
train_eval_loader = None # non-shuffled, no sampler — for per-epoch train logging
|
||||
train_bilat_loader = None
|
||||
md_only_loader = None # image-free loader for md_warmup phase
|
||||
if run_single:
|
||||
single_sampler = build_balanced_sampler(eye_train) if use_balanced else None
|
||||
train_single_loader = make_loader(
|
||||
@@ -761,6 +768,20 @@ class V2HyperTower:
|
||||
shuffle=False,
|
||||
**loader_kw,
|
||||
)
|
||||
if single_warmup_md > 0:
|
||||
# MD-only loader: drop image_1 so PIL never opens files during md_warmup.
|
||||
# Always use balanced sampling for md_warmup — MD features alone are weaker
|
||||
# than images and collapse to majority class without class balancing.
|
||||
slots_md_only = {k: v for k, v in slots_eye.items() if k != "image_1"}
|
||||
md_warmup_sampler = single_sampler if single_sampler is not None else build_balanced_sampler(eye_train)
|
||||
md_only_loader = make_loader(
|
||||
eye_train, slots_md_only,
|
||||
image_transform=None,
|
||||
image_preprocessor=None,
|
||||
shuffle=True,
|
||||
sampler=md_warmup_sampler,
|
||||
**loader_kw,
|
||||
)
|
||||
if run_bilat:
|
||||
bilat_sampler = build_balanced_sampler(bilat_train) if use_balanced else None
|
||||
train_bilat_loader = make_loader(
|
||||
@@ -907,7 +928,7 @@ class V2HyperTower:
|
||||
print(
|
||||
f" [fold {fold+1}] single_train_n={len(eye_train)} (eye-level) "
|
||||
f"val_n={len(bilat_val)} "
|
||||
f"single_warmup={single_warmup_tower}+{single_warmup_fused} total={total_single_epochs}",
|
||||
f"single_warmup=md{single_warmup_md}+twr{single_warmup_tower}+fus{single_warmup_fused} total={total_single_epochs}",
|
||||
flush=True,
|
||||
)
|
||||
else:
|
||||
@@ -919,17 +940,20 @@ class V2HyperTower:
|
||||
)
|
||||
|
||||
# ---- epoch loop ------------------------------------------------
|
||||
_prev_phase_single = "inactive" # used to detect md_warmup → next phase transition
|
||||
for epoch in range(total_epochs):
|
||||
if not run_single:
|
||||
phase_single, main_epoch_single, single_active = "inactive", 0, False
|
||||
elif epoch < single_warmup_tower:
|
||||
elif epoch < single_warmup_md:
|
||||
phase_single, main_epoch_single, single_active = "md_warmup", 0, True
|
||||
elif epoch < (single_warmup_md + single_warmup_tower):
|
||||
phase_single, main_epoch_single, single_active = "tower_warmup", 0, True
|
||||
elif epoch < (single_warmup_tower + single_warmup_fused):
|
||||
elif epoch < (single_warmup_md + single_warmup_tower + single_warmup_fused):
|
||||
phase_single, main_epoch_single, single_active = "fused_warmup", 0, True
|
||||
elif epoch < total_single_epochs:
|
||||
phase_single, main_epoch_single, single_active = (
|
||||
"main",
|
||||
epoch - single_warmup_tower - single_warmup_fused + 1,
|
||||
epoch - single_warmup_md - single_warmup_tower - single_warmup_fused + 1,
|
||||
True,
|
||||
)
|
||||
else:
|
||||
@@ -951,8 +975,9 @@ class V2HyperTower:
|
||||
phase_bilat, main_epoch_bilat, bilat_active = "done", main_epochs, False
|
||||
|
||||
if run_single and single_active:
|
||||
_active_loader = md_only_loader if phase_single == "md_warmup" else train_single_loader
|
||||
sl_loss, sl_acc = train_single_epoch(
|
||||
single, train_single_loader, opt_single, device,
|
||||
single, _active_loader, opt_single, device,
|
||||
phase=phase_single, bcd_prob=float(args.bcd_prob),
|
||||
)
|
||||
else:
|
||||
@@ -966,7 +991,9 @@ class V2HyperTower:
|
||||
else:
|
||||
bl_loss, bl_acc = nan, nan
|
||||
|
||||
if run_single and tower_mode == "single":
|
||||
_skip_val_eval = (phase_single == "md_warmup")
|
||||
|
||||
if run_single and tower_mode == "single" and not _skip_val_eval:
|
||||
y_cl, p_cl, p_cl_img, p_cl_md = collect_probs_single_components(
|
||||
single, val_loader, device, aggregate_patient=False
|
||||
)
|
||||
@@ -980,7 +1007,7 @@ class V2HyperTower:
|
||||
en_acc = en_auc = nan
|
||||
en_n = 0
|
||||
en_acc_img = en_acc_md = en_auc_img = en_auc_md = nan
|
||||
elif run_single and tower_mode == "ensemble":
|
||||
elif run_single and tower_mode == "ensemble" and not _skip_val_eval:
|
||||
(y_en,
|
||||
_p_en_f_od, _p_en_i_od, _p_en_m_od,
|
||||
_p_en_f_os, _p_en_i_os, _p_en_m_os,
|
||||
@@ -1010,7 +1037,7 @@ class V2HyperTower:
|
||||
cl_acc_img = cl_acc_md = en_acc_img = en_acc_md = nan
|
||||
cl_auc_img = cl_auc_md = en_auc_img = en_auc_md = nan
|
||||
|
||||
if run_bilat:
|
||||
if run_bilat and not _skip_val_eval:
|
||||
y_bi, p_bi, p_bi_img, p_bi_md = collect_probs_bilateral_components(
|
||||
bilateral, val_loader, device
|
||||
)
|
||||
@@ -1034,7 +1061,7 @@ class V2HyperTower:
|
||||
p_cl_h = p_cl_h_img = p_cl_h_md = _z2
|
||||
p_en_h = p_en_h_img = p_en_h_md = _z2
|
||||
|
||||
if holdout_loader is not None:
|
||||
if holdout_loader is not None and not _skip_val_eval:
|
||||
if run_single and tower_mode == "single":
|
||||
y_cl_h, p_cl_h, p_cl_h_img, p_cl_h_md = collect_probs_single_components(
|
||||
single, holdout_loader, device, aggregate_patient=False
|
||||
@@ -1092,7 +1119,7 @@ class V2HyperTower:
|
||||
tr_fe_corr = tr_fe_err = tr_n = 0
|
||||
y_tr = np.array([], dtype=np.int64)
|
||||
p_tr_f = p_tr_i = p_tr_m = np.zeros((0, num_classes), dtype=np.float32)
|
||||
if run_single and train_eval_loader is not None:
|
||||
if run_single and train_eval_loader is not None and not _skip_val_eval:
|
||||
y_tr, p_tr_f, p_tr_i, p_tr_m, tr_ids = collect_probs_eye_level(
|
||||
single, train_eval_loader, device, return_ids=True
|
||||
)
|
||||
@@ -1284,16 +1311,33 @@ class V2HyperTower:
|
||||
**cm_row,
|
||||
}, optional_cols=epoch_fields)
|
||||
|
||||
# ---- md_warmup progress bar (replaces per-epoch print) --------
|
||||
if phase_single == "md_warmup":
|
||||
_bar_w = 30
|
||||
_filled = int(_bar_w * (epoch + 1) / single_warmup_md)
|
||||
_bar = "#" * _filled + "-" * (_bar_w - _filled)
|
||||
_bar_msg = (
|
||||
f" [fold {fold+1}] md_warmup [{_bar}] "
|
||||
f"{epoch + 1}/{single_warmup_md} loss={sl_loss:.4f}"
|
||||
)
|
||||
print(f"\r{_bar_msg}", end="", flush=True)
|
||||
fold_logger.info(_bar_msg)
|
||||
_prev_phase_single = phase_single
|
||||
continue # skip normal log block entirely
|
||||
|
||||
if _prev_phase_single == "md_warmup":
|
||||
print() # seal the progress bar line
|
||||
|
||||
if args.log_every > 0 and (epoch + 1) % args.log_every == 0:
|
||||
hld_auc = target_holdout_single_auc if run_single else bi_auc_h
|
||||
hld_suffix = f" hld_auc={hld_auc:.4f}" if holdout_loader is not None else ""
|
||||
|
||||
# Human-readable phase progress for console logs.
|
||||
if phase_single == "tower_warmup":
|
||||
single_phase_epoch = epoch + 1
|
||||
single_phase_epoch = epoch - single_warmup_md + 1
|
||||
single_phase_total = single_warmup_tower
|
||||
elif phase_single == "fused_warmup":
|
||||
single_phase_epoch = epoch - single_warmup_tower + 1
|
||||
single_phase_epoch = epoch - single_warmup_md - single_warmup_tower + 1
|
||||
single_phase_total = single_warmup_fused
|
||||
else:
|
||||
single_phase_epoch = main_epoch_single
|
||||
@@ -1343,6 +1387,8 @@ class V2HyperTower:
|
||||
print(msg, flush=True)
|
||||
fold_logger.info(msg)
|
||||
|
||||
_prev_phase_single = phase_single
|
||||
|
||||
fold_logger.close()
|
||||
|
||||
if args.save_checkpoints:
|
||||
|
||||
Reference in New Issue
Block a user