Download code/deep_elipse_model.py from qinxunhui/deepellipse-circleid: direct link, hf CLI and curl.
- Browser
- Download file 36.1 kB
-
https://huggingface.co/qinxunhui/deepellipse-circleid/resolve/main/code/deep_elipse_model.py
- Command line
-
hf download hf://qinxunhui/deepellipse-circleid/code/deep_elipse_model.py
-
curl -L -o deep_elipse_model.py https://huggingface.co/qinxunhui/deepellipse-circleid/resolve/main/code/deep_elipse_model.py
36.1 kB
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
DeepEllipse:§4.1 + §4.2 + §4.3 端到端书写者识别模型。
- 输入:exp6_data_loader 的 (fea_tensor, img_tensor, pen_ids, labels)
- 骨干:Section41Encoder → Section42Encoder → Section43Encoder
- §4.1.3 默认 ConvNeXt(冻结),权重按平台自动选择(Windows: model/…,Linux: checkpoints/model_convnext/…)
- 头:Linear(f_ellipse → num_classes),默认 CrossEntropyLoss 训练
用法:
python deep_elipse_model.py --mode train --debug
python deep_elipse_model.py --mode val --checkpoint checkpoints/deep_ellipse/best.pth
python deep_elipse_model.py --mode test --checkpoint checkpoints/deep_ellipse/best.pth
"""
from __future__ import annotations
import argparse
import os
import time
from dataclasses import asdict, dataclass, field
from pathlib import Path
from typing import Dict, List, Optional, Tuple, Union
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.data import DataLoader
from tqdm import tqdm
from elipse_feature import DEFAULT_IMG_SIZE, FeaTensor
from exp6_data_loader import Config as DataConfig, create_dataloaders
from section41 import Section41Config, Section41Encoder
from section42 import Section42Config, Section42Encoder
from section43 import (
Section43Config,
Section43Encoder,
forward_pipeline_batch_from_dataloader,
)
from section43_v2 import Section43ConfigV2
from section43_v3 import Section43ConfigV3, build_section43_encoder
from section43_v3_1 import Section43ConfigV3_1, Section43EncoderV3_1
from section42_cwt import (
CWTPhiEncoder,
Section42CWTConfig,
forward_pipeline_batch_from_dataloader_cwt,
)
from resnet_point_feature import default_resnet_checkpoint
from convnext_point_feature import default_convnext_checkpoint
PathLike = Union[str, Path]
# ---------------------------------------------------------------------------
# 配置
# ---------------------------------------------------------------------------
@dataclass
class DeepEllipseConfig:
"""端到端 DeepEllipse 超参。"""
# §4.x 子模块(None → 各 Section 默认)
cfg41: Optional[Section41Config] = None
cfg42: Optional[Section42Config] = None
cfg43: Optional[Section43Config] = None
# §4.3 方案:v1 / v2 / v3 / v3.1 四选一
use_scheme_v2: bool = False
use_scheme_v3: bool = False
use_scheme_v3_1: bool = False
seq_backend: Optional[str] = None # Table7: transformer|conformer|zipformer|dolphin|synsig
v2_pre_smooth: bool = True
v2_pool_mode: str = "sp" # max | mean | sp
v3_pool_mode: str = "avg" # avg | max
v3_1_stage_pool: str = "dbsa" # global | seg_sigmoid | dbsa | cwt_angular | cwt_dual
cwt_scale_affinity: bool = False # E2:各 stage 独立 scale 子集 MLP
cwt_scale_index_mode: str = "uniform" # uniform | physics(H3 E2-phys)
angular_norm: str = "sum_relu" # softmax | sum_relu(默认 sum_relu,E1 消融用 softmax)
e3_cross_query: str = "cwt_cnn" # phi | cwt_cnn(仅 cwt_dual 路1)
e3_cross_query_path2: str = "maxpool" # maxpool | orig_cnn(仅 cwt_dual 路2)
e3_cwt_branch: bool = True # False → E3a 仅路1(512)
# E2b(cwt_angular 扩展)
e2_cwt_branch: bool = False
cwt_phi_valid_emb: int = 0
e2_cross_path2: str = "orig_cnn" # maxpool | orig_cnn
gcc_enabled: bool = False
gcc_gate_init: float = -2.0
gcc_d_stage_emb: int = 16
gcc_dw_kernels_bu: tuple = (3, 5, 7)
gcc_dw_kernels_td: tuple = (7, 5, 3)
phi_tx_layers: int = 0
phi_tx_heads: int = 4
phi_tx_ffn: int = 256
phi_tx_dropout: float = 0.1
num_classes: int = 45
num_pens: int = 16
half_norm: float = DEFAULT_IMG_SIZE / 2.0
dropout: float = 0.1
deep_backend: str = "convnext"
resnet_checkpoint: str = field(default_factory=default_resnet_checkpoint)
convnext_checkpoint: str = field(default_factory=default_convnext_checkpoint)
freeze41_deep: bool = True
# 训练
batch_size: int = 16
num_epochs: int = 30
lr: float = 1e-4
weight_decay: float = 2e-4
grad_clip: float = 1.0
num_workers: int = 0
pin_memory: bool = True
checkpoint_dir: str = "checkpoints/deep_ellipse"
log_every: int = 20
sample_ratio: float = 1.0
min_samples: int = 0
device: str = field(
default_factory=lambda: "cuda" if torch.cuda.is_available() else "cpu"
)
def build_section_configs(self) -> Tuple[Section41Config, Section42Config, Section43Config]:
cfg41 = self.cfg41 or Section41Config(
deep_backend=self.deep_backend,
resnet_checkpoint=self.resnet_checkpoint,
convnext_checkpoint=self.convnext_checkpoint,
)
cfg42 = self.cfg42 or Section42Config()
scheme_flags = (self.use_scheme_v2, self.use_scheme_v3, self.use_scheme_v3_1)
if sum(int(x) for x in scheme_flags) > 1:
raise ValueError("use_scheme_v2 / v3 / v3_1 至多一个为 True")
if self.use_scheme_v3_1:
base = vars(self.cfg43) if self.cfg43 is not None else {}
cfg43 = Section43ConfigV3_1(
**{
k: v
for k, v in base.items()
if k in Section43ConfigV3_1.__dataclass_fields__
},
pool_mode=self.v3_pool_mode, # type: ignore[arg-type]
stage_pool_mode=self.v3_1_stage_pool,
cwt_scale_affinity=self.cwt_scale_affinity,
cwt_scale_index_mode=self.cwt_scale_index_mode,
angular_norm=self.angular_norm, # type: ignore[arg-type]
e3_cross_query=self.e3_cross_query,
e3_cross_query_path2=self.e3_cross_query_path2,
e3_cwt_branch=self.e3_cwt_branch,
e2_cwt_branch=self.e2_cwt_branch,
cwt_phi_valid_emb=self.cwt_phi_valid_emb,
e2_cross_path2=self.e2_cross_path2,
gcc_enabled=self.gcc_enabled,
gcc_gate_init=self.gcc_gate_init,
gcc_d_stage_emb=self.gcc_d_stage_emb,
gcc_dw_kernels_bu=self.gcc_dw_kernels_bu,
gcc_dw_kernels_td=self.gcc_dw_kernels_td,
phi_tx_layers=self.phi_tx_layers,
phi_tx_heads=self.phi_tx_heads,
phi_tx_ffn=self.phi_tx_ffn,
phi_tx_dropout=self.phi_tx_dropout,
)
elif self.use_scheme_v3 and self.seq_backend:
from section43_v3_seq import Section43ConfigV3Seq
out_dim = 256 if self.seq_backend in ("synsig", "dolphin") else 768
cfg43 = Section43ConfigV3Seq(
pool_mode=self.v3_pool_mode,
seq_backend=self.seq_backend, # type: ignore[arg-type]
backend_out_dim=out_dim,
)
elif self.use_scheme_v3:
base = vars(self.cfg43) if self.cfg43 is not None else {}
cfg43 = Section43ConfigV3(
**{
k: v
for k, v in base.items()
if k in Section43ConfigV3.__dataclass_fields__
},
pool_mode=self.v3_pool_mode, # type: ignore[arg-type]
)
elif self.use_scheme_v2:
base = vars(self.cfg43) if self.cfg43 is not None else {}
cfg43 = Section43ConfigV2(
**{
k: v
for k, v in base.items()
if k in Section43ConfigV2.__dataclass_fields__
},
use_pre_smooth=self.v2_pre_smooth,
pool_mode=self.v2_pool_mode, # type: ignore[arg-type]
)
else:
cfg43 = self.cfg43 or Section43Config()
return cfg41, cfg42, cfg43
# ---------------------------------------------------------------------------
# 模型
# ---------------------------------------------------------------------------
class DeepEllipseModel(nn.Module):
"""
DataLoader → §4.1 → §4.2 → §4.3 → 分类 logits。
forward 返回 logits;`return_details=True` 时附带各段中间量。
"""
def __init__(self, config: DeepEllipseConfig):
super().__init__()
self.config = config
cfg41, cfg42, cfg43 = config.build_section_configs()
self.enc41 = Section41Encoder(cfg41)
self.enc42 = Section42Encoder(cfg42)
if config.use_scheme_v3_1:
# 直接实例化,避免旧版 section43_v3.build_section43_encoder 无 use_v3_1 参数
self.enc43 = Section43EncoderV3_1(cfg43) # type: ignore[arg-type]
elif config.use_scheme_v3 and config.seq_backend:
from section43_v3_seq import Section43EncoderV3Seq
self.enc43 = Section43EncoderV3Seq(cfg43) # type: ignore[arg-type]
else:
self.enc43 = build_section43_encoder(
cfg43,
use_v2=config.use_scheme_v2,
use_v3=config.use_scheme_v3,
)
self.cwt_encoder: Optional[CWTPhiEncoder] = None
if config.use_scheme_v3_1 and config.v3_1_stage_pool in ("cwt_angular", "cwt_dual"):
cwt_cfg = Section42CWTConfig(
d_phi=cfg43.d_phi, # type: ignore[union-attr]
n_scales=cfg43.cwt_n_scales, # type: ignore[union-attr]
wavelet=cfg43.cwt_wavelet, # type: ignore[union-attr]
scale_affinity=cfg43.cwt_scale_affinity, # type: ignore[union-attr]
scale_index_mode=cfg43.cwt_scale_index_mode, # type: ignore[union-attr]
)
self.cwt_encoder = CWTPhiEncoder(cwt_cfg)
if config.freeze41_deep:
for p in self.enc41.deep_branch.parameters():
p.requires_grad = False
dropout_rate = 0.2 if config.use_scheme_v3_1 else config.dropout
if config.use_scheme_v2 or config.use_scheme_v3 or config.use_scheme_v3_1:
d_out = cfg43.d_ellipse
else:
d_out = cfg43.d_h
self.pen_embed = nn.Embedding(config.num_pens, cfg43.d_pen)
self.dropout = nn.Dropout(dropout_rate)
self.classifier = nn.Linear(d_out, config.num_classes)
self._init_classifier()
def _init_classifier(self) -> None:
nn.init.xavier_uniform_(self.classifier.weight)
nn.init.zeros_(self.classifier.bias)
@property
def half_norm(self) -> float:
return self.config.half_norm
def pen_embedding(self, pen_ids: Optional[torch.Tensor]) -> Optional[torch.Tensor]:
if pen_ids is None:
return None
idx = pen_ids.long().clamp(0, self.config.num_pens - 1)
return self.pen_embed(idx)
@property
def d_ellipse(self) -> int:
if (
self.config.use_scheme_v2
or self.config.use_scheme_v3
or self.config.use_scheme_v3_1
):
return self.enc43.config.d_ellipse
return self.enc43.config.d_h
@property
def d_geo(self) -> int:
return self.enc41.config.d_geo
@property
def d_pen(self) -> int:
return self.enc43.config.d_pen
def encode_branch(
self,
fea_tensor: FeaTensor,
img_tensor: torch.Tensor,
pen_ids: Optional[torch.Tensor] = None,
*,
training: bool = False,
cwt_scalo: Optional[torch.Tensor] = None,
) -> Tuple[Dict[str, torch.Tensor], Dict[str, torch.Tensor], Dict[str, torch.Tensor]]:
"""§4.1–§4.3 特征(无分类头)。
CWT 模式:§4.1/§4.2 与 DBSA 相同;仅在 enc43 前注入 phi_cwt;
enc43.pyramid + stage_proj 与旧 checkpoint 键名一致,可 partial 加载微调。
"""
e_pen = self.pen_embedding(pen_ids)
if self.cwt_encoder is not None:
return forward_pipeline_batch_from_dataloader_cwt(
fea_tensor,
img_tensor,
self.enc41,
self.enc42,
self.enc43,
half_norm=self.half_norm,
e_pen=e_pen,
training=training,
cwt_phi_encoder=self.cwt_encoder,
cwt_scalo_batch=cwt_scalo,
)
return forward_pipeline_batch_from_dataloader(
fea_tensor,
img_tensor,
self.enc41,
self.enc42,
self.enc43,
half_norm=self.half_norm,
e_pen=e_pen,
training=training,
)
def encode_for_fusion(
self,
fea_tensor: FeaTensor,
img_tensor: torch.Tensor,
pen_ids: Optional[torch.Tensor] = None,
*,
training: bool = False,
cwt_scalo: Optional[torch.Tensor] = None,
) -> Dict[str, torch.Tensor]:
"""
供 §5.1 融合:f_ellipse (§4.3 d_h)、g_vec (§4.1)、E_pen(共享 pen_embed)。
"""
o41, o42, o43 = self.encode_branch(
fea_tensor, img_tensor, pen_ids, training=training, cwt_scalo=cwt_scalo
)
e_pen = self.pen_embedding(pen_ids)
return {
"f_ellipse": o43["f_ellipse"],
"g_vec": o41["g_vec"],
"e_pen": e_pen,
"o41": o41,
"o42": o42,
"o43": o43,
}
def forward(
self,
fea_tensor: FeaTensor,
img_tensor: torch.Tensor,
pen_ids: Optional[torch.Tensor] = None,
*,
training: bool = False,
return_details: bool = False,
cwt_scalo: Optional[torch.Tensor] = None,
) -> Union[torch.Tensor, Tuple[torch.Tensor, Dict[str, torch.Tensor]]]:
o41, o42, o43 = self.encode_branch(
fea_tensor,
img_tensor,
pen_ids,
training=training,
cwt_scalo=cwt_scalo,
)
feat = self.dropout(o43["f_ellipse"])
logits = self.classifier(feat)
if not return_details:
return logits
details = {
"o41": o41,
"o42": o42,
"o43": o43,
"f_ellipse": o43["f_ellipse"],
}
return logits, details
def trainable_parameter_groups(self) -> List[Dict]:
"""AdamW 参数组:§4.1/42/43 + 头。"""
return [{"params": [p for p in self.parameters() if p.requires_grad]}]
def count_parameters(self) -> Tuple[int, int]:
total = sum(p.numel() for p in self.parameters())
trainable = sum(p.numel() for p in self.parameters() if p.requires_grad)
return total, trainable
def save_checkpoint(
self,
path: PathLike,
*,
epoch: int,
optimizer: torch.optim.Optimizer,
val_acc: float,
label_classes: List[str],
extra: Optional[dict] = None,
) -> None:
path = Path(path)
path.parent.mkdir(parents=True, exist_ok=True)
torch.save(
{
"epoch": epoch,
"val_acc": val_acc,
"model_state_dict": self.state_dict(),
"optimizer_state_dict": optimizer.state_dict(),
"label_encoder_classes": label_classes,
"config": self.config,
"extra": extra or {},
},
path,
)
@classmethod
def load_from_checkpoint(
cls,
path: PathLike,
*,
num_classes: Optional[int] = None,
map_location: Optional[str] = None,
strict: bool = True,
) -> Tuple["DeepEllipseModel", dict]:
payload = torch.load(path, map_location=map_location, weights_only=False)
cfg = payload.get("config")
if cfg is None:
cfg = DeepEllipseConfig(num_classes=num_classes or 45)
elif isinstance(cfg, dict):
cfg = DeepEllipseConfig(**cfg)
elif num_classes is not None:
cfg.num_classes = num_classes
if num_classes is not None:
cfg.num_classes = num_classes
model = cls(cfg)
model.load_state_dict(payload["model_state_dict"], strict=strict)
return model, payload
# ---------------------------------------------------------------------------
# 损失 / 指标
# ---------------------------------------------------------------------------
def classification_loss(
logits: torch.Tensor,
labels: torch.Tensor,
criterion: Optional[nn.Module] = None,
) -> torch.Tensor:
"""默认交叉熵;labels 中 -1 为无效标签(跳过)。"""
crit = criterion or nn.CrossEntropyLoss()
valid = labels >= 0
if not valid.any():
return logits.sum() * 0.0
return crit(logits[valid], labels[valid])
@torch.no_grad()
def batch_accuracy(logits: torch.Tensor, labels: torch.Tensor) -> Tuple[int, int]:
valid = labels >= 0
if not valid.any():
return 0, 0
pred = logits[valid].argmax(dim=-1)
correct = (pred == labels[valid]).sum().item()
return correct, valid.sum().item()
# ---------------------------------------------------------------------------
# 训练 / 验证 / 测试
# ---------------------------------------------------------------------------
def train_one_epoch(
model: DeepEllipseModel,
loader: DataLoader,
criterion: nn.Module,
optimizer: torch.optim.Optimizer,
device: torch.device,
*,
max_batches: Optional[int] = None,
grad_clip: float = 1.0,
) -> Tuple[float, float]:
model.train()
total_loss = 0.0
correct = 0
total = 0
n_batches = 0
pbar = tqdm(enumerate(loader), total=len(loader), desc="train", leave=False)
for batch_idx, batch in pbar:
if max_batches is not None and batch_idx >= max_batches:
break
input_all, labels, pen_ids, _writers, _image_ids = batch
fea_tensor, _ske, img_tensor, cwt_scalo = input_all
labels = labels.to(device)
pen_ids = pen_ids.to(device)
img_tensor = img_tensor.to(device)
if cwt_scalo is not None:
cwt_scalo = cwt_scalo.to(device)
optimizer.zero_grad(set_to_none=True)
logits = model(
fea_tensor, img_tensor, pen_ids, training=True, cwt_scalo=cwt_scalo
)
loss = classification_loss(logits, labels, criterion)
loss.backward()
if grad_clip > 0:
nn.utils.clip_grad_norm_(model.parameters(), grad_clip)
optimizer.step()
bs_correct, bs_total = batch_accuracy(logits.detach(), labels)
correct += bs_correct
total += bs_total
total_loss += loss.item()
n_batches += 1
pbar.set_postfix(loss=f"{loss.item():.4f}", acc=f"{bs_correct/max(bs_total,1):.3f}")
avg_loss = total_loss / max(n_batches, 1)
acc = correct / max(total, 1)
return avg_loss, acc
@torch.no_grad()
def evaluate(
model: DeepEllipseModel,
loader: DataLoader,
criterion: nn.Module,
device: torch.device,
*,
max_batches: Optional[int] = None,
) -> Tuple[float, float]:
model.eval()
total_loss = 0.0
correct = 0
total = 0
n_batches = 0
total = len(loader) if max_batches is None else min(len(loader), max_batches)
for batch_idx, batch in enumerate(tqdm(loader, total=total, desc="eval", leave=False)):
if max_batches is not None and batch_idx >= max_batches:
break
input_all, labels, pen_ids, _writers, _image_ids = batch
fea_tensor, _ske, img_tensor, cwt_scalo = input_all
labels = labels.to(device)
pen_ids = pen_ids.to(device)
img_tensor = img_tensor.to(device)
if cwt_scalo is not None:
cwt_scalo = cwt_scalo.to(device)
logits = model(
fea_tensor, img_tensor, pen_ids, training=False, cwt_scalo=cwt_scalo
)
loss = classification_loss(logits, labels, criterion)
bs_correct, bs_total = batch_accuracy(logits, labels)
correct += bs_correct
total += bs_total
total_loss += loss.item()
n_batches += 1
avg_loss = total_loss / max(n_batches, 1)
acc = correct / max(total, 1)
return avg_loss, acc
@torch.no_grad()
def infer_batch(
model: DeepEllipseModel,
loader: DataLoader,
device: torch.device,
*,
max_batches: Optional[int] = None,
) -> List[dict]:
"""推理:返回每条样本的 pred / prob / writer / image_id。"""
model.eval()
results: List[dict] = []
for batch_idx, batch in enumerate(loader):
if max_batches is not None and batch_idx >= max_batches:
break
input_all, labels, pen_ids, writers, image_ids = batch
fea_tensor, _ske, img_tensor, cwt_scalo = input_all
pen_ids = pen_ids.to(device)
img_tensor = img_tensor.to(device)
if cwt_scalo is not None:
cwt_scalo = cwt_scalo.to(device)
logits = model(
fea_tensor, img_tensor, pen_ids, training=False, cwt_scalo=cwt_scalo
)
probs = F.softmax(logits, dim=-1)
preds = probs.argmax(dim=-1)
for i in range(logits.size(0)):
results.append(
{
"image_id": image_ids[i],
"writer_id": writers[i],
"label": int(labels[i].item()),
"pred": int(preds[i].item()),
"confidence": float(probs[i, preds[i]].item()),
}
)
return results
# ---------------------------------------------------------------------------
# 主流程
# ---------------------------------------------------------------------------
def build_model_and_loaders(
cfg: DeepEllipseConfig,
*,
debug: bool = False,
) -> Tuple[DeepEllipseModel, DataLoader, DataLoader, Optional[DataLoader], object]:
sample_ratio = 0.02 if debug else cfg.sample_ratio
min_samples = DataConfig.DEBUG_MIN_SAMPLES if debug else cfg.min_samples
batch_size = 4 if debug else cfg.batch_size
num_workers = 0 if debug else cfg.num_workers
train_loader, val_loader, test_loader, label_encoder = create_dataloaders(
batch_size=batch_size,
num_workers=num_workers,
pin_memory=cfg.pin_memory and torch.cuda.is_available(),
sample_ratio=sample_ratio,
min_samples=min_samples,
use_precomputed=DataConfig.USE_PRECOMPUTED_FEA,
verbose=not debug,
)
num_classes = len(label_encoder.classes_)
cfg.num_classes = num_classes
model = DeepEllipseModel(cfg)
if debug:
# 验证集 CSV 未采样时 batch 极多;调试阶段复用训练子集
val_loader = train_loader
return model, train_loader, val_loader, test_loader, label_encoder
def run_train(cfg: DeepEllipseConfig, *, debug: bool = False, resume: Optional[str] = None) -> None:
device = torch.device(cfg.device)
if debug:
cfg.num_epochs = min(cfg.num_epochs, 2)
model, train_loader, val_loader, _test_loader, label_encoder = build_model_and_loaders(
cfg, debug=debug
)
model.to(device)
total, trainable = model.count_parameters()
print(f"DeepEllipse 参数量: total={total:,} trainable={trainable:,}")
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.AdamW(
model.trainable_parameter_groups()[0]["params"],
lr=cfg.lr,
weight_decay=cfg.weight_decay,
)
start_epoch = 0
best_acc = 0.0
if resume:
_m, payload = DeepEllipseModel.load_from_checkpoint(resume, map_location=cfg.device)
model.load_state_dict(payload["model_state_dict"], strict=True)
if "optimizer_state_dict" in payload:
optimizer.load_state_dict(payload["optimizer_state_dict"])
start_epoch = int(payload.get("epoch", 0)) + 1
best_acc = float(payload.get("val_acc", 0.0))
print(f"续训: epoch={start_epoch}, best_val_acc={best_acc:.4f}")
ckpt_dir = Path(cfg.checkpoint_dir)
ckpt_dir.mkdir(parents=True, exist_ok=True)
classes = label_encoder.classes_.tolist()
max_batches = 15 if debug else None
for epoch in range(start_epoch, cfg.num_epochs):
t0 = time.time()
train_loss, train_acc = train_one_epoch(
model,
train_loader,
criterion,
optimizer,
device,
max_batches=max_batches,
grad_clip=cfg.grad_clip,
)
val_loss, val_acc = evaluate(
model,
val_loader,
criterion,
device,
max_batches=max_batches,
)
elapsed = time.time() - t0
print(
f"Epoch {epoch + 1}/{cfg.num_epochs} "
f"train loss={train_loss:.4f} acc={train_acc:.4f} "
f"val loss={val_loss:.4f} acc={val_acc:.4f} "
f"({elapsed:.1f}s)"
)
last_path = ckpt_dir / "last.pth"
model.save_checkpoint(
last_path,
epoch=epoch,
optimizer=optimizer,
val_acc=val_acc,
label_classes=classes,
)
if val_acc >= best_acc:
best_acc = val_acc
best_path = ckpt_dir / "best.pth"
model.save_checkpoint(
best_path,
epoch=epoch,
optimizer=optimizer,
val_acc=val_acc,
label_classes=classes,
)
print(f" -> best saved: {best_path} (acc={best_acc:.4f})")
print(f"训练结束, best val acc = {best_acc:.4f}")
def run_eval(
cfg: DeepEllipseConfig,
checkpoint: str,
split: str = "val",
*,
debug: bool = False,
) -> None:
device = torch.device(cfg.device)
_model, train_loader, val_loader, test_loader, label_encoder = build_model_and_loaders(
cfg, debug=debug
)
cfg.num_classes = len(label_encoder.classes_)
model, _payload = DeepEllipseModel.load_from_checkpoint(
checkpoint,
num_classes=cfg.num_classes,
map_location=cfg.device,
)
model.to(device)
if split == "train":
loader = train_loader
elif split == "test" and test_loader is not None:
loader = test_loader
else:
loader = val_loader
criterion = nn.CrossEntropyLoss()
max_batches = 10 if debug else None
loss, acc = evaluate(model, loader, criterion, device, max_batches=max_batches)
print(f"[{split}] loss={loss:.4f} acc={acc:.4f} checkpoint={checkpoint}")
def run_infer_test(
cfg: DeepEllipseConfig,
checkpoint: Optional[str] = None,
*,
debug: bool = False,
) -> None:
"""快速推理 smoke test(可用随机初始化权重)。"""
device = torch.device(cfg.device)
model, train_loader, val_loader, _test_loader, label_encoder = build_model_and_loaders(
cfg, debug=debug
)
model.to(device)
if checkpoint and Path(checkpoint).is_file():
state = torch.load(checkpoint, map_location=cfg.device, weights_only=False)
model.load_state_dict(state["model_state_dict"], strict=True)
print(f"loaded: {checkpoint}")
else:
print("infer smoke: 随机初始化权重")
loader = val_loader
max_batches = 5 if debug else 20
results = infer_batch(model, loader, device, max_batches=max_batches)
correct = sum(1 for r in results if r["label"] >= 0 and r["pred"] == r["label"])
labeled = sum(1 for r in results if r["label"] >= 0)
print(f"infer samples={len(results)} acc={correct/max(labeled,1):.4f}")
# 打印前 5 条
for r in results[:5]:
wname = label_encoder.classes_[r["pred"]] if r["pred"] < len(label_encoder.classes_) else "?"
print(
f" id={r['image_id']} writer={r['writer_id']} "
f"label={r['label']} pred={r['pred']}({wname}) conf={r['confidence']:.3f}"
)
def parse_args() -> argparse.Namespace:
p = argparse.ArgumentParser(description="DeepEllipse §4.1–4.3 训练与推理")
p.add_argument(
"--mode",
choices=("train", "val", "test", "infer"),
default="train",
help="train=训练; val/test=评估; infer=推理 smoke",
)
p.add_argument("--checkpoint", type=str, default=None)
p.add_argument("--resume", type=str, default=None)
p.add_argument("--debug", action="store_true", help="小样本/少 epoch 快速验证")
p.add_argument("--device", type=str, default=None)
p.add_argument("--epochs", type=int, default=None)
p.add_argument("--batch-size", type=int, default=None)
p.add_argument("--lr", type=float, default=None)
p.add_argument(
"--deep",
choices=("patch", "resnet", "convnext"),
default="convnext",
help=f"§4.1.3 深度分支(默认 convnext + {default_convnext_checkpoint()})",
)
p.add_argument(
"--resnet-ckpt",
type=str,
default=None,
help=f"ResNet 权重路径(默认 {default_resnet_checkpoint()})",
)
p.add_argument(
"--convnext-ckpt",
type=str,
default=None,
help=f"ConvNeXt 权重路径(默认 {default_convnext_checkpoint()})",
)
p.add_argument("--sample-ratio", type=float, default=None)
p.add_argument(
"--scheme-v2",
action="store_true",
help="§4.3 v2(可配合 --pool-mode / --no-pre-smooth 做消融)",
)
p.add_argument(
"--no-pre-smooth",
action="store_true",
help="v2:关闭角向预平滑(默认开启)",
)
p.add_argument(
"--pool-mode",
choices=("max", "mean", "sp"),
default="sp",
help="v2 序列池化:max/mean→128维,sp→512维",
)
p.add_argument(
"--scheme-v3",
action="store_true",
help="§4.3 v3:ConvNeXt-Tiny 1D 序列金字塔(默认 f_ellipse=768)",
)
p.add_argument(
"--scheme-v3-1",
action="store_true",
help="§4.3 v3.1:SGPSR 频谱引导金字塔 stage 路由(f_ellipse=128)",
)
p.add_argument(
"--v3-pool",
choices=("avg", "max"),
default="avg",
help="v3 / v3.1 全局池化方式",
)
p.add_argument(
"--v3-1-stage-pool",
choices=("global", "seg_sigmoid", "dbsa", "cwt_angular", "cwt_dual"),
default="dbsa",
help="v3.1:dbsa=DBSA-v2;cwt_angular=CWT-GAPA;cwt_dual=E3 双域 1024",
)
p.add_argument(
"--angular-norm",
choices=("softmax", "sum_relu"),
default="sum_relu",
help="CWT 角向池化归一化:sum_relu(默认)/ softmax(E1 消融)",
)
p.add_argument(
"--e3-cross-query",
choices=("phi", "cwt_cnn"),
default="cwt_cnn",
help="cwt_dual 路1:phi(φ→H_orig) 或 cwt_cnn(H_cwt→H_orig)",
)
p.add_argument(
"--e3-cross-query-path2",
choices=("maxpool", "orig_cnn"),
default="maxpool",
help="cwt_dual 路2:maxpool 或 orig_cnn(H_orig→H_cwt)",
)
p.add_argument(
"--e3-no-cwt-branch",
action="store_true",
help="cwt_dual 消融 E3a:关闭路2 CWT MaxPool,仅 GAPA 512",
)
p.add_argument(
"--cwt-scale-affinity",
action="store_true",
help="CWT-GAPA:各 stage 独立 scale 子集编码(E2/E2b)",
)
p.add_argument(
"--cwt-scale-index-mode",
choices=("uniform", "physics"),
default="uniform",
help="E2 尺度分组:uniform | physics",
)
p.add_argument(
"--e2-cwt-branch",
action="store_true",
help="E2b:cwt_angular 启用路2 CWT 域 orig→cwt,输出 1024",
)
p.add_argument(
"--e2-valid-phi",
action="store_true",
help="E2b:φ MLP 输入拼接 valid embedding(8 维)",
)
p.add_argument(
"--e2-phi-tx-layers",
type=int,
default=0,
help="E2b:每 stage φ 角向 Transformer 层数(0=关闭,推荐 2)",
)
p.add_argument(
"--e2-cross-path2",
choices=("maxpool", "orig_cnn"),
default="orig_cnn",
help="E2b 路2:maxpool 或 orig_cnn(H_orig→H_cwt)",
)
p.add_argument(
"--e2b",
action="store_true",
help="E2b 快捷:scaleAff + validPhi + phiTx2L + sum_relu + dual orig_cnn",
)
return p.parse_args()
def main() -> None:
args = parse_args()
cfg = DeepEllipseConfig(deep_backend=args.deep)
if args.resnet_ckpt:
cfg.resnet_checkpoint = args.resnet_ckpt
if args.convnext_ckpt:
cfg.convnext_checkpoint = args.convnext_ckpt
if args.device:
cfg.device = args.device
if args.epochs is not None:
cfg.num_epochs = args.epochs
if args.batch_size is not None:
cfg.batch_size = args.batch_size
if args.lr is not None:
cfg.lr = args.lr
if args.sample_ratio is not None:
cfg.sample_ratio = args.sample_ratio
scheme_count = sum(int(x) for x in (args.scheme_v2, args.scheme_v3, args.scheme_v3_1))
if scheme_count > 1:
raise ValueError("--scheme-v2 / --scheme-v3 / --scheme-v3-1 至多选一个")
if args.scheme_v3_1:
cfg.use_scheme_v3_1 = True
cfg.v3_pool_mode = args.v3_pool
cfg.v3_1_stage_pool = args.v3_1_stage_pool
cfg.cwt_scale_affinity = args.cwt_scale_affinity
cfg.cwt_scale_index_mode = args.cwt_scale_index_mode
cfg.angular_norm = args.angular_norm
cfg.e3_cross_query = args.e3_cross_query
cfg.e3_cross_query_path2 = args.e3_cross_query_path2
cfg.e3_cwt_branch = not args.e3_no_cwt_branch
if getattr(args, "e2b", False):
cfg.v3_1_stage_pool = "cwt_angular"
cfg.cwt_scale_affinity = True
cfg.angular_norm = "sum_relu"
cfg.e2_cwt_branch = True
cfg.cwt_phi_valid_emb = 8
cfg.phi_tx_layers = 2
cfg.e2_cross_path2 = "orig_cnn"
else:
cfg.e2_cwt_branch = getattr(args, "e2_cwt_branch", False)
cfg.cwt_phi_valid_emb = 8 if getattr(args, "e2_valid_phi", False) else 0
cfg.phi_tx_layers = getattr(args, "e2_phi_tx_layers", 0)
cfg.e2_cross_path2 = getattr(args, "e2_cross_path2", "orig_cnn")
if args.v3_1_stage_pool == "cwt_dual" and cfg.angular_norm == "softmax":
cfg.angular_norm = "sum_relu"
if cfg.e2_cwt_branch and cfg.angular_norm == "softmax":
cfg.angular_norm = "sum_relu"
cfg.dropout = 0.2
_, _, cfg43 = cfg.build_section_configs()
print(
f"§4.3 v3.1: {cfg43.scheme_label()}, "
f"f_ellipse dim = {cfg43.d_ellipse}, dropout = {cfg.dropout}"
)
elif args.scheme_v3:
cfg.use_scheme_v3 = True
cfg.v3_pool_mode = args.v3_pool
_, _, cfg43 = cfg.build_section_configs()
print(f"§4.3 v3: {cfg43.scheme_label()}, f_ellipse dim = {cfg43.d_ellipse}")
elif args.scheme_v2:
cfg.use_scheme_v2 = True
cfg.v2_pre_smooth = not args.no_pre_smooth
cfg.v2_pool_mode = args.pool_mode
_, _, cfg43 = cfg.build_section_configs()
print(f"§4.3 v2: {cfg43.scheme_label()}, f_ellipse dim = {cfg43.d_ellipse}")
if args.debug:
cfg.num_epochs = 2
cfg.batch_size = 4
cfg.num_workers = 0
cfg.sample_ratio = 0.01
cfg.min_samples = 8
print("debug mode: sample_ratio=0.01, batch=4, epochs=2")
if args.mode == "train":
run_train(cfg, debug=args.debug, resume=args.resume)
elif args.mode in ("val", "test"):
ckpt = args.checkpoint or str(Path(cfg.checkpoint_dir) / "best.pth")
if not Path(ckpt).is_file():
raise FileNotFoundError(f"checkpoint not found: {ckpt}")
run_eval(cfg, ckpt, split=args.mode, debug=args.debug)
elif args.mode == "infer":
run_infer_test(cfg, checkpoint=args.checkpoint, debug=args.debug)
else:
raise ValueError(args.mode)
if __name__ == "__main__":
main()
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
DeepEllipse:§4.1 + §4.2 + §4.3 端到端书写者识别模型。
- 输入:exp6_data_loader 的 (fea_tensor, img_tensor, pen_ids, labels)
- 骨干:Section41Encoder → Section42Encoder → Section43Encoder
- §4.1.3 默认 ConvNeXt(冻结),权重按平台自动选择(Windows: model/…,Linux: checkpoints/model_convnext/…)
- 头:Linear(f_ellipse → num_classes),默认 CrossEntropyLoss 训练
用法:
python deep_elipse_model.py --mode train --debug
python deep_elipse_model.py --mode val --checkpoint checkpoints/deep_ellipse/best.pth
python deep_elipse_model.py --mode test --checkpoint checkpoints/deep_ellipse/best.pth
"""
from __future__ import annotations
import argparse
import os
import time
from dataclasses import asdict, dataclass, field
from pathlib import Path
from typing import Dict, List, Optional, Tuple, Union
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.data import DataLoader
from tqdm import tqdm
from elipse_feature import DEFAULT_IMG_SIZE, FeaTensor
from exp6_data_loader import Config as DataConfig, create_dataloaders
from section41 import Section41Config, Section41Encoder
from section42 import Section42Config, Section42Encoder
from section43 import (
Section43Config,
Section43Encoder,
forward_pipeline_batch_from_dataloader,
)
from section43_v2 import Section43ConfigV2
from section43_v3 import Section43ConfigV3, build_section43_encoder
from section43_v3_1 import Section43ConfigV3_1, Section43EncoderV3_1
from section42_cwt import (
CWTPhiEncoder,
Section42CWTConfig,
forward_pipeline_batch_from_dataloader_cwt,
)
from resnet_point_feature import default_resnet_checkpoint
from convnext_point_feature import default_convnext_checkpoint
PathLike = Union[str, Path]
# ---------------------------------------------------------------------------
# 配置
# ---------------------------------------------------------------------------
class DeepEllipseConfig:
"""端到端 DeepEllipse 超参。"""
# §4.x 子模块(None → 各 Section 默认)
cfg41: Optional[Section41Config] = None
cfg42: Optional[Section42Config] = None
cfg43: Optional[Section43Config] = None
# §4.3 方案:v1 / v2 / v3 / v3.1 四选一
use_scheme_v2: bool = False
use_scheme_v3: bool = False
use_scheme_v3_1: bool = False
seq_backend: Optional[str] = None # Table7: transformer|conformer|zipformer|dolphin|synsig
v2_pre_smooth: bool = True
v2_pool_mode: str = "sp" # max | mean | sp
v3_pool_mode: str = "avg" # avg | max
v3_1_stage_pool: str = "dbsa" # global | seg_sigmoid | dbsa | cwt_angular | cwt_dual
cwt_scale_affinity: bool = False # E2:各 stage 独立 scale 子集 MLP
cwt_scale_index_mode: str = "uniform" # uniform | physics(H3 E2-phys)
angular_norm: str = "sum_relu" # softmax | sum_relu(默认 sum_relu,E1 消融用 softmax)
e3_cross_query: str = "cwt_cnn" # phi | cwt_cnn(仅 cwt_dual 路1)
e3_cross_query_path2: str = "maxpool" # maxpool | orig_cnn(仅 cwt_dual 路2)
e3_cwt_branch: bool = True # False → E3a 仅路1(512)
# E2b(cwt_angular 扩展)
e2_cwt_branch: bool = False
cwt_phi_valid_emb: int = 0
e2_cross_path2: str = "orig_cnn" # maxpool | orig_cnn
gcc_enabled: bool = False
gcc_gate_init: float = -2.0
gcc_d_stage_emb: int = 16
gcc_dw_kernels_bu: tuple = (3, 5, 7)
gcc_dw_kernels_td: tuple = (7, 5, 3)
phi_tx_layers: int = 0
phi_tx_heads: int = 4
phi_tx_ffn: int = 256
phi_tx_dropout: float = 0.1
num_classes: int = 45
num_pens: int = 16
half_norm: float = DEFAULT_IMG_SIZE / 2.0