""" 论文 §4.2 段级频谱特征 — PyTorch 可训练封装。 FFT(扩窗分窗 + 整圆)在 elipse_feature / DataLoader 中 CPU 预计算; 本模块实现:全局谱池化 → 门控融合 → BN+LN+Conv1d → 谱轴 MHA → f̄_freq。 默认:N_θ=360, M=10, s_l=36, L_win=72, w=5(与论文 §4.2 / §7.1.3 一致)。 """ from __future__ import annotations from dataclasses import dataclass from pathlib import Path from typing import Dict, List, Optional, Tuple, Union import numpy as np import torch import torch.nn as nn from elipse_feature import ( DEFAULT_N_SEGMENTS, DEFAULT_N_THETA, FeaFields, FeaTensor, S_RES_CHANNELS, SegmentGeometry, build_segment_geometry, compute_fft_amplitude_segmented, compute_spectral_fft_features, extract_ellipse_features, preprocess_rgb_to_skeleton, responsibility_indices, unpack_fea_fields, window_indices, ) PathLike = Union[str, Path] CHANNEL_LABELS = [r"$\Delta r$", r"$\Delta\theta$", r"$\kappa$", r"$\tau$"] compute_fft_amplitude_batch = compute_fft_amplitude_segmented @dataclass class Section42Config: n_theta: int = DEFAULT_N_THETA n_segments: int = DEFAULT_N_SEGMENTS global_pool_win: int = 5 c1: int = 32 d_freq: int = 64 conv_kernel: int = 3 num_heads: int = 4 # --------------------------------------------------------------------------- # §4.2.1 全局谱频轴池化 # --------------------------------------------------------------------------- class GlobalSpectrumPool(nn.Module): """A_global (N_θ/2×4) → 沿频轴 AvgPool1d(w, stride=w) → (s_l×4)。""" def __init__(self, kernel: int = 5): super().__init__() self.pool = nn.AvgPool1d(kernel_size=kernel, stride=kernel) def forward(self, amp_global: torch.Tensor) -> torch.Tensor: """ amp_global: (B, N_θ/2, 4) 或 (N_θ/2, 4) 返回 amp_global_bar: (B, s_l, 4) """ if amp_global.dim() == 2: amp_global = amp_global.unsqueeze(0) x = amp_global.transpose(1, 2).contiguous() x = self.pool(x) return x.transpose(1, 2).contiguous() # --------------------------------------------------------------------------- # §4.2.3 全局—局部频率门控融合 # --------------------------------------------------------------------------- class FrequencyGateFusion(nn.Module): """α⊙A_loc + (1-α)⊙Ā_global,α=σ(W[loc;glob]+b)。""" def __init__(self, c_in: int = S_RES_CHANNELS): super().__init__() self.proj = nn.Linear(2 * c_in, c_in) def forward( self, amp_loc: torch.Tensor, amp_glob_bar: torch.Tensor, ) -> torch.Tensor: """ amp_loc: (B, M, s_l, 4) amp_glob_bar: (B, s_l, 4) 返回 amp_fuse: (B, M, s_l, 4) """ g = amp_glob_bar.unsqueeze(1).expand_as(amp_loc) alpha = torch.sigmoid(self.proj(torch.cat([amp_loc, g], dim=-1))) return alpha * amp_loc + (1.0 - alpha) * g # --------------------------------------------------------------------------- # §4.2.4 融合谱域编码 # --------------------------------------------------------------------------- class SpectralConvNet(nn.Module): """ BN → LN → Conv1d(4→C1) → Conv1d(C1→d_freq),频率轴长度保持 s_l。 输入 (B, 4, s_l),输出 (B, d_freq, s_l)。 """ def __init__( self, s_l: int, c_in: int = S_RES_CHANNELS, c1: int = 32, d_freq: int = 64, kernel: int = 3, ): super().__init__() pad = kernel // 2 self.s_l = s_l self.bn = nn.BatchNorm1d(c_in) self.ln = nn.LayerNorm([c_in, s_l]) self.conv1 = nn.Conv1d(c_in, c1, kernel_size=kernel, padding=pad) self.conv2 = nn.Conv1d(c1, d_freq, kernel_size=kernel, padding=pad) self.act = nn.GELU() def forward(self, amp: torch.Tensor) -> torch.Tensor: x = self.bn(amp) x = self.ln(x) x = self.act(self.conv1(x)) x = self.conv2(x) return x class SpectralAxisAttention(nn.Module): """f' = LayerNorm(f + MHA(f)),沿谱轴(长度 s_l)。""" def __init__(self, d_freq: int = 64, num_heads: int = 4): super().__init__() self.mha = nn.MultiheadAttention( d_freq, num_heads, batch_first=True, ) self.ln = nn.LayerNorm(d_freq) def forward(self, f: torch.Tensor) -> torch.Tensor: """f: (N, d_freq, s_l) → f': 同形状。""" x = f.transpose(1, 2) attn, _ = self.mha(x, x, x, need_weights=False) return self.ln(x + attn).transpose(1, 2) class Section42Encoder(nn.Module): """§4.2 完整编码器。""" def __init__(self, config: Optional[Section42Config] = None): super().__init__() self.config = config or Section42Config() self.geom = build_segment_geometry( self.config.n_theta, self.config.n_segments, ) self.global_pool = GlobalSpectrumPool(self.config.global_pool_win) self.gate_fusion = FrequencyGateFusion() self.spectral = SpectralConvNet( s_l=self.geom.s_l, c1=self.config.c1, d_freq=self.config.d_freq, kernel=self.config.conv_kernel, ) self.spectral_attn = SpectralAxisAttention( self.config.d_freq, self.config.num_heads, ) def _normalize_amp_shapes( self, amp_segment: torch.Tensor, amp_global: torch.Tensor, device: torch.device, dtype: torch.dtype, ) -> Tuple[torch.Tensor, torch.Tensor]: """统一为 (B, M, s_l, 4) 与 (B, N_θ/2, 4)。""" amp_seg = amp_segment.to(device=device, dtype=dtype) amp_glo = amp_global.to(device=device, dtype=dtype) if amp_seg.dim() == 3: amp_seg = amp_seg.unsqueeze(0) if amp_seg.shape[-1] == S_RES_CHANNELS and amp_seg.shape[-2] != self.geom.s_l: amp_seg = amp_seg.permute(0, 1, 3, 2) if amp_glo.dim() == 2: amp_glo = amp_glo.unsqueeze(0) if amp_glo.shape[-1] != S_RES_CHANNELS and amp_glo.shape[-2] == S_RES_CHANNELS: amp_glo = amp_glo.transpose(-1, -2) return amp_seg, amp_glo def _fuse_from_numpy_fft( self, s_res: torch.Tensor, ) -> Tuple[torch.Tensor, torch.Tensor]: """无预计算幅值时,CPU 侧算 amp_segment / amp_global。""" b = s_res.shape[0] seg_list: List[torch.Tensor] = [] glo_list: List[torch.Tensor] = [] for i in range(b): sp = compute_spectral_fft_features( s_res[i].detach().cpu().numpy(), self.geom.n_segments, ) seg_list.append(torch.from_numpy(sp.amp_segment)) glo_list.append(torch.from_numpy(sp.amp_global)) device, dtype = s_res.device, s_res.dtype return ( torch.stack(seg_list, dim=0).to(device=device, dtype=dtype), torch.stack(glo_list, dim=0).to(device=device, dtype=dtype), ) def forward( self, s_res: torch.Tensor, *, amp_segment: Optional[torch.Tensor] = None, amp_global: Optional[torch.Tensor] = None, return_amp: bool = True, ) -> Dict[str, torch.Tensor]: """ s_res: (N_theta, 4) 或 (B, N_theta, 4) amp_segment: 可选 (B, M, s_l, 4) 或 (M, s_l, 4) amp_global: 可选 (B, N_θ/2, 4) 或 (N_θ/2, 4) 返回: f_freq: (B, M, d_freq, s_l) — §4.2.4 经 MHA 后的 f'_freq f_bar: (B, M, d_freq) — §4.2.5 谱轴均值,供 §4.3 Query amp_fuse: (B, M, s_l, 4) 可选 """ batched = s_res.dim() == 3 if not batched: s_res = s_res.unsqueeze(0) b, n, c = s_res.shape assert n == self.geom.n_theta and c == S_RES_CHANNELS m_count = self.geom.n_segments device = s_res.device dtype = s_res.dtype if amp_segment is not None and amp_global is not None: amp_seg, amp_glo = self._normalize_amp_shapes( amp_segment, amp_global, device, dtype ) else: amp_seg, amp_glo = self._fuse_from_numpy_fft(s_res) amp_glob_bar = self.global_pool(amp_glo) amp_fuse = self.gate_fusion(amp_seg, amp_glob_bar) amp_conv = amp_fuse.permute(0, 1, 3, 2).reshape( b * m_count, S_RES_CHANNELS, self.geom.s_l ) f_flat = self.spectral(amp_conv) f_flat = self.spectral_attn(f_flat) f_freq = f_flat.reshape(b, m_count, self.config.d_freq, self.geom.s_l) f_bar = f_freq.mean(dim=-1) out: Dict[str, torch.Tensor] = { "f_freq": f_freq, "f_bar": f_bar, "amp_fuse": amp_fuse, "amp_global_bar": amp_glob_bar, } if return_amp: out["amp"] = amp_seg out["amp_global"] = amp_glo if not batched: out = {k: v.squeeze(0) for k, v in out.items()} return out def forward_from_fea( self, fea_tensor: FeaTensor, *, return_amp: bool = True, ) -> Dict[str, torch.Tensor]: """ 使用 DataLoader 输出的 fea_tensor 运行 §4.2。 输入:amp_segment (M,s_l,4)、amp_global (N_θ/2,4);s_res 仅作形状校验/回退。 """ fields = ( fea_tensor.fields if isinstance(fea_tensor, FeaTensor) else unpack_fea_fields(fea_tensor) ) device = next(self.parameters()).device s_res = fields.s_res.float().to(device) amp_segment = fields.amp_segment.float().to(device) amp_global = fields.amp_global.float().to(device) return self.forward( s_res, amp_segment=amp_segment, amp_global=amp_global, return_amp=return_amp, ) def trainable_parameters(self) -> List[torch.nn.Parameter]: return [p for p in self.parameters() if p.requires_grad] def forward_from_image( self, rgb: PathLike, *, img_size: int = 100, ) -> Tuple[dict, object, Dict[str, torch.Tensor]]: prep = preprocess_rgb_to_skeleton(rgb, size=img_size) bundle = extract_ellipse_features( prep["skeleton"], n_theta=self.config.n_theta, half_norm=img_size / 2.0, ) s_res = torch.from_numpy(bundle.sequence.s_res).float() amp_pre = amp_glo = None if bundle.spectral is not None: amp_pre = torch.from_numpy(bundle.spectral.amp_segment).float() amp_glo = torch.from_numpy(bundle.spectral.amp_global).float() return prep, bundle, self.forward( s_res, amp_segment=amp_pre, amp_global=amp_glo, ) def save_checkpoint(self, path: Union[str, Path], extra: Optional[dict] = None) -> None: torch.save( { "config": self.config, "state_dict": self.state_dict(), "extra": extra or {}, }, path, ) @classmethod def load_checkpoint( cls, path: Union[str, Path], map_location: Optional[str] = None, strict: bool = True, ) -> "Section42Encoder": payload = torch.load(path, map_location=map_location, weights_only=False) cfg = payload.get("config") or Section42Config() if isinstance(cfg, dict): cfg = Section42Config(**cfg) model = cls(cfg) model.load_state_dict(payload["state_dict"], strict=strict) return model def extract_section42( s_res: np.ndarray, config: Optional[Section42Config] = None, encoder: Optional[Section42Encoder] = None, device: Optional[str] = None, amp_segment: Optional[np.ndarray] = None, amp_global: Optional[np.ndarray] = None, ) -> Dict[str, np.ndarray]: enc = encoder or Section42Encoder(config) if device: enc = enc.to(device) enc.eval() with torch.no_grad(): t = torch.from_numpy(s_res).float() amp_s = torch.from_numpy(amp_segment).float() if amp_segment is not None else None amp_g = torch.from_numpy(amp_global).float() if amp_global is not None else None if device: t = t.to(device) if amp_s is not None: amp_s = amp_s.to(device) if amp_g is not None: amp_g = amp_g.to(device) out = enc( t, amp_segment=amp_s, amp_global=amp_g, return_amp=True, ) return {k: v.detach().cpu().numpy() for k, v in out.items()} def demo(image_path: str = "data/images/20938.png") -> Dict[str, np.ndarray]: enc = Section42Encoder() enc.eval() with torch.no_grad(): _, bundle, out_t = enc.forward_from_image(image_path) out = {k: v.detach().cpu().numpy() for k, v in out_t.items()} g = enc.geom print(f"§4.2 demo: {image_path}") print(f" M={g.n_segments}, s_l={g.s_l}, L_win={g.l_win}, d_freq={enc.config.d_freq}") print(f" amp_fuse {out['amp_fuse'].shape}") print(f" amp_global_bar {out['amp_global_bar'].shape}") print(f" f_freq {out['f_freq'].shape}") print(f" f_bar {out['f_bar'].shape}") return out def demo_from_dataloader( sample_ratio: float = 0.001, batch_size: int = 2, config: Optional[Section42Config] = None, device: str = "cpu", ) -> Dict[str, torch.Tensor]: """从 exp6_data_loader 取 fea_tensor,调试 §4.2(amp_segment + amp_global)。""" from exp6_data_loader import create_dataloaders cfg = config or Section42Config() enc = Section42Encoder(cfg).to(device) enc.eval() g = enc.geom train_loader, _, _, _ = create_dataloaders( batch_size=batch_size, num_workers=0, sample_ratio=sample_ratio, verbose=False, ) batch = next(iter(train_loader)) input_all, labels, pen_ids, writers, image_ids = batch fea_tensor, ske_tensor, img_tensor = input_all print("[section42] dataloader batch") print(f" fea_tensor.data: {fea_tensor.data.shape}") print(f" img_tensor: {tuple(img_tensor.shape)} label: {labels.tolist()}") print(f" M={g.n_segments}, s_l={g.s_l}, w={cfg.global_pool_win}, d_freq={cfg.d_freq}") fea_one = FeaTensor(data=fea_tensor.data[0]) fields = fea_one.fields print(f" s_res: {tuple(fields.s_res.shape)}") print(f" amp_segment: {tuple(fields.amp_segment.shape)}") print(f" amp_global: {tuple(fields.amp_global.shape)}") with torch.no_grad(): out_one = enc.forward_from_fea(fea_one) n_trainable = sum(p.numel() for p in enc.trainable_parameters()) print("[section42] forward_from_fea (single)") print(f" writer: {writers[0]} image_id: {image_ids[0]}") print(f" trainable params: {n_trainable:,}") print(f" amp_global_bar: {tuple(out_one['amp_global_bar'].shape)}") print(f" amp_fuse: {tuple(out_one['amp_fuse'].shape)}") print(f" f_freq: {tuple(out_one['f_freq'].shape)}") print(f" f_bar: {tuple(out_one['f_bar'].shape)}") print("[section42] batch forward") fields_b = fea_tensor.fields s_b = fields_b.s_res.float().to(device) amp_s = fields_b.amp_segment.float().to(device) amp_g = fields_b.amp_global.float().to(device) with torch.no_grad(): out_b = enc.forward(s_b, amp_segment=amp_s, amp_global=amp_g) print(f" batch s_res: {tuple(s_b.shape)}") print(f" batch amp_segment: {tuple(amp_s.shape)}") print(f" batch amp_global: {tuple(amp_g.shape)}") print(f" batch f_freq: {tuple(out_b['f_freq'].shape)}") print(f" batch f_bar: {tuple(out_b['f_bar'].shape)}") return out_one def main() -> None: import argparse p = argparse.ArgumentParser(description="§4.2 segment spectral encoder") p.add_argument("image", nargs="?", default="data/images/20938.png") p.add_argument("--save", type=str, default=None) p.add_argument("--device", default="cpu") p.add_argument( "--from-dataloader", action="store_true", help="使用 exp6_data_loader 输出的 fea_tensor 调试 §4.2", ) p.add_argument("--sample-ratio", type=float, default=0.001) p.add_argument("--batch-size", type=int, default=2) args = p.parse_args() if args.from_dataloader: demo_from_dataloader( sample_ratio=args.sample_ratio, batch_size=args.batch_size, device=args.device, ) return out = demo(args.image) if args.save: np.savez(args.save, **out) print("saved:", args.save) if __name__ == "__main__": demo_from_dataloader()