""" §4.2 扩展:CWT 角向谱先验 φ(t)。 从 S_res 或预计算 scalogram 构造逐角位 Query 特征,供 §4.3 CWT-GAPA 使用。 """ from __future__ import annotations from dataclasses import dataclass from typing import Dict, List, Optional, Tuple, TYPE_CHECKING import numpy as np import torch import torch.nn as nn from elipse_feature import ( CWT_N_SCALES_DEFAULT, CWT_WAVELET_DEFAULT, S_RES_CHANNELS, compute_cwt_scalogram, default_cwt_scale_indices, default_cwt_scales, ) from section41 import ( Section41Encoder, ) from section42 import Section42Encoder from section43 import ( build_enc43_valid_mask_kwargs, forward_pipeline_batch_to_upstream, ) from section43_v3 import ConvNeXt1DPyramid if TYPE_CHECKING: from section43 import Section43Encoder @dataclass class Section42CWTConfig: d_phi: int = 64 n_scales: int = CWT_N_SCALES_DEFAULT wavelet: str = CWT_WAVELET_DEFAULT mlp_hidden: int = 128 dropout: float = 0.1 scale_affinity: bool = False scale_index_mode: str = "uniform" # uniform | physics n_stages: int = 4 def _numpy_cwt_batch(s_res: torch.Tensor, cfg: Section42CWTConfig) -> torch.Tensor: """s_res (B, N, 4) → cwt_scalo (B, 4, K, N) on CPU numpy。""" arr = s_res.detach().cpu().numpy() scales = default_cwt_scales(cfg.n_scales) batch = [] for i in range(arr.shape[0]): batch.append(compute_cwt_scalogram(arr[i], scales=scales, wavelet=cfg.wavelet)) out = np.stack(batch, axis=0) return torch.from_numpy(out) class CWTPhiEncoder(nn.Module): """ scalogram |W(a,b)| → φ(t)。 统一 MLP(scale_affinity=False):四通道×K 尺度展平 → φ(t) ∈ R^{d_phi}。 scale_affinity=True 时由 §4.3 CWTAngularStagePool 按 stage 子集编码。 """ def __init__(self, cfg: Optional[Section42CWTConfig] = None): super().__init__() self.config = cfg or Section42CWTConfig() c = self.config in_dim = S_RES_CHANNELS * c.n_scales self.mlp = nn.Sequential( nn.Linear(in_dim, c.mlp_hidden), nn.GELU(), nn.Dropout(c.dropout), nn.Linear(c.mlp_hidden, c.d_phi), ) @property def raw_dim(self) -> int: return S_RES_CHANNELS * self.config.n_scales def encode_scalogram(self, cwt_scalo: torch.Tensor) -> torch.Tensor: """ cwt_scalo: (B, 4, K, N_theta) → phi_cwt: (B, N_theta, d_phi) """ b, _, k, n = cwt_scalo.shape flat = cwt_scalo.permute(0, 3, 1, 2).reshape(b * n, -1) phi = self.mlp(flat).reshape(b, n, -1) return phi def forward( self, s_res: torch.Tensor, *, cwt_scalo: Optional[torch.Tensor] = None, return_scalogram: bool = False, ) -> Dict[str, torch.Tensor]: """ s_res: (B, N_theta, 4) 或 (N_theta, 4) cwt_scalo: 可选预计算 (B, 4, K, N_theta) """ batched = s_res.dim() == 3 if not batched: s_res = s_res.unsqueeze(0) if cwt_scalo is not None and cwt_scalo.dim() == 3: cwt_scalo = cwt_scalo.unsqueeze(0) if cwt_scalo is None: cwt_scalo = _numpy_cwt_batch(s_res, self.config).to( device=s_res.device, dtype=s_res.dtype ) else: cwt_scalo = cwt_scalo.to(device=s_res.device, dtype=s_res.dtype) out: Dict[str, torch.Tensor] = {"cwt_scalo": cwt_scalo} if not self.config.scale_affinity: out["phi_cwt"] = self.encode_scalogram(cwt_scalo) if not batched: squeezed: Dict[str, torch.Tensor] = {} for k, v in out.items(): squeezed[k] = v.squeeze(0) out = squeezed return out class CWTStagePhiEncoder(nn.Module): """scale_affinity=True:每 stage 独立 MLP,输入为该 stage 尺度子集。""" def __init__(self, cfg: Section42CWTConfig, scale_indices: List[List[int]]): super().__init__() self.scale_indices = scale_indices self.mlps = nn.ModuleList() for idx in scale_indices: in_dim = S_RES_CHANNELS * len(idx) self.mlps.append( nn.Sequential( nn.Linear(in_dim, cfg.mlp_hidden), nn.GELU(), nn.Dropout(cfg.dropout), nn.Linear(cfg.mlp_hidden, cfg.d_phi), ) ) def forward(self, cwt_scalo: torch.Tensor) -> List[torch.Tensor]: """ cwt_scalo: (B, 4, K, N) → list of phi_s_full (B, N, d_phi),长度 = n_stages """ b, _, _, n = cwt_scalo.shape outputs: List[torch.Tensor] = [] for mlp, idx in zip(self.mlps, self.scale_indices): sub = cwt_scalo[:, :, idx, :] flat = sub.permute(0, 3, 1, 2).reshape(b * n, -1) phi = mlp(flat).reshape(b, n, -1) outputs.append(phi) return outputs class CWTStagePhiEncoderValid(nn.Module): """ E2b:scale 子集展平 + valid embedding → MLP → φ_raw。 valid_ang: (B, N) 或 None(默认全 1)。 """ def __init__( self, cfg: Section42CWTConfig, scale_indices: List[List[int]], valid_emb_dim: int = 8, ): super().__init__() self.valid_emb_dim = valid_emb_dim self.scale_indices = scale_indices self.valid_emb = nn.Embedding(2, valid_emb_dim) self.mlps = nn.ModuleList() for idx in scale_indices: in_dim = S_RES_CHANNELS * len(idx) + valid_emb_dim self.mlps.append( nn.Sequential( nn.Linear(in_dim, cfg.mlp_hidden), nn.GELU(), nn.Dropout(cfg.dropout), nn.Linear(cfg.mlp_hidden, cfg.d_phi), ) ) def forward( self, cwt_scalo: torch.Tensor, valid_ang: Optional[torch.Tensor] = None, ) -> List[torch.Tensor]: b, _, _, n = cwt_scalo.shape if valid_ang is None: valid_ang = torch.ones( b, n, device=cwt_scalo.device, dtype=cwt_scalo.dtype ) elif valid_ang.dim() == 1: valid_ang = valid_ang.unsqueeze(0).expand(b, -1) valid_idx = valid_ang.long().clamp(0, 1) valid_emb = self.valid_emb(valid_idx) outputs: List[torch.Tensor] = [] for mlp, idx in zip(self.mlps, self.scale_indices): sub = cwt_scalo[:, :, idx, :] flat = sub.permute(0, 3, 1, 2).reshape(b * n, -1) vemb = valid_emb.reshape(b * n, self.valid_emb_dim) x = torch.cat([flat, vemb], dim=-1) phi = mlp(x).reshape(b, n, -1) outputs.append(phi) return outputs def build_scale_indices( cfg: Section42CWTConfig, *, n_theta: int = 360, ) -> List[List[int]]: if cfg.scale_index_mode == "physics": try: from elipse_feature import physics_cwt_scale_indices except ImportError as exc: raise ImportError( "cwt scale_index_mode=physics 需要新版 elipse_feature.py " "(含 physics_cwt_scale_indices);请同步 elipse_feature.py " "或改用 scale_index_mode=uniform(E2b 默认)" ) from exc return physics_cwt_scale_indices( cfg.n_scales, cfg.n_stages, n_theta=n_theta, wavelet=cfg.wavelet, ) return default_cwt_scale_indices(cfg.n_scales, cfg.n_stages) def flatten_scalogram(cwt_scalo: torch.Tensor) -> torch.Tensor: """ cwt_scalo (B, 4, K, N_theta) → (B, 4*K, N_theta) 供 1D CNN 金字塔。 """ b, c, k, n = cwt_scalo.shape return cwt_scalo.permute(0, 3, 1, 2).reshape(b, n, c * k).transpose(1, 2) class CWTPyramid(nn.Module): """ Scalogram 1D CNN 金字塔,结构与 ConvNeXt1DPyramid 同构(360→180→90→45→22)。 """ def __init__( self, in_channels: int, dims: Tuple[int, ...] = (96, 192, 384, 768), depths: Tuple[int, ...] = (3, 3, 9, 3), *, stem_kernel: int = 4, stem_stride: int = 2, dw_kernel: int = 7, mlp_ratio: int = 4, ): super().__init__() self.pyramid = ConvNeXt1DPyramid( in_channels=in_channels, dims=dims, depths=depths, stem_kernel=stem_kernel, stem_stride=stem_stride, dw_kernel=dw_kernel, mlp_ratio=mlp_ratio, ) def forward(self, cwt_scalo: torch.Tensor) -> List[torch.Tensor]: """ cwt_scalo: (B, 4, K, N) → list of H_cwt^s (B, C_s, L_s),长度 = n_stages """ x = flatten_scalogram(cwt_scalo) _, pyramid_feats = self.pyramid(x) return pyramid_feats def cwt_pyramid_in_channels(n_scales: int, n_ch: int = S_RES_CHANNELS) -> int: return n_ch * n_scales class CWTStageProjector(nn.Module): """各 CWT pyramid stage 投影到 d_h。""" def __init__(self, stage_dims: Tuple[int, ...], d_h: int): super().__init__() self.projs = nn.ModuleList([nn.Linear(c, d_h) for c in stage_dims]) def forward(self, stage_idx: int, x: torch.Tensor) -> torch.Tensor: """x: (B, C_s, L_s) → (B, L_s, d_h)""" return self.projs[stage_idx](x.transpose(1, 2)) def enrich_o42_with_cwt( o42: Dict[str, torch.Tensor], s_res_b: torch.Tensor, cwt_phi_encoder: nn.Module, cwt_scalo_batch: Optional[torch.Tensor], device: torch.device, ) -> Dict[str, torch.Tensor]: """将 CWT 角向特征注入 §4.2 输出,供 §4.3 CWT-GAPA 使用。""" cwt_out = cwt_phi_encoder( s_res_b, cwt_scalo=cwt_scalo_batch.to(device) if cwt_scalo_batch is not None else None, ) if "phi_cwt" in cwt_out: o42["phi_cwt"] = cwt_out["phi_cwt"] o42["cwt_scalo"] = cwt_out["cwt_scalo"] return o42 def forward_pipeline_batch_from_dataloader_cwt( fea_tensor, img_tensor: torch.Tensor, enc41: Section41Encoder, enc42: Section42Encoder, enc43: "Section43Encoder", *, half_norm: float, e_pen: Optional[torch.Tensor] = None, training: bool = False, cwt_phi_encoder: nn.Module, cwt_scalo_batch: Optional[torch.Tensor] = None, ) -> Tuple[Dict[str, torch.Tensor], Dict[str, torch.Tensor], Dict[str, torch.Tensor]]: """ CWT-GAPA:复用 section43 原始 §4.1/§4.2 流水线,仅在 enc43 前注入 CWT。 §4.3 ConvNeXt1D 金字塔与 stage_proj 与 DBSA/global 完全一致,便于从旧 checkpoint 微调。 """ device = next(enc43.parameters()).device o41, o42 = forward_pipeline_batch_to_upstream( fea_tensor, img_tensor, enc41, enc42, half_norm=half_norm ) s_res_b = fea_tensor.fields.s_res.float().to(device) o42 = enrich_o42_with_cwt(o42, s_res_b, cwt_phi_encoder, cwt_scalo_batch, device) enc43_kwargs = build_enc43_valid_mask_kwargs(enc43, fea_tensor, device) o43 = enc43.forward_from_upstream( o41, o42, e_pen, training=training, return_details=not training, **enc43_kwargs ) return o41, o42, o43 if __name__ == "__main__": enc = CWTPhiEncoder() s_res = torch.randn(2, 360, 4) out = enc(s_res, return_scalogram=True) print("phi_cwt", tuple(out["phi_cwt"].shape)) print("cwt_scalo", tuple(out["cwt_scalo"].shape))