File size: 15,166 Bytes
76d61a0 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 | from __future__ import annotations
import torch
from torch import nn
import torch.nn.functional as F
class CIBCELoss(nn.Module):
def __init__(
self,
output_key: str = "ci",
label_key: str = "ci",
weight: float = 1.0,
eps: float = 1e-6,
):
super().__init__()
self.output_key = str(output_key)
self.label_key = str(label_key)
self.weight = float(weight)
self.eps = float(eps)
def forward(self, outputs: dict[str, torch.Tensor], labels: dict[str, torch.Tensor]) -> torch.Tensor:
preds = torch.clamp(outputs[self.output_key].float(), self.eps, 1.0 - self.eps)
targets = labels[self.label_key].float()
if targets.ndim == 4 and targets.shape[1] == 1:
targets = targets[:, 0]
loss = self.weight * F.binary_cross_entropy(preds, targets)
self.last_components = {"bce": float(loss.detach().cpu())}
return loss
class CIFocalLoss(nn.Module):
def __init__(
self,
output_key: str = "ci",
label_key: str = "ci",
weight: float = 1.0,
alpha: float = 0.25,
gamma: float = 2.0,
eps: float = 1e-6,
):
super().__init__()
self.output_key = str(output_key)
self.label_key = str(label_key)
self.weight = float(weight)
self.alpha = float(alpha)
self.gamma = float(gamma)
self.eps = float(eps)
def forward(self, outputs: dict[str, torch.Tensor], labels: dict[str, torch.Tensor]) -> torch.Tensor:
preds = torch.clamp(outputs[self.output_key].float(), self.eps, 1.0 - self.eps)
targets = labels[self.label_key].float().clamp(0.0, 1.0)
if targets.ndim == 4 and targets.shape[1] == 1:
targets = targets[:, 0]
bce = F.binary_cross_entropy(preds, targets, reduction="none")
pt = targets * preds + (1.0 - targets) * (1.0 - preds)
alpha_t = targets * self.alpha + (1.0 - targets) * (1.0 - self.alpha)
loss = self.weight * (alpha_t * (1.0 - pt).pow(self.gamma) * bce).mean()
self.last_components = {"focal": float(loss.detach().cpu())}
return loss
class CIFocalTverskyLoss(nn.Module):
def __init__(
self,
output_key: str = "ci",
label_key: str = "ci",
weight: float = 1.0,
focal_weight: float = 0.5,
tversky_weight: float = 0.5,
fp_weight: float = 0.7,
fn_weight: float = 0.3,
focal_gamma: float = 2.0,
tversky_gamma: float = 1.0,
positive_alpha: float = 0.35,
target_threshold: float = 0.5,
eps: float = 1e-6,
):
super().__init__()
self.output_key = str(output_key)
self.label_key = str(label_key)
self.weight = float(weight)
self.focal_weight = float(focal_weight)
self.tversky_weight = float(tversky_weight)
self.fp_weight = float(fp_weight)
self.fn_weight = float(fn_weight)
self.focal_gamma = float(focal_gamma)
self.tversky_gamma = float(tversky_gamma)
self.positive_alpha = float(positive_alpha)
self.target_threshold = float(target_threshold)
self.eps = float(eps)
def forward(self, outputs: dict[str, torch.Tensor], labels: dict[str, torch.Tensor]) -> torch.Tensor:
probs = torch.clamp(outputs[self.output_key].float(), self.eps, 1.0 - self.eps)
targets = labels[self.label_key].float().clamp(0.0, 1.0)
if targets.ndim == 4 and targets.shape[1] == 1:
targets = targets[:, 0]
bce = F.binary_cross_entropy(probs, targets, reduction="none")
pt = targets * probs + (1.0 - targets) * (1.0 - probs)
alpha_t = targets * self.positive_alpha + (1.0 - targets) * (1.0 - self.positive_alpha)
focal = alpha_t * (1.0 - pt).pow(self.focal_gamma) * bce
focal_loss = focal.mean()
reduce_dims = tuple(range(1, probs.dim()))
y_hard = (targets >= self.target_threshold).float()
tp = (probs * y_hard).sum(dim=reduce_dims)
fp = (probs * (1.0 - y_hard)).sum(dim=reduce_dims)
fn = ((1.0 - probs) * y_hard).sum(dim=reduce_dims)
tversky = (tp + self.eps) / (tp + self.fp_weight * fp + self.fn_weight * fn + self.eps)
tversky_loss = (1.0 - tversky).pow(self.tversky_gamma).mean()
loss = self.focal_weight * focal_loss + self.tversky_weight * tversky_loss
total = self.weight * loss
self.last_components = {
"focal": float((self.weight * self.focal_weight * focal_loss).detach().cpu()),
"tversky": float((self.weight * self.tversky_weight * tversky_loss).detach().cpu()),
"total": float(total.detach().cpu()),
}
return total
class BTMaskedSmoothL1Loss(nn.Module):
def __init__(
self,
output_key: str = "bt",
label_key: str = "bt",
weight: float = 1.0,
lead_weights: list[float] | None = None,
beta: float = 1.0,
):
super().__init__()
self.output_key = str(output_key)
self.label_key = str(label_key)
self.weight = float(weight)
self.beta = float(beta)
self.register_buffer("lead_weights", torch.tensor(lead_weights or [], dtype=torch.float32), persistent=False)
def forward(self, outputs: dict[str, torch.Tensor], labels: dict[str, torch.Tensor]) -> torch.Tensor:
preds = outputs[self.output_key].float()
targets = labels[self.label_key].float()
valid = torch.isfinite(targets)
if not bool(valid.any()):
zero = preds.sum() * 0.0
self.last_components = {"bt_smooth_l1": 0.0}
return zero
loss = F.smooth_l1_loss(preds, torch.nan_to_num(targets), beta=self.beta, reduction="none")
if self.lead_weights.numel() > 0:
if self.lead_weights.numel() != preds.shape[1]:
raise ValueError(f"lead_weights length {self.lead_weights.numel()} != T {preds.shape[1]}")
view_shape = (1, preds.shape[1]) + (1,) * (preds.ndim - 2)
loss = loss * self.lead_weights.to(preds.device).view(view_shape)
loss = self.weight * loss[valid].mean()
self.last_components = {"bt_smooth_l1": float(loss.detach().cpu())}
return loss
class BTMaskedMSELoss(nn.Module):
def __init__(
self,
output_key: str = "bt",
label_key: str = "bt",
weight: float = 1.0,
lead_weights: list[float] | None = None,
):
super().__init__()
self.output_key = str(output_key)
self.label_key = str(label_key)
self.weight = float(weight)
self.register_buffer("lead_weights", torch.tensor(lead_weights or [], dtype=torch.float32), persistent=False)
def forward(self, outputs: dict[str, torch.Tensor], labels: dict[str, torch.Tensor]) -> torch.Tensor:
preds = outputs[self.output_key].float()
targets = labels[self.label_key].float()
valid = torch.isfinite(targets)
if not bool(valid.any()):
zero = preds.sum() * 0.0
self.last_components = {"bt_mse": 0.0}
return zero
loss = (preds - torch.nan_to_num(targets)) ** 2
if self.lead_weights.numel() > 0:
if self.lead_weights.numel() != preds.shape[1]:
raise ValueError(f"lead_weights length {self.lead_weights.numel()} != T {preds.shape[1]}")
view_shape = (1, preds.shape[1]) + (1,) * (preds.ndim - 2)
loss = loss * self.lead_weights.to(preds.device).view(view_shape)
loss = self.weight * loss[valid].mean()
self.last_components = {"bt_mse": float(loss.detach().cpu())}
return loss
class CIBTLoss(nn.Module):
def __init__(
self,
ci_loss: dict,
bt_loss: dict | None = None,
normalize_terms: bool = False,
ref_batches: int = 100,
eps: float = 1e-8,
):
super().__init__()
ci_cfg = dict(ci_loss)
bt_cfg = dict(bt_loss or {"name": "bt_mse", "weight": 0.05})
self.ci_weight = float(ci_cfg.pop("weight", 1.0))
self.bt_weight = float(bt_cfg.pop("weight", 1.0))
self.ci_loss = _build_one_loss(ci_cfg, default_name=str(ci_cfg.get("name", "bce")))
self.bt_loss = _build_one_loss(bt_cfg, default_name="bt_mse")
self.normalize_terms = bool(normalize_terms)
self.ref_batches = int(ref_batches)
self.eps = float(eps)
self.register_buffer("ref_count", torch.tensor(0.0), persistent=True)
self.register_buffer("ci_ref_sum", torch.tensor(0.0), persistent=True)
self.register_buffer("bt_ref_sum", torch.tensor(0.0), persistent=True)
self.register_buffer("ci_ref", torch.tensor(1.0), persistent=True)
self.register_buffer("bt_ref", torch.tensor(1.0), persistent=True)
self.last_components: dict[str, float] = {}
def forward(self, outputs: dict[str, torch.Tensor], labels: dict[str, torch.Tensor]) -> torch.Tensor:
ci = self.ci_loss(outputs, labels)
bt = self.bt_loss(outputs, labels)
if self.normalize_terms:
self._update_refs(ci, bt)
ci_term = ci / self.ci_ref.clamp_min(self.eps)
bt_term = bt / self.bt_ref.clamp_min(self.eps)
else:
ci_term = ci
bt_term = bt
ci_weighted = self.ci_weight * ci_term
bt_weighted = self.bt_weight * bt_term
total = ci_weighted + bt_weighted
components = {
"ci": float(ci_weighted.detach().cpu()),
"bt": float(bt_weighted.detach().cpu()),
"total": float(total.detach().cpu()),
"ci_raw": float(ci.detach().cpu()),
"bt_raw": float(bt.detach().cpu()),
"ci_norm": float(ci_term.detach().cpu()),
"bt_norm": float(bt_term.detach().cpu()),
"ci_weight": float(self.ci_weight),
"bt_weight": float(self.bt_weight),
"ci_ref": float(self.ci_ref.detach().cpu()),
"bt_ref": float(self.bt_ref.detach().cpu()),
"ref_count": float(self.ref_count.detach().cpu()),
"ref_ready": float(self._ref_ready()),
}
for name, value in getattr(self.ci_loss, "last_components", {}).items():
components[f"ci_{name}"] = float(value)
for name, value in getattr(self.bt_loss, "last_components", {}).items():
component_name = str(name) if str(name).startswith("bt_") else f"bt_{name}"
components[component_name] = float(value)
self.last_components = components
return total
def _ref_ready(self) -> bool:
return bool(self.normalize_terms and self.ref_count.item() >= max(self.ref_batches, 1))
def _update_refs(self, ci: torch.Tensor, bt: torch.Tensor) -> None:
if self.ref_batches <= 0:
if self.ref_count.item() < 1:
self.ci_ref.copy_(ci.detach().abs().clamp_min(self.eps))
self.bt_ref.copy_(bt.detach().abs().clamp_min(self.eps))
self.ref_count.fill_(1.0)
return
if self.ref_count.item() >= self.ref_batches:
return
self.ci_ref_sum.add_(ci.detach().abs())
self.bt_ref_sum.add_(bt.detach().abs())
self.ref_count.add_(1.0)
denom = self.ref_count.clamp_min(1.0)
self.ci_ref.copy_((self.ci_ref_sum / denom).clamp_min(self.eps))
self.bt_ref.copy_((self.bt_ref_sum / denom).clamp_min(self.eps))
class CombinedLoss(nn.Module):
def __init__(self, losses: dict[str, nn.Module]):
super().__init__()
if not losses:
raise ValueError("CombinedLoss requires at least one loss")
self.losses = nn.ModuleDict(losses)
self.last_components: dict[str, float] = {}
def forward(self, outputs: dict[str, torch.Tensor], labels: dict[str, torch.Tensor]) -> torch.Tensor:
total = None
components = {}
for name, loss_fn in self.losses.items():
value = loss_fn(outputs, labels)
components[name] = float(value.detach().cpu())
for sub_name, sub_value in getattr(loss_fn, "last_components", {}).items():
component_name = str(sub_name) if name == "loss" else f"{name}_{sub_name}"
components[component_name] = float(sub_value)
total = value if total is None else total + value
self.last_components = components
if total is None:
raise RuntimeError("no losses were evaluated")
return total
def build_loss(config: dict) -> nn.Module:
config = dict(config)
if str(config.get("name", "")).lower() == "ci_bt":
return _build_one_loss(config, default_name="ci_bt")
if "losses" in config:
return CombinedLoss(
{
str(component_name): _build_one_loss(dict(component_cfg), default_name=str(component_name))
for component_name, component_cfg in dict(config["losses"]).items()
}
)
return CombinedLoss({"loss": _build_one_loss(config, default_name=str(config.get("name", "binary_focal")))})
def _build_one_loss(config: dict, default_name: str) -> nn.Module:
name = str(config.get("name", default_name)).lower()
params = dict(config.get("params", {}))
output_key = str(config.get("output_key", "bt" if name.startswith("bt") or "smooth_l1" in name else "ci"))
label_key = str(config.get("label_key", config.get("target_label", output_key)))
weight = float(config.get("weight", 1.0))
if name in {"bce", "binary_bce", "binary_cross_entropy", "ci_bce"}:
return CIBCELoss(output_key=output_key, label_key=label_key, weight=weight, **params)
if name in {"binary_focal", "focal", "binary_focal_loss", "ci", "ci_focal"}:
return CIFocalLoss(output_key=output_key, label_key=label_key, weight=weight, **params)
if name in {"focal_tversky", "tversky_focal", "ci_focal_tversky", "far_aware_tversky_focal", "far_aware", "ci_far_aware"}:
return CIFocalTverskyLoss(output_key=output_key, label_key=label_key, weight=weight, **params)
if name in {"masked_lead_smooth_l1", "bt_smooth_l1", "bt", "bt_masked_smooth_l1"}:
return BTMaskedSmoothL1Loss(output_key=output_key, label_key=label_key, weight=weight, **params)
if name in {"bt_mse", "masked_bt_mse", "bt_masked_mse"}:
return BTMaskedMSELoss(output_key=output_key, label_key=label_key, weight=weight, **params)
if name == "ci_bt":
if "ci_loss" not in config:
raise ValueError("ci_bt loss requires a ci_loss config block")
bt_loss = config.get("bt_loss")
if bt_loss is None:
bt_loss = {
"name": "bt_mse",
"output_key": config.get("bt_output_key", "bt"),
"label_key": config.get("bt_label_key", "bt"),
"weight": config.get("bt_weight", 0.05),
}
return CIBTLoss(
ci_loss=dict(config["ci_loss"]),
bt_loss=dict(bt_loss),
**params,
)
raise ValueError(f"unknown loss: {name}")
|