File size: 9,353 Bytes
f621d73 75c3625 f621d73 75c3625 f621d73 75c3625 f621d73 75c3625 f621d73 75c3625 f621d73 75c3625 f621d73 75c3625 f621d73 75c3625 f621d73 75c3625 f621d73 75c3625 f621d73 75c3625 f621d73 75c3625 f621d73 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 | from typing import Any, Optional
import numpy as np
import torch
from torch import nn, optim
import lightning.pytorch as pl
import torchvision.models.video as tvmv
import sklearn.metrics as skm
class SyntaxLightningModule(pl.LightningModule):
"""
LightningModule for training the 3D backbone on SYNTAX score.
Architecture:
- backbone: ResNet3D (r3d_18) from torchvision
- output linear layer with two units
[0] binary classification logit for significant disease
[1] regression output for SYNTAX score (log1p)
Training modes:
- pretrain (weight_path is None):
freeze the backbone and train only the final fc layer
- finetune (weight_path is set):
load checkpoint weights and fine-tune the full network
"""
def __init__(
self,
num_classes: int,
lr: float,
weight_decay: float = 0.0,
max_epochs: Optional[int] = None,
weight_path: Optional[str] = None,
sigma_a: float = 0.0,
sigma_b: float = 1.0,
**kwargs,
):
super().__init__()
self.save_hyperparameters()
self.num_classes = int(num_classes)
self.lr = float(lr)
self.weight_decay = float(weight_decay)
self.max_epochs = max_epochs
self.weight_path = weight_path
self.sigma_a = float(sigma_a)
self.sigma_b = float(sigma_b)
self.model = tvmv.r3d_18(weights=tvmv.R3D_18_Weights.DEFAULT)
in_features = self.model.fc.in_features
self.model.fc = nn.Linear(in_features=in_features, out_features=self.num_classes, bias=True)
if self.weight_path is not None:
self._load_backbone_weights(self.weight_path)
self.loss_clf = nn.BCEWithLogitsLoss(reduction="none")
self.loss_reg = nn.MSELoss(reduction="none")
self._y_true = []
self._y_prob = []
self._y_pred = []
self._t_true = []
self._t_pred = []
def _load_backbone_weights(self, weight_path: str) -> None:
"""
Load backbone weights from either:
- a Lightning checkpoint (dict with a 'state_dict' key)
- or a raw state_dict saved via model.state_dict() (.pt/.pth)
Prints the source type and key statistics.
"""
obj = torch.load(weight_path, map_location="cpu", weights_only=False)
if isinstance(obj, dict) and "state_dict" in obj:
state_dict = obj["state_dict"]
state_dict = {k.replace("model.", ""): v for k, v in state_dict.items()}
src_type = "lightning_checkpoint"
else:
state_dict = obj
src_type = "raw_state_dict"
incompatible = self.model.load_state_dict(state_dict, strict=False)
loaded_keys = [k for k in state_dict.keys() if k not in incompatible.missing_keys]
print(
f"[Backbone] Loaded weights from '{weight_path}' "
f"(type={src_type}): {len(loaded_keys)} params, "
f"missing={len(incompatible.missing_keys)}, "
f"unexpected={len(incompatible.unexpected_keys)}"
)
if incompatible.missing_keys:
print(f"[Backbone] Missing keys example: {incompatible.missing_keys[:5]}")
if incompatible.unexpected_keys:
print(f"[Backbone] Unexpected keys example: {incompatible.unexpected_keys[:5]}")
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""
Input:
x: (B, C, T, H, W)
Output:
y_hat: (B, 2) — [clf_logit, reg_output]
"""
return self.model(x)
def training_step(self, batch: Any, batch_idx: int) -> torch.Tensor:
"""
One backbone training step.
"""
x, y, target, sample_weight, path, original_label = batch
y_hat = self(x)
yp_clf = y_hat[:, 0:1]
yp_reg = y_hat[:, 1:2]
weights_clf = torch.where(y > 0, 1.0, 0.45).to(y.dtype)
clf_loss = (self.loss_clf(yp_clf, y) * weights_clf).mean()
reg_loss_raw = self.loss_reg(yp_reg, target)
sigma = self.sigma_a * target + self.sigma_b
reg_loss = (reg_loss_raw / (sigma ** 2)).mean()
loss = clf_loss + 0.5 * reg_loss
y_prob = torch.sigmoid(yp_clf).detach()
y_pred = (y_prob > 0.5).int().cpu().numpy()
y_true = y.detach().int().cpu().numpy()
self.log("train_clf_loss", clf_loss, prog_bar=True, sync_dist=True)
self.log("train_reg_loss", reg_loss, prog_bar=True, sync_dist=True)
self.log("train_loss", loss, prog_bar=True, sync_dist=True)
self.log("train_f1", skm.f1_score(y_true, y_pred, zero_division=0),
prog_bar=True, sync_dist=True)
self.log("train_acc", skm.accuracy_score(y_true, y_pred),
prog_bar=True, sync_dist=True)
return loss
def validation_step(self, batch: Any, batch_idx: int) -> torch.Tensor:
"""
One backbone validation step.
"""
x, y, target, sample_weight, path, original_label = batch
y_hat = self(x)
yp_clf = y_hat[:, 0:1]
yp_reg = y_hat[:, 1:2]
clf_loss = self.loss_clf(yp_clf, y).mean()
reg_loss_raw = self.loss_reg(yp_reg, target)
sigma = self.sigma_a * target + self.sigma_b
reg_loss = (reg_loss_raw / (sigma ** 2)).mean()
loss = clf_loss + 0.5 * reg_loss
y_prob = torch.sigmoid(yp_clf).float()
self._y_true.append(float(y[..., 0].float().cpu()))
self._y_prob.append(float(y_prob[..., 0].cpu()))
self._y_pred.append(int((y_prob[..., 0] > 0.5).cpu()))
self._t_true.append(float(target[..., 0].float().cpu()))
self._t_pred.append(float(yp_reg[..., 0].cpu()))
self.log("val_loss", loss, prog_bar=True, sync_dist=True)
self.log("val_clf_loss", clf_loss, prog_bar=False, sync_dist=True)
self.log("val_reg_loss", reg_loss, prog_bar=False, sync_dist=True)
return loss
def on_validation_epoch_end(self) -> None:
"""
Compute and log metrics at the end of validation.
"""
if len(self._t_true) > 0:
rmse = skm.root_mean_squared_error(self._t_true, self._t_pred)
mae = skm.mean_absolute_error(self._t_true, self._t_pred)
self.log("val_rmse", rmse, prog_bar=True, sync_dist=True)
self.log("val_reg_mae", mae, prog_bar=True, sync_dist=True)
if len(set(self._y_true)) > 1:
auc = skm.roc_auc_score(self._y_true, self._y_prob)
f1 = skm.f1_score(self._y_true, self._y_pred, zero_division=0)
acc = skm.accuracy_score(self._y_true, self._y_pred)
self.log("val_auc", auc, prog_bar=True, sync_dist=True)
self.log("val_f1", f1, prog_bar=True, sync_dist=True)
self.log("val_acc", acc, prog_bar=True, sync_dist=True)
self._y_true.clear()
self._y_prob.clear()
self._y_pred.clear()
self._t_true.clear()
self._t_pred.clear()
def on_train_epoch_end(self) -> None:
"""
Log the current learning rate.
"""
opt = self.optimizers()
self.log(
"lr",
opt.optimizer.param_groups[0]["lr"],
on_step=False,
on_epoch=True,
sync_dist=True,
)
def configure_optimizers(self):
"""
Configure the optimizer and OneCycleLR.
If weight_path is None:
train only self.model.fc (pretrain).
Otherwise:
train the full model (fine-tuning).
"""
if self.weight_path is None:
for p in self.parameters():
p.requires_grad = False
for p in self.model.fc.parameters():
p.requires_grad = True
params = self.model.fc.parameters()
else:
for p in self.parameters():
p.requires_grad = True
params = self.parameters()
optimizer = optim.AdamW(params, lr=self.lr, weight_decay=self.weight_decay)
if self.max_epochs is not None and getattr(self, "trainer", None) is not None:
total_steps = self.trainer.estimated_stepping_batches
scheduler = optim.lr_scheduler.OneCycleLR(
optimizer=optimizer,
max_lr=self.lr,
total_steps=total_steps,
)
return {
"optimizer": optimizer,
"lr_scheduler": {
"scheduler": scheduler,
"interval": "step",
},
}
return optimizer
def predict_step(self, batch: Any, batch_idx: int, dataloader_idx: int = 0) -> Any:
"""
Backbone inference step.
"""
x, y, target, sample_weight, path, original_label = batch
y_hat = self(x)
yp_clf = y_hat[:, 0:1]
yp_reg = y_hat[:, 1:2]
y_prob = torch.sigmoid(yp_clf)
return {
"y": y,
"y_pred": (y_prob > 0.5).int(),
"y_prob": y_prob,
"y_reg": yp_reg,
"target": target,
"original_label": original_label,
"path": path,
}
|