""" 论文 §4.3 段级动态多尺度融合 — PyTorch 可训练封装。 §4.3.1 角向下采样 → 共享 Φ_ms + FiLM(k) → 段内对齐 H̃_k^(m) §4.3.2 频率引导尺度注意力(Query: f̄_freq, g_vec, E_pen) §4.3.3 段融合 → Transformer → f_ellipse 默认 K={1,2,4,8,16,32},D_h=128,L_ell=3(与 §7.1.3 一致)。 """ from __future__ import annotations from dataclasses import dataclass, field from pathlib import Path from typing import Dict, List, Optional, Sequence, Tuple, Union import numpy as np import torch import torch.nn as nn import torch.nn.functional as F from elipse_feature import DEFAULT_IMG_SIZE, FeaTensor from section41 import ( Section41Config, Section41Encoder, build_u_man, extract_section41, fea_fields_from_dataloader, img_tensor_to_bgr, keypoints_px_from_u_man, ) from section42 import ( Section42Config, Section42Encoder, build_segment_geometry, responsibility_indices, ) PathLike = Union[str, Path] DEFAULT_SCALES: Tuple[int, ...] = (1, 2, 4, 8, 16, 32) @dataclass class Section43Config: n_theta: int = 360 n_segments: int = 10 scales: Tuple[int, ...] = DEFAULT_SCALES d_seq: int = 128 c1: int = 64 d_h: int = 128 d_e: int = 16 d_freq: int = 64 d_geo: int = 32 d_pen: int = 16 h_scale: int = 4 d_prime: int = 32 l_ell: int = 3 n_heads_ell: int = 4 ffn_ell: int = 256 conv1_kernel: int = 7 # --------------------------------------------------------------------------- # §4.3.1.1 角向下采样 # --------------------------------------------------------------------------- def compute_Lk_raw(n_theta: int, k: int) -> int: """ 角向下采样后粗序列长度 L_k^raw ≈ N_θ / k。 推理:角标 t_i = i·k,i = 0, …, L-1,L = ⌊N_θ/k⌋(N_θ=360 时 k∈{8,16,32} 恰为 45/22/11)。 旧实现用训练约束公式 I_k=⌊(N-1)/k⌋-(k-1),在 k≥8 时严重偏短,已弃用。 """ return max(1, n_theta // k) def multiscale_sample_indices( n_theta: int, k: int, *, training: bool = False, eps: Optional[int] = None, ) -> torch.Tensor: """ 返回角标序列 (L_k^raw,),长度 L ≈ N_θ/k。 推理:t_i = i·k,i=0,…,L-1。 训练:t_i = (i+ε)·k,ε∼U{0,…,k-1} 与同次前向共享,再 clip 到 [0, N_θ-1]。 """ L = compute_Lk_raw(n_theta, k) if training: e = int(torch.randint(0, k, (1,)).item()) if eps is None else int(eps) idx = [min((i + e) * k, n_theta - 1) for i in range(L)] else: idx = [i * k for i in range(L)] return torch.tensor(idx, dtype=torch.long) # --------------------------------------------------------------------------- # §4.3.1.2 共享多尺度 CNN + FiLM # --------------------------------------------------------------------------- class SharedMultiscaleStem(nn.Module): """ψ^(1) + FiLM(k) + MaxPool + ψ^tail,输入 (B, L, d_seq) → (B, L_pool, d_h)。""" def __init__(self, cfg: Section43Config, n_scales: int): super().__init__() self.cfg = cfg pad1 = cfg.conv1_kernel // 2 self.conv1 = nn.Conv1d(cfg.d_seq, cfg.c1, kernel_size=cfg.conv1_kernel, padding=pad1) self.scale_embed = nn.Embedding(n_scales, cfg.d_e) self.film_gamma = nn.Linear(cfg.d_e, cfg.c1) self.film_beta = nn.Linear(cfg.d_e, cfg.c1) self.pool = nn.MaxPool1d(kernel_size=2, stride=2) pad3 = 1 self.conv2 = nn.Conv1d(cfg.c1, cfg.d_h, kernel_size=3, padding=pad3) self.conv3 = nn.Conv1d(cfg.d_h, cfg.d_h, kernel_size=3, padding=pad3) self.act = nn.GELU() def forward(self, bar_s: torch.Tensor, scale_id: int) -> torch.Tensor: """ bar_s: (B, L_raw, d_seq) 返回 H_k: (B, L_pool, d_h) """ x = bar_s.transpose(1, 2) z = self.act(self.conv1(x)) e = self.scale_embed(torch.tensor(scale_id, device=z.device, dtype=torch.long)) gamma = self.film_gamma(e).unsqueeze(-1) beta = self.film_beta(e).unsqueeze(-1) z = gamma * z + beta z = self.pool(z) z = self.act(self.conv2(z)) z = self.act(self.conv3(z)) return z.transpose(1, 2) def _interp1d_positions(feat: torch.Tensor, positions: torch.Tensor) -> torch.Tensor: """ feat: (B, L, C),positions: (s_l,) 浮点索引 ∈ [0, L-1] 返回 (B, s_l, C) """ b, L, c = feat.shape s_l = positions.shape[0] pos = positions.clamp(0, max(L - 1, 0)) i0 = pos.long() i1 = torch.clamp(i0 + 1, max=L - 1) w = (pos - i0.float()).view(1, s_l, 1) f0 = feat[:, i0, :] f1 = feat[:, i1, :] return f0 * (1.0 - w) + f1 * w def align_H_to_segment( H_k: torch.Tensor, seg_theta_idx: np.ndarray, n_theta: int, k: int, ) -> torch.Tensor: """ §4.3.1.3:H_k (B, L_pool, d_h) 对齐到段内 s_l 个角锚 → (B, s_l, d_h)。 角锚 θ 映射到粗网格池化后坐标:θ / k / 2。 """ L_pool = H_k.shape[1] pos = torch.tensor(seg_theta_idx, dtype=torch.float32, device=H_k.device) / float(k) / 2.0 pos = pos.clamp(0, max(L_pool - 1, 0)) return _interp1d_positions(H_k, pos) # --------------------------------------------------------------------------- # §4.3.2 频率引导尺度注意力 # --------------------------------------------------------------------------- class FreqGuidedScaleAttention(nn.Module): def __init__(self, cfg: Section43Config, n_scales: int): super().__init__() self.cfg = cfg self.n_scales = n_scales d_qinp = cfg.d_freq + cfg.d_geo + cfg.d_pen self.q_proj = nn.Linear(d_qinp, cfg.h_scale * cfg.d_prime) self.w_k = nn.ModuleList( [nn.Linear(cfg.d_h, cfg.d_prime, bias=False) for _ in range(cfg.h_scale)] ) def forward( self, f_bar: torch.Tensor, g_vec: torch.Tensor, e_pen: torch.Tensor, t_stack: torch.Tensor, ) -> Tuple[torch.Tensor, torch.Tensor]: """ f_bar: (B, d_freq) 段级频域均值 g_vec: (B, d_geo) e_pen: (B, d_pen) t_stack: (B, n_scales, d_h) 各尺度段摘要 t_k^(m) 返回: W_scale: (B, n_scales) 归一化尺度权重 beta_heads: (B, h_scale, n_scales) """ u = torch.cat([f_bar, g_vec, e_pen], dim=-1) q_all = self.q_proj(u) q_heads = q_all.view(-1, self.cfg.h_scale, self.cfg.d_prime) beta_list = [] scale = self.cfg.d_prime**0.5 for hi in range(self.cfg.h_scale): q = q_heads[:, hi, :] k_proj = self.w_k[hi](t_stack) logits = torch.einsum("bd,bkd->bk", q, k_proj) / scale beta_list.append(F.softmax(logits, dim=-1)) beta_heads = torch.stack(beta_list, dim=1) beta_mean = beta_heads.mean(dim=1) W_scale = beta_mean / (beta_mean.sum(dim=-1, keepdim=True) + 1e-12) return W_scale, beta_heads # --------------------------------------------------------------------------- # §4.3 完整编码器 # --------------------------------------------------------------------------- class Section43Encoder(nn.Module): """§4.3 段级动态多尺度融合(全部可训练)。""" def __init__(self, config: Optional[Section43Config] = None): super().__init__() self.config = config or Section43Config() cfg = self.config self.scales = list(cfg.scales) self.n_scales = len(self.scales) self.geom = build_segment_geometry(cfg.n_theta, cfg.n_segments) self.s_l = self.geom.s_l self.M = self.geom.n_segments self.ms_stem = SharedMultiscaleStem(cfg, self.n_scales) self.scale_attn = FreqGuidedScaleAttention(cfg, self.n_scales) enc_layer = nn.TransformerEncoderLayer( d_model=cfg.d_h, nhead=cfg.n_heads_ell, dim_feedforward=cfg.ffn_ell, batch_first=True, norm_first=True, activation="gelu", ) self.ell_transformer = nn.TransformerEncoder(enc_layer, num_layers=cfg.l_ell) def _encode_scale( self, s_seq: torch.Tensor, k: int, scale_id: int, training: bool, ) -> torch.Tensor: """单尺度 H_k: (B, L_pool, d_h)。""" b, n, _ = s_seq.shape idx = multiscale_sample_indices(n, k, training=training).to(s_seq.device) bar_s = s_seq[:, idx, :] return self.ms_stem(bar_s, scale_id) def _align_all_segments( self, H_k: torch.Tensor, k: int, ) -> torch.Tensor: """(B, M, s_l, d_h)。""" b = H_k.shape[0] out = [] for m in range(self.M): seg_idx = responsibility_indices(m, self.geom) h_m = align_H_to_segment(H_k, seg_idx, self.config.n_theta, k) out.append(h_m.unsqueeze(1)) return torch.cat(out, dim=1) def forward( self, s_seq: torch.Tensor, f_freq: torch.Tensor, g_vec: torch.Tensor, e_pen: Optional[torch.Tensor] = None, *, training: bool = False, return_details: bool = True, ) -> Dict[str, torch.Tensor]: """ s_seq: (B, N_theta, d_seq) f_freq: (B, M, d_freq, s_l) — 与 section42 输出一致 g_vec: (B, d_geo) e_pen: (B, d_pen) 或 None(未知笔型用零向量) """ cfg = self.config batched = s_seq.dim() == 3 if not batched: s_seq = s_seq.unsqueeze(0) f_freq = f_freq.unsqueeze(0) g_vec = g_vec.unsqueeze(0) b = s_seq.size(0) device = s_seq.device if e_pen is None: e_pen = torch.zeros(b, cfg.d_pen, device=device, dtype=s_seq.dtype) elif e_pen.dim() == 1: e_pen = e_pen.unsqueeze(0).expand(b, -1) f_bar = f_freq.mean(dim=-1) H_aligned: List[torch.Tensor] = [] for sid, k in enumerate(self.scales): H_k = self._encode_scale(s_seq, k, sid, training) H_tilde = self._align_all_segments(H_k, k) H_aligned.append(H_tilde) H_stack = torch.stack(H_aligned, dim=2) F_seg = torch.zeros(b, self.M, self.s_l, cfg.d_h, device=device, dtype=s_seq.dtype) W_all = torch.zeros(b, self.M, self.n_scales, device=device, dtype=s_seq.dtype) beta_all = torch.zeros(b, self.M, cfg.h_scale, self.n_scales, device=device) for m in range(self.M): t_stack = H_stack[:, m, :, :].mean(dim=2) W_m, beta_m = self.scale_attn(f_bar[:, m], g_vec, e_pen, t_stack) W_all[:, m] = W_m beta_all[:, m] = beta_m for sid in range(self.n_scales): F_seg[:, m] = F_seg[:, m] + W_m[:, sid : sid + 1, None] * H_stack[:, m, sid] H_ell = F_seg.reshape(b, self.M * self.s_l, cfg.d_h) H_ell = self.ell_transformer(H_ell) f_ellipse = H_ell.max(dim=1).values out: Dict[str, torch.Tensor] = { "f_ellipse": f_ellipse, "H_ell": H_ell, "F_segment": F_seg, "W_scale": W_all, } if return_details: out["H_aligned"] = H_stack out["beta_heads"] = beta_all out["f_bar"] = f_bar if not batched: out = {k: v.squeeze(0) for k, v in out.items()} return out def forward_from_upstream( self, o41: Dict[str, torch.Tensor], o42: Dict[str, torch.Tensor], e_pen: Optional[torch.Tensor] = None, *, training: bool = False, return_details: bool = True, **kwargs, ) -> Dict[str, torch.Tensor]: """§4.1 + §4.2 输出字典 → §4.3(s_seq / f_freq / g_vec)。""" return self.forward( o41["s_seq"], o42["f_freq"], o41["g_vec"], e_pen, training=training, return_details=return_details, ) 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, ) -> "Section43Encoder": payload = torch.load(path, map_location=map_location, weights_only=False) cfg = payload.get("config") or Section43Config() if isinstance(cfg, dict): cfg = Section43Config(**cfg) model = cls(cfg) model.load_state_dict(payload["state_dict"], strict=strict) return model class EllipseBranch43(nn.Module): """§4.1 → §4.2 → §4.3 串联(训练时可仅优化 section43 或端到端)。""" def __init__( self, cfg41: Optional[Section41Config] = None, cfg42: Optional[Section42Config] = None, cfg43: Optional[Section43Config] = None, *, freeze41_deep: bool = True, ): super().__init__() self.enc41 = Section41Encoder(cfg41) self.enc42 = Section42Encoder(cfg42) self.enc43 = Section43Encoder(cfg43) if freeze41_deep: for p in self.enc41.deep_branch.parameters(): p.requires_grad = False def forward( self, rgb: PathLike, e_pen: Optional[torch.Tensor] = None, *, training: bool = False, device: Optional[torch.device] = None, ) -> Dict[str, torch.Tensor]: from elipse_feature import extract_ellipse_features, preprocess_rgb_to_skeleton prep = preprocess_rgb_to_skeleton(rgb, size=self.enc41.config.img_size) bundle = extract_ellipse_features( prep["skeleton"], n_theta=self.enc43.config.n_theta, half_norm=self.enc41.config.img_size / 2.0, ) s_res = torch.from_numpy(bundle.sequence.s_res).float() if device: s_res = s_res.to(device) with torch.no_grad(): h_deep = self.enc41.deep_branch( prep["rgb"], bundle.sequence.s_base_px, patch_size=self.enc41.config.patch_size, ) g = torch.from_numpy(bundle.global_feat.as_vector()).float() u = torch.from_numpy(build_u_man(bundle.sequence)).float() if device: g, u, h_deep = g.to(device), u.to(device), h_deep.to(device) o41 = self.enc41.forward_tensors(g, u, h_deep) o42 = self.enc42(s_res.unsqueeze(0) if s_res.dim() == 2 else s_res) s_seq = o41["s_seq"] if s_seq.dim() == 2: s_seq = s_seq.unsqueeze(0) g_vec = o41["g_vec"].unsqueeze(0) f_freq = o42["f_freq"].unsqueeze(0) if o42["f_freq"].dim() == 3 else o42["f_freq"] else: g_vec = o41["g_vec"] f_freq = o42["f_freq"] o43 = self.enc43( s_seq, f_freq, g_vec, e_pen, training=training, ) return {"prep": prep, "bundle": bundle, **{f"o41_{k}": v for k, v in o41.items()}, **{f"o42_{k}": v for k, v in o42.items()}, **o43} def extract_section43( s_seq: torch.Tensor, f_freq: torch.Tensor, g_vec: torch.Tensor, e_pen: Optional[torch.Tensor] = None, config: Optional[Section43Config] = None, encoder: Optional[Section43Encoder] = None, device: Optional[str] = None, training: bool = False, ) -> Dict[str, torch.Tensor]: enc = encoder or Section43Encoder(config) if device: enc = enc.to(device) s_seq = s_seq.to(device) f_freq = f_freq.to(device) g_vec = g_vec.to(device) enc.eval() if not training else enc.train() return enc(s_seq, f_freq, g_vec, e_pen, training=training) def forward_pipeline_from_dataloader( fea_tensor: FeaTensor, img_tensor: torch.Tensor, enc41: Section41Encoder, enc42: Section42Encoder, enc43: Section43Encoder, *, half_norm: float = DEFAULT_IMG_SIZE / 2.0, e_pen: Optional[torch.Tensor] = None, training: bool = False, ) -> Tuple[Dict[str, torch.Tensor], Dict[str, torch.Tensor], Dict[str, torch.Tensor]]: """ DataLoader 单样本:fea_tensor + img_tensor → §4.1 → §4.2 → §4.3。 """ fields = fea_fields_from_dataloader(fea_tensor) if fields.u_man.dim() == 3: raise NotImplementedError("单样本 FeaTensor 请传入 (N,8) 而非 batch 维") device = next(enc43.parameters()).device g_t = fields.g_global.float().to(device) u_t = fields.u_man.float().to(device) pts = keypoints_px_from_u_man(u_t, half_norm=half_norm) bgr = img_tensor_to_bgr(img_tensor) h_deep = enc41.extract_h_deep(bgr, pts) o41 = enc41.forward_tensors(g_t, u_t, h_deep) o42 = enc42.forward_from_fea(fea_tensor) 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 def build_enc43_valid_mask_kwargs( enc43: Section43Encoder, fea_tensor: FeaTensor, device: torch.device, ) -> Dict[str, torch.Tensor]: """为 enc43.forward_from_upstream 构造 valid_mask(v2/v3.1)。""" enc43_kwargs: Dict[str, torch.Tensor] = {} try: from section43_v2 import build_pool_valid_mask, v2_needs_valid_mask from section43_v3_1 import v3_1_needs_valid_mask if v2_needs_valid_mask(enc43) or v3_1_needs_valid_mask(enc43): valid = fea_tensor.fields.valid.float().to(device) if valid.dim() == 1: valid = valid.unsqueeze(0) if getattr(enc43.config, "stage_pool_mode", None) in ("cwt_angular", "cwt_dual"): enc43_kwargs["valid_mask"] = valid else: enc43_kwargs["valid_mask"] = build_pool_valid_mask( valid, enc43.geom, enc43.M, enc43.s_l, ) except ImportError: pass return enc43_kwargs def forward_pipeline_batch_to_upstream( fea_tensor: FeaTensor, img_tensor: torch.Tensor, enc41: Section41Encoder, enc42: Section42Encoder, *, half_norm: float = DEFAULT_IMG_SIZE / 2.0, ) -> Tuple[Dict[str, torch.Tensor], Dict[str, torch.Tensor]]: """DataLoader batch:§4.1 + §4.2,不含 §4.3(供 CWT 等扩展在 enc43 前注入 o42)。""" device = next(enc42.parameters()).device fields = fea_tensor.fields g_b = fields.g_global.float().to(device) u_b = fields.u_man.float().to(device) if enc41.deep_branch.backend == "resnet": h_b = enc41.extract_h_deep_batch(img_tensor, u_b, half_norm=half_norm) else: h_list = [] for i in range(img_tensor.shape[0]): pts = keypoints_px_from_u_man(u_b[i], half_norm=half_norm) bgr = img_tensor_to_bgr(img_tensor[i]) h_list.append(enc41.extract_h_deep(bgr, pts)) h_b = torch.stack(h_list, dim=0) o41 = enc41.forward_tensors(g_b, u_b, h_b) s_res_b = fields.s_res.float().to(device) o42 = enc42.forward( s_res_b, amp_segment=fields.amp_segment.float().to(device), amp_global=fields.amp_global.float().to(device), ) return o41, o42 def forward_pipeline_batch_from_dataloader( fea_tensor: FeaTensor, img_tensor: torch.Tensor, enc41: Section41Encoder, enc42: Section42Encoder, enc43: Section43Encoder, *, half_norm: float = DEFAULT_IMG_SIZE / 2.0, e_pen: Optional[torch.Tensor] = None, training: bool = False, ) -> Tuple[Dict[str, torch.Tensor], Dict[str, torch.Tensor], Dict[str, torch.Tensor]]: """DataLoader batch:逐样本 deep + batched §4.1/§4.2/§4.3(§4.3 ConvNeXt1D 路径未改)。""" device = next(enc43.parameters()).device o41, o42 = forward_pipeline_batch_to_upstream( fea_tensor, img_tensor, enc41, enc42, half_norm=half_norm ) 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 def demo_from_dataloader( sample_ratio: float = 0.001, batch_size: int = 2, half_norm: float = DEFAULT_IMG_SIZE / 2.0, cfg41: Optional[Section41Config] = None, cfg42: Optional[Section42Config] = None, cfg43: Optional[Section43Config] = None, device: str = "cpu", deep_backend: str = "resnet", ) -> Dict[str, torch.Tensor]: """ 从 exp6_data_loader 取样本:§4.1(s_seq, g_vec)+ §4.2(f_freq)→ §4.3。 """ from exp6_data_loader import create_dataloaders cfg41 = cfg41 or Section41Config( deep_backend=deep_backend, resnet_checkpoint=( None if deep_backend == "patch" else Section41Config().resnet_checkpoint ), ) cfg42 = cfg42 or Section42Config() cfg43 = cfg43 or Section43Config() enc41 = Section41Encoder(cfg41).to(device) enc42 = Section42Encoder(cfg42).to(device) enc43 = Section43Encoder(cfg43).to(device) enc41.eval() enc42.eval() enc43.eval() 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("[section43] dataloader batch") print(f" fea_tensor.data: {fea_tensor.data.shape}") print(f" img_tensor: {tuple(img_tensor.shape)} label: {labels.tolist()}") print(f" deep_backend: {cfg41.deep_backend} scales: {cfg43.scales}") fea_one = FeaTensor(data=fea_tensor.data[0]) img_one = img_tensor[0] fields = fea_one.fields print(f" g_global: {tuple(fields.g_global.shape)} u_man: {tuple(fields.u_man.shape)}") print(f" amp_segment: {tuple(fields.amp_segment.shape)} amp_global: {tuple(fields.amp_global.shape)}") with torch.no_grad(): o41, o42, o43 = forward_pipeline_from_dataloader( fea_one, img_one, enc41, enc42, enc43, half_norm=half_norm, ) n41 = sum(p.numel() for p in enc41.trainable_parameters()) n42 = sum(p.numel() for p in enc42.trainable_parameters()) n43 = sum(p.numel() for p in enc43.parameters()) print("[section43] pipeline single (41→42→43)") print(f" writer: {writers[0]} image_id: {image_ids[0]}") print(f" trainable §4.1: {n41:,} §4.2: {n42:,} §4.3: {n43:,}") print(f" s_seq: {tuple(o41['s_seq'].shape)}") print(f" g_vec: {tuple(o41['g_vec'].shape)}") print(f" f_freq: {tuple(o42['f_freq'].shape)}") print(f" f_bar: {tuple(o42['f_bar'].shape)}") print(f" f_ellipse: {tuple(o43['f_ellipse'].shape)}") print(f" F_segment: {tuple(o43['F_segment'].shape)}") print(f" W_scale: {tuple(o43['W_scale'].shape)}") print(f" H_ell: {tuple(o43['H_ell'].shape)}") ws = o43["W_scale"] if ws.dim() == 2: w0 = ws[0].detach().cpu().numpy() else: w0 = ws.detach().cpu().numpy().ravel() print(f" 段0尺度权重: {dict(zip(DEFAULT_SCALES, [f'{float(x):.3f}' for x in w0]))}") print("[section43] pipeline batch") with torch.no_grad(): o41_b, o42_b, o43_b = forward_pipeline_batch_from_dataloader( fea_tensor, img_tensor, enc41, enc42, enc43, half_norm=half_norm, ) print(f" batch s_seq: {tuple(o41_b['s_seq'].shape)}") print(f" batch f_freq: {tuple(o42_b['f_freq'].shape)}") print(f" batch f_ellipse: {tuple(o43_b['f_ellipse'].shape)}") print(f" batch W_scale: {tuple(o43_b['W_scale'].shape)}") return o43 def demo(image_path: str = "data/images/20938.png", device: str = "cpu") -> Dict[str, torch.Tensor]: branch = EllipseBranch43().to(device) branch.eval() with torch.no_grad(): out = branch.forward(image_path, device=torch.device(device)) n_params = sum(p.numel() for p in branch.enc43.parameters()) print(f"§4.3 demo: {image_path}") print(f" enc43 参数: {n_params:,}") print(f" f_ellipse {out['f_ellipse'].shape}") print(f" W_scale {out['W_scale'].shape}") print(f" F_segment {out['F_segment'].shape}") print(f" H_ell {out['H_ell'].shape}") ws = out["W_scale"] if ws.dim() == 3: ws = ws[0, 0] elif ws.dim() == 2: ws = ws[0] w0 = ws.detach().cpu().numpy().ravel() print(f" 段0尺度权重: {dict(zip(DEFAULT_SCALES, [f'{float(x):.3f}' for x in w0]))}") return out def main() -> None: import argparse p = argparse.ArgumentParser(description="§4.3 multiscale fusion encoder") p.add_argument("image", nargs="?", default="data/images/20938.png") p.add_argument("--device", default="cpu") p.add_argument("--save", type=str, default=None) p.add_argument("--ckpt", type=str, default=None) p.add_argument( "--from-dataloader", action="store_true", help="使用 exp6_data_loader → §4.1 → §4.2 → §4.3 调试", ) p.add_argument("--sample-ratio", type=float, default=0.001) p.add_argument("--batch-size", type=int, default=2) p.add_argument( "--deep", choices=("resnet", "patch"), default="resnet", help=f"§4.1.3 深度分支(默认 resnet + 平台默认权重)", ) args = p.parse_args() if args.from_dataloader: demo_from_dataloader( sample_ratio=args.sample_ratio, batch_size=args.batch_size, device=args.device, deep_backend=args.deep, ) return out = demo(args.image, args.device) if args.ckpt: EllipseBranch43().enc43.save_checkpoint(args.ckpt) print("saved enc43:", args.ckpt) if args.save: np.savez( args.save, f_ellipse=out["f_ellipse"].cpu().numpy(), W_scale=out["W_scale"].cpu().numpy(), ) print("saved:", args.save) if __name__ == "__main__": demo_from_dataloader()