""" 论文 §4.1 椭圆几何建模与序列化表征 — PyTorch 可训练封装。 §4.1.1 MLP_geo(g_global) → g_vec [可训练] §4.1.2 MLP_man(u_man) → s_man [可训练] §4.1.3 ResNet grid_sample → raw_deep [冻结] proj_deep(raw_deep) → h_deep [可训练] §4.1.4 W_g, MHCA, Transformer → s_seq [可训练] 训练时:optimizer 仅传入 model.trainable_parameters()。 """ from __future__ import annotations from dataclasses import dataclass, field from pathlib import Path from typing import Dict, List, Optional, 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, DEFAULT_N_THETA, EllipseFeatureBundle, FeaFields, FeaTensor, SequenceEllipseFeatures, extract_ellipse_features, preprocess_rgb_to_skeleton, unpack_fea_fields, ) from resnet_point_feature import default_resnet_checkpoint from convnext_point_feature import default_convnext_checkpoint PathLike = Union[str, Path, np.ndarray] U_MAN_DIM = 8 G_GLOBAL_DIM = 6 @dataclass class Section41Config: n_theta: int = DEFAULT_N_THETA img_size: int = DEFAULT_IMG_SIZE d_geo: int = 32 d_man: int = 64 d_deep: int = 128 d_attn: int = 128 d_seq: int = 128 h_geo: int = 4 l_seq: int = 2 n_heads_seq: int = 4 ffn_dim: int = 256 patch_size: int = 7 mlp_geo_hidden: Tuple[int, ...] = (32,) mlp_man_hidden: Tuple[int, ...] = (64,) deep_backend: str = "convnext" resnet_checkpoint: Optional[str] = field(default_factory=default_resnet_checkpoint) convnext_checkpoint: Optional[str] = field(default_factory=default_convnext_checkpoint) @dataclass class Section41Output: """§4.1 输出(张量字段在 eval 时可 .numpy())。""" prep: dict bundle: Optional[EllipseFeatureBundle] g_global: np.ndarray g_vec: np.ndarray u_man: np.ndarray s_man: np.ndarray h_deep: np.ndarray z: np.ndarray u_fuse: np.ndarray s_seq: np.ndarray config: Section41Config # --------------------------------------------------------------------------- # 工具 # --------------------------------------------------------------------------- def build_u_man(sequence: SequenceEllipseFeatures) -> np.ndarray: """u_man(t) = [Δr, Δθ, κ, τ, x, y, φ, valid]。""" valid = (1.0 - sequence.interp_flag).astype(np.float32) phi_norm = (sequence.theta_ideal / (2.0 * np.pi)).astype(np.float32) return np.concatenate( [ sequence.s_res[:, :4].astype(np.float32), sequence.s_base.astype(np.float32), phi_norm[:, None], valid[:, None], ], axis=1, ) def keypoints_px_from_u_man( u_man: Union[np.ndarray, torch.Tensor], half_norm: float = DEFAULT_IMG_SIZE / 2.0, ) -> np.ndarray: """ 由 u_man[:, 4:6](归一化 s_base)还原像素坐标 (x, y),供 §4.1.3 ResNet 采样。 half_norm 须与 DataLoader 中 extract_ellipse_features 一致(默认 50)。 """ if isinstance(u_man, torch.Tensor): s_base = u_man[..., 4:6].detach().cpu().numpy() else: s_base = np.asarray(u_man)[..., 4:6] return (s_base * half_norm).astype(np.float32) def keypoints_px_tensor( u_man: torch.Tensor, half_norm: float = DEFAULT_IMG_SIZE / 2.0, ) -> torch.Tensor: """u_man (..., 8) → 像素坐标 (..., 2),保留 device/dtype 供批量 ResNet 采样。""" kps = u_man[..., 4:6].float() * half_norm return kps def img_tensor_to_bgr( img_tensor: torch.Tensor, mean: Tuple[float, ...] = (0.485, 0.456, 0.406), std: Tuple[float, ...] = (0.229, 0.224, 0.225), ) -> np.ndarray: """DataLoader 输出的归一化 RGB (3,H,W) → BGR uint8,供 deep_branch 使用。""" import cv2 t = img_tensor.detach().cpu().float() if t.dim() != 3: raise ValueError(f"img_tensor 须为 (3,H,W),got {tuple(t.shape)}") for c in range(3): t[c] = t[c] * std[c] + mean[c] rgb = (t.permute(1, 2, 0).numpy() * 255.0).clip(0, 255).astype(np.uint8) return cv2.cvtColor(rgb, cv2.COLOR_RGB2BGR) def fea_fields_from_dataloader(fea_tensor: FeaTensor) -> FeaFields: """从 DataLoader 输出的 FeaTensor 反解各字段。""" if isinstance(fea_tensor, FeaTensor): return fea_tensor.fields return unpack_fea_fields(fea_tensor) def _build_mlp( in_dim: int, hidden: Tuple[int, ...], out_dim: int, ) -> nn.Sequential: layers: List[nn.Module] = [] dims = [in_dim, *hidden, out_dim] for i in range(len(dims) - 2): layers.extend([nn.Linear(dims[i], dims[i + 1]), nn.ReLU(inplace=True)]) layers.append(nn.Linear(dims[-2], dims[-1])) return nn.Sequential(*layers) def _freeze_module(module: nn.Module) -> None: module.eval() for p in module.parameters(): p.requires_grad = False # --------------------------------------------------------------------------- # §4.1.3 冻结深度分支 # --------------------------------------------------------------------------- class FrozenDeepBranch(nn.Module): """ §4.1.3 冻结视觉骨干:仅 ResNet(或 patch 裁剪)输出 raw_deep,不含可训练层。 线性投影 proj_deep 在 Section41Encoder 中。 """ def __init__(self, config: Section41Config): super().__init__() from resnet_point_feature import RESNET_DEEP_RAW_DIM, ResNetFeatureExtractor from convnext_point_feature import CONVNEXT_DEEP_RAW_DIM, ConvNeXtFeatureExtractor self.backend = config.deep_backend.lower() if self.backend not in ("resnet", "convnext", "patch"): raise ValueError(f"deep_backend 需为 resnet、convnext 或 patch, got {self.backend}") if self.backend == "resnet": self.resnet = ResNetFeatureExtractor( checkpoint_path=config.resnet_checkpoint, ) self.convnext = None _freeze_module(self.resnet) self.raw_dim = RESNET_DEEP_RAW_DIM elif self.backend == "convnext": self.resnet = None self.convnext = ConvNeXtFeatureExtractor( checkpoint_path=config.convnext_checkpoint, ) _freeze_module(self.convnext) self.raw_dim = CONVNEXT_DEEP_RAW_DIM else: self.resnet = None self.convnext = None self.raw_dim = 3 * config.patch_size * config.patch_size @torch.no_grad() def _extract_raw_resnet( self, rgb_bgr: np.ndarray, points_xy: np.ndarray, ) -> torch.Tensor: from resnet_point_feature import _rgb_bgr_to_tensor device = next(self.parameters()).device image_tensor, _, _ = _rgb_bgr_to_tensor(rgb_bgr, device) kps = torch.from_numpy(points_xy.astype(np.float32)).unsqueeze(0).to(device) _, _, h, w = image_tensor.shape kps[..., 0] = kps[..., 0].clamp(0, w - 1) kps[..., 1] = kps[..., 1].clamp(0, h - 1) return self.resnet(image_tensor, kps).squeeze(0) @torch.no_grad() def _extract_raw_resnet_batch( self, images: torch.Tensor, keypoints: torch.Tensor, ) -> torch.Tensor: """images [B,3,H,W],keypoints [B,N,2] → raw_deep [B,N,raw_dim]。""" _, _, h, w = images.shape kps = keypoints.clone() kps[..., 0] = kps[..., 0].clamp(0, w - 1) kps[..., 1] = kps[..., 1].clamp(0, h - 1) return self.resnet(images, kps) @torch.no_grad() def _extract_raw_convnext( self, rgb_bgr: np.ndarray, points_xy: np.ndarray, ) -> torch.Tensor: from resnet_point_feature import _rgb_bgr_to_tensor device = next(self.parameters()).device image_tensor, _, _ = _rgb_bgr_to_tensor(rgb_bgr, device) kps = torch.from_numpy(points_xy.astype(np.float32)).unsqueeze(0).to(device) _, _, h, w = image_tensor.shape kps[..., 0] = kps[..., 0].clamp(0, w - 1) kps[..., 1] = kps[..., 1].clamp(0, h - 1) return self.convnext(image_tensor, kps).squeeze(0) @torch.no_grad() def _extract_raw_convnext_batch( self, images: torch.Tensor, keypoints: torch.Tensor, ) -> torch.Tensor: _, _, h, w = images.shape kps = keypoints.clone() kps[..., 0] = kps[..., 0].clamp(0, w - 1) kps[..., 1] = kps[..., 1].clamp(0, h - 1) return self.convnext(images, kps) @torch.no_grad() def _extract_raw_patch( self, rgb_bgr: np.ndarray, points_xy: np.ndarray, patch_size: int, ) -> torch.Tensor: import cv2 h, w = rgb_bgr.shape[:2] r = patch_size // 2 rgb = cv2.cvtColor(rgb_bgr, cv2.COLOR_BGR2RGB).astype(np.float32) / 255.0 n = points_xy.shape[0] patches = [] for t in range(n): cx = int(round(points_xy[t, 0])) cy = int(round(points_xy[t, 1])) patch = np.zeros((patch_size, patch_size, 3), dtype=np.float32) for dy in range(-r, r + 1): for dx in range(-r, r + 1): y = int(np.clip(cy + dy, 0, h - 1)) x = int(np.clip(cx + dx, 0, w - 1)) patch[dy + r, dx + r] = rgb[y, x] patches.append(patch.reshape(-1)) return torch.from_numpy(np.stack(patches, axis=0)) def forward( self, rgb_bgr: np.ndarray, points_xy: np.ndarray, patch_size: int = 7, ) -> torch.Tensor: """返回 raw_deep: (N, raw_dim),无梯度。""" if self.backend == "resnet": return self._extract_raw_resnet(rgb_bgr, points_xy) if self.backend == "convnext": return self._extract_raw_convnext(rgb_bgr, points_xy) return self._extract_raw_patch(rgb_bgr, points_xy, patch_size) # --------------------------------------------------------------------------- # §4.1.4 可训练注意力与 Transformer # --------------------------------------------------------------------------- class GeoMultiHeadCrossAttention(nn.Module): """全局 q_global 广播,K/V 来自 s_man(论文 §4.1.4 MHCA)。""" def __init__(self, d_man: int, d_attn: int, n_heads: int): super().__init__() if d_attn % n_heads != 0: raise ValueError("d_attn 须能被 n_heads 整除") self.n_heads = n_heads self.d_head = d_attn // n_heads self.w_k = nn.Linear(d_man, d_attn, bias=False) self.w_v = nn.Linear(d_man, d_attn, bias=False) self.w_o = nn.Linear(d_attn, d_attn, bias=True) def forward(self, q_global: torch.Tensor, s_man: torch.Tensor) -> torch.Tensor: """ q_global: (d_attn,) 或 (B, d_attn) s_man: (N, d_man) 或 (B, N, d_man) 返回 z: (N, d_attn) 或 (B, N, d_attn) """ batched = q_global.dim() == 2 if not batched: q_global = q_global.unsqueeze(0) s_man = s_man.unsqueeze(0) b, n, _ = s_man.shape h, d = self.n_heads, self.d_head k = self.w_k(s_man).view(b, n, h, d) v = self.w_v(s_man).view(b, n, h, d) q = q_global.view(b, h, d) scale = d**0.5 scores = torch.einsum("bhd,bnhd->bhn", q, k) / scale attn = F.softmax(scores, dim=-1) z_heads = torch.einsum("bhn,bnhd->bhd", attn, v) z = z_heads.unsqueeze(1).expand(-1, n, -1, -1).reshape(b, n, h * d) z = self.w_o(z) if not batched: z = z.squeeze(0) return z class Section41Encoder(nn.Module): """ §4.1 完整编码器(nn.Module)。 可训练:mlp_geo, mlp_man, proj_deep, w_g, mhca, fuse_norm, transformer, proj_seq 冻结:deep_branch(ResNet 骨干) """ def __init__(self, config: Optional[Section41Config] = None): super().__init__() self.config = config or Section41Config() cfg = self.config d_fuse = cfg.d_attn + cfg.d_deep self.mlp_geo = _build_mlp(G_GLOBAL_DIM, cfg.mlp_geo_hidden, cfg.d_geo) self.mlp_man = _build_mlp(U_MAN_DIM, cfg.mlp_man_hidden, cfg.d_man) self.w_g = nn.Linear(cfg.d_geo, cfg.d_attn, bias=False) self.mhca = GeoMultiHeadCrossAttention(cfg.d_man, cfg.d_attn, cfg.h_geo) self.deep_branch = FrozenDeepBranch(cfg) self.proj_deep = nn.Linear(self.deep_branch.raw_dim, cfg.d_deep, bias=True) self.fuse_norm = nn.LayerNorm(d_fuse) enc_layer = nn.TransformerEncoderLayer( d_model=d_fuse, nhead=cfg.n_heads_seq, dim_feedforward=cfg.ffn_dim, batch_first=True, norm_first=True, activation="gelu", ) self.transformer = nn.TransformerEncoder(enc_layer, num_layers=cfg.l_seq) self.proj_seq = nn.Linear(d_fuse, cfg.d_seq) def trainable_parameters(self) -> List[nn.Parameter]: """§4.1.1 / §4.1.2 / §4.1.3 proj_deep / §4.1.4 可训练参数。""" return [p for p in self.parameters() if p.requires_grad] def trainable_named_parameters(self) -> List[Tuple[str, nn.Parameter]]: return [(n, p) for n, p in self.named_parameters() if p.requires_grad] def frozen_parameters(self) -> List[nn.Parameter]: """仅冻结的 ResNet 骨干参数。""" return [p for p in self.deep_branch.parameters() if not p.requires_grad] def extract_h_deep( self, rgb_bgr: np.ndarray, points_xy: np.ndarray, ) -> torch.Tensor: """§4.1.3:冻结 backbone 采样 raw_deep,再经可训练 proj_deep → h_deep。""" with torch.no_grad(): raw = self.deep_branch( rgb_bgr, points_xy, patch_size=self.config.patch_size, ) return self.proj_deep(raw) def extract_h_deep_batch( self, img_tensor: torch.Tensor, u_man: torch.Tensor, half_norm: float = DEFAULT_IMG_SIZE / 2.0, ) -> torch.Tensor: """§4.1.3 批量:DataLoader 的 img_tensor [B,3,H,W] + u_man [B,N,8]。""" if self.deep_branch.backend not in ("resnet", "convnext"): raise NotImplementedError( "extract_h_deep_batch 仅支持 deep_backend=resnet 或 convnext" ) device = next(self.parameters()).device images = img_tensor.to(device) kps = keypoints_px_tensor(u_man, half_norm=half_norm).to(device) with torch.no_grad(): if self.deep_branch.backend == "resnet": raw = self.deep_branch._extract_raw_resnet_batch(images, kps) else: raw = self.deep_branch._extract_raw_convnext_batch(images, kps) return self.proj_deep(raw) def forward_tensors( self, g_global: torch.Tensor, u_man: torch.Tensor, h_deep: torch.Tensor, ) -> Dict[str, torch.Tensor]: """ 核心可微前向(训练用)。 g_global: (6,) 或 (B, 6) u_man: (N, 8) 或 (B, N, 8) h_deep: (N, d_deep) 或 (B, N, d_deep),由 extract_h_deep 产生(proj_deep 可反传) """ batched = g_global.dim() == 2 if not batched: g_global = g_global.unsqueeze(0) u_man = u_man.unsqueeze(0) h_deep = h_deep.unsqueeze(0) g_vec = self.mlp_geo(g_global) s_man = self.mlp_man(u_man) q_global = self.w_g(g_vec) z = self.mhca(q_global, s_man) u = torch.cat([z, h_deep], dim=-1) h = self.fuse_norm(u) h = self.transformer(h) s_seq = self.proj_seq(h) if not batched: s_seq = s_seq.squeeze(0) g_vec = g_vec.squeeze(0) s_man = s_man.squeeze(0) z = z.squeeze(0) u = u.squeeze(0) return { "g_vec": g_vec, "s_man": s_man, "z": z, "u_fuse": u, "s_seq": s_seq, } def _compute_geometry( self, rgb: PathLike, ) -> Tuple[dict, EllipseFeatureBundle, np.ndarray, np.ndarray, np.ndarray]: cfg = self.config prep = preprocess_rgb_to_skeleton(rgb, size=cfg.img_size) hn = cfg.img_size / 2.0 bundle = extract_ellipse_features( prep["skeleton"], n_theta=cfg.n_theta, half_norm=hn, ) g_global = bundle.global_feat.as_vector() u_man = build_u_man(bundle.sequence) return prep, bundle, g_global, u_man, bundle.sequence.s_base_px def forward( self, rgb: PathLike, *, return_numpy: bool = True, ) -> Section41Output: """单张图完整 §4.1 流程。""" prep, bundle, g_global, u_man, pts = self._compute_geometry(rgb) device = next(self.parameters()).device g_t = torch.from_numpy(g_global).float().to(device) u_t = torch.from_numpy(u_man).float().to(device) h_deep = self.extract_h_deep(prep["rgb"], pts) out_t = self.forward_tensors(g_t, u_t, h_deep) def _np(x: torch.Tensor) -> np.ndarray: return x.detach().cpu().numpy() if return_numpy: return Section41Output( prep=prep, bundle=bundle, g_global=g_global, g_vec=_np(out_t["g_vec"]), u_man=u_man, s_man=_np(out_t["s_man"]), h_deep=_np(h_deep), z=_np(out_t["z"]), u_fuse=_np(out_t["u_fuse"]), s_seq=_np(out_t["s_seq"]), config=self.config, ) raise NotImplementedError("return_numpy=False 时可扩展返回张量字典") def forward_from_fea( self, fea_tensor: FeaTensor, img_bgr: Optional[np.ndarray] = None, img_tensor: Optional[torch.Tensor] = None, half_norm: float = DEFAULT_IMG_SIZE / 2.0, return_numpy: bool = True, ) -> Union[Section41Output, Dict[str, torch.Tensor]]: """ 使用 DataLoader 输出的 fea_tensor 运行 §4.1。 Parameters ---------- fea_tensor : FeaTensor,含 g_global / u_man 等(列向量打包) img_bgr : 与特征同坐标系的 BGR 图;与 img_tensor 二选一 img_tensor : DataLoader 返回的归一化 RGB (3,H,W) half_norm : u_man 中 s_base 的归一化半宽(默认 50,对应 100×100 图) """ if img_bgr is None: if img_tensor is None: raise ValueError("forward_from_fea 需要 img_bgr 或 img_tensor") img_bgr = img_tensor_to_bgr(img_tensor) fields = fea_fields_from_dataloader(fea_tensor) if fields.u_man.dim() == 3: raise NotImplementedError("batch FeaTensor 请逐样本调用 forward_from_fea") g_global = fields.g_global.detach().cpu().numpy().astype(np.float32) u_man = fields.u_man.detach().cpu().numpy().astype(np.float32) points_px = keypoints_px_from_u_man(u_man, half_norm=half_norm) device = next(self.parameters()).device g_t = torch.from_numpy(g_global).float().to(device) u_t = torch.from_numpy(u_man).float().to(device) h_deep = self.extract_h_deep( img_bgr, points_px, ) out_t = self.forward_tensors(g_t, u_t, h_deep) if not return_numpy: out_t["g_global"] = g_t out_t["u_man"] = u_t out_t["h_deep"] = h_deep out_t["points_px"] = torch.from_numpy(points_px) return out_t def _np(x: torch.Tensor) -> np.ndarray: return x.detach().cpu().numpy() return Section41Output( prep={"rgb": img_bgr}, bundle=None, g_global=g_global, g_vec=_np(out_t["g_vec"]), u_man=u_man, s_man=_np(out_t["s_man"]), h_deep=_np(h_deep), z=_np(out_t["z"]), u_fuse=_np(out_t["u_fuse"]), s_seq=_np(out_t["s_seq"]), config=self.config, ) def save_checkpoint(self, path: Union[str, Path], extra: Optional[dict] = None) -> None: payload = { "config": self.config, "state_dict": self.state_dict(), "extra": extra or {}, } torch.save(payload, path) @classmethod def load_checkpoint( cls, path: Union[str, Path], map_location: Optional[str] = None, strict: bool = True, ) -> "Section41Encoder": payload = torch.load(path, map_location=map_location, weights_only=False) cfg = payload.get("config") or Section41Config() if not isinstance(cfg, Section41Config): cfg = Section41Config(**cfg) if isinstance(cfg, dict) else Section41Config() model = cls(cfg) model.load_state_dict(payload["state_dict"], strict=strict) return model def extract_section41( rgb: PathLike, config: Optional[Section41Config] = None, encoder: Optional[Section41Encoder] = None, device: Optional[Union[str, torch.device]] = None, ) -> Section41Output: enc = encoder or Section41Encoder(config) if device is not None: enc = enc.to(device) enc.eval() return enc.forward(rgb) def demo( image_path: str = "data/images/20938.png", config: Optional[Section41Config] = None, device: str = "cpu", ) -> Section41Output: enc = Section41Encoder(config).to(device) n_trainable = sum(p.numel() for p in enc.trainable_parameters()) n_frozen = sum(p.numel() for p in enc.frozen_parameters()) n_total = sum(p.numel() for p in enc.parameters()) out = enc.forward(image_path) c = out.config print(f"§4.1 demo: {image_path} [deep={enc.deep_branch.backend}]") print(f" 可训练参数: {n_trainable:,} / 冻结: {n_frozen:,} / 总: {n_total:,}") print(f" proj_deep trainable: {enc.proj_deep.weight.requires_grad}") print(f" g_vec {out.g_vec.shape}") print(f" u_man {out.u_man.shape}") print(f" s_man {out.s_man.shape}") print(f" h_deep {out.h_deep.shape}") print(f" z (MHCA) {out.z.shape}") print(f" s_seq {out.s_seq.shape}") if out.bundle is not None: print(f" interp {int(out.bundle.sequence.interp_flag.sum())} / {c.n_theta}") return out def demo_from_dataloader( sample_ratio: float = 0.001, batch_size: int = 2, half_norm: float = DEFAULT_IMG_SIZE / 2.0, config: Optional[Section41Config] = None, device: str = "cpu", ) -> Section41Output: """ 从 exp6_data_loader 取样本,用 fea_tensor(g_global + u_man + 还原采样点)调试 §4.1。 """ from itertools import islice from exp6_data_loader import Config as DataConfig, create_dataloaders cfg = config or Section41Config() enc = Section41Encoder(cfg).to(device) enc.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("[section41] dataloader batch") print(f" fea_tensor.data: {fea_tensor.data.shape}") print(f" img_tensor: {tuple(img_tensor.shape)} label: {labels.tolist()}") # 取 batch 中第一个样本 fea_one = FeaTensor(data=fea_tensor.data[0]) img_one = img_tensor[0] fields = fea_one.fields points_px = keypoints_px_from_u_man(fields.u_man, half_norm=half_norm) print(f" g_global: {tuple(fields.g_global.shape)}") print(f" u_man: {tuple(fields.u_man.shape)}") print(f" s_res: {tuple(fields.s_res.shape)} valid sum: {fields.valid.sum().item():.0f}") print(f" points_px:{points_px.shape} range x=[{points_px[:,0].min():.1f},{points_px[:,0].max():.1f}] " f"y=[{points_px[:,1].min():.1f},{points_px[:,1].max():.1f}]") out = enc.forward_from_fea( fea_one, img_tensor=img_one, half_norm=half_norm, ) n_trainable = sum(p.numel() for p in enc.trainable_parameters()) print("[section41] forward_from_fea") print(f" writer: {writers[0]} image_id: {image_ids[0]}") print(f" trainable params: {n_trainable:,}") print(f" g_vec: {out.g_vec.shape}") print(f" s_man: {out.s_man.shape}") print(f" h_deep: {out.h_deep.shape}") print(f" z: {out.z.shape}") print(f" s_seq: {out.s_seq.shape}") # 批量 smoke test(逐样本 deep + batched forward_tensors 的可训练部分) print("[section41] batch smoke (trainable path only)") fields_b = fea_tensor.fields g_b = fields_b.g_global.float().to(device) u_b = fields_b.u_man.float().to(device) 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(enc.extract_h_deep(bgr, pts)) h_b = torch.stack(h_list, dim=0) out_b = enc.forward_tensors(g_b, u_b, h_b) print(f" batch s_seq: {tuple(out_b['s_seq'].shape)}") return out def main() -> None: import argparse p = argparse.ArgumentParser(description="§4.1 ellipse sequence encoder (PyTorch)") p.add_argument("image", nargs="?", default="data/images/20938.png") p.add_argument("--save", type=str, default=None, help="保存输出 .npz") p.add_argument("--ckpt", type=str, default=None, help="保存可训练权重 .pth") p.add_argument( "--deep", choices=("convnext", "resnet", "patch"), default="convnext", ) p.add_argument( "--resnet-ckpt", type=str, default=None, help=f"ResNet 权重(默认 {default_resnet_checkpoint()})", ) p.add_argument( "--convnext-ckpt", type=str, default=None, help=f"ConvNeXt 权重(默认 {default_convnext_checkpoint()})", ) 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 §4.1", ) p.add_argument("--sample-ratio", type=float, default=0.001) p.add_argument("--batch-size", type=int, default=2) args = p.parse_args() cfg = Section41Config( deep_backend=args.deep, resnet_checkpoint=args.resnet_ckpt or default_resnet_checkpoint(), convnext_checkpoint=args.convnext_ckpt or default_convnext_checkpoint(), ) if args.from_dataloader: demo_from_dataloader( sample_ratio=args.sample_ratio, batch_size=args.batch_size, config=cfg, device=args.device, ) return enc = Section41Encoder(cfg).to(args.device) out = demo(args.image, config=cfg, device=args.device) if args.ckpt: enc.save_checkpoint(args.ckpt) print("checkpoint:", args.ckpt) if args.save: np.savez( args.save, g_vec=out.g_vec, s_man=out.s_man, h_deep=out.h_deep, z=out.z, s_seq=out.s_seq, u_man=out.u_man, s_res=out.bundle.sequence.s_res, ) print("saved:", args.save) if __name__ == "__main__": demo_from_dataloader()