#!/usr/bin/env python3 # -*- coding: utf-8 -*- """ ConvNeXt-Tiny 多层特征图双线性插值深度点特征(§4.1.3)。 采样 stage1–3(stride 4/8/16),拼接 672 维 raw 特征。 """ from __future__ import annotations import argparse import os import random from pathlib import Path from typing import List, Optional, Sequence, Tuple, Union import cv2 import numpy as np import torch import torch.nn as nn import torch.nn.functional as F from torchvision import models from elipse_feature import ( DEFAULT_IMG_SIZE, DEFAULT_N_THETA, EllipseFeatureBundle, extract_ellipse_features, preprocess_rgb_to_skeleton, ) from resnet_point_feature import ( _batch_rgb_bgr_to_tensor, _rgb_bgr_to_tensor, extract_ellipse_keypoints_from_bgr, ) PathLike = Union[str, Path, np.ndarray] CONVNEXT_DEEP_RAW_DIM = 672 # stage1(96) + stage2(192) + stage3(384) CONVNEXT_CHECKPOINT_WINDOWS = "model/model_epoch_60_convnext_image.pth" CONVNEXT_CHECKPOINT_LINUX = "./checkpoints/0619_convnext_rgb_aug/model_epoch60_best_0.517.pth" def default_convnext_checkpoint() -> str: if os.name == "nt": return CONVNEXT_CHECKPOINT_WINDOWS return CONVNEXT_CHECKPOINT_LINUX DEFAULT_CHECKPOINT = Path(default_convnext_checkpoint()) def _default_device() -> torch.device: return torch.device("cuda" if torch.cuda.is_available() else "cpu") def _load_features_only(model: nn.Module, state_dict: dict) -> int: cleaned = {} for k, v in state_dict.items(): name = k[7:] if k.startswith("module.") else k if name.startswith("features.") or name.startswith("backbone.features."): key = name.split("backbone.", 1)[-1] if "backbone." in name else name if key.startswith("features."): cleaned[key] = v if not cleaned: for k, v in state_dict.items(): name = k[7:] if k.startswith("module.") else k if ".features." in name: suffix = name.split(".features.", 1)[1] cleaned[f"features.{suffix}"] = v missing, unexpected = model.load_state_dict(cleaned, strict=False) loaded = len(cleaned) if loaded: print(f"[ok] loaded {loaded} feature tensors (strict=False)") if unexpected: print(f"[info] unexpected keys ignored: {len(unexpected)}") return loaded class ConvNeXtFeatureExtractor(nn.Module): """ConvNeXt-Tiny stage1–3 双线性插值采样(stride 4/8/16)。""" def __init__( self, checkpoint_path: Optional[PathLike] = None, device: Optional[torch.device] = None, ): super().__init__() self.device = device or _default_device() backbone = models.convnext_tiny(weights=None) self.features = backbone.features ckpt = Path(checkpoint_path) if checkpoint_path else DEFAULT_CHECKPOINT if ckpt.is_file(): state = torch.load(ckpt, map_location=self.device, weights_only=False) state_dict = state.get("model_state_dict", state) n = _load_features_only(self, state_dict) if n == 0: print(f"[warn] no features.* in {ckpt}, using random init") else: print(f"[ok] loaded checkpoint: {ckpt}") else: print(f"[warn] checkpoint not found: {ckpt}, using random init") self.strides = {"s1": 4, "s2": 8, "s3": 16} self.out_channels = {"s1": 96, "s2": 192, "s3": 384} self.total_dim = sum(self.out_channels.values()) def _forward_feature_maps(self, images: torch.Tensor) -> dict: x = self.features[0](images) x = self.features[1](x) fmaps = {"s1": x} x = self.features[2](x) x = self.features[3](x) fmaps["s2"] = x x = self.features[4](x) x = self.features[5](x) fmaps["s3"] = x return fmaps def forward(self, images: torch.Tensor, keypoints: torch.Tensor) -> torch.Tensor: """ images : [B, 3, H, W] keypoints : [B, N, 2] 像素坐标 (x, y) returns : [B, N, total_dim] """ feature_maps = self._forward_feature_maps(images) sampled_features_list = [] for name in ("s1", "s2", "s3"): fmap = feature_maps[name] stride = self.strides[name] _, _, fh, fw = fmap.shape scaled_kps = keypoints / stride if fw > 1: norm_x = (scaled_kps[..., 0] / (fw - 1)) * 2 - 1 else: norm_x = torch.zeros_like(scaled_kps[..., 0]) if fh > 1: norm_y = (scaled_kps[..., 1] / (fh - 1)) * 2 - 1 else: norm_y = torch.zeros_like(scaled_kps[..., 1]) norm_x = torch.clamp(norm_x, -1, 1) norm_y = torch.clamp(norm_y, -1, 1) grid = torch.stack([norm_x, norm_y], dim=-1).unsqueeze(2) sampled = F.grid_sample( fmap, grid, mode="bilinear", align_corners=True, padding_mode="border", ) sampled = sampled.squeeze(-1).permute(0, 2, 1) sampled_features_list.append(sampled) return torch.cat(sampled_features_list, dim=-1) @torch.no_grad() def extract_convnext_point_features( rgb_bgr: np.ndarray, points_xy: np.ndarray, extractor: Optional[ConvNeXtFeatureExtractor] = None, checkpoint_path: Optional[PathLike] = None, device: Optional[torch.device] = None, ) -> torch.Tensor: device = device or _default_device() if extractor is None: extractor = ConvNeXtFeatureExtractor(checkpoint_path=checkpoint_path, device=device) extractor.to(device) extractor.eval() 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 extractor(image_tensor, kps).squeeze(0) @torch.no_grad() def extract_convnext_point_features_batch( rgb_bgr_list: Sequence[np.ndarray], points_xy_list: Sequence[np.ndarray], extractor: Optional[ConvNeXtFeatureExtractor] = None, checkpoint_path: Optional[PathLike] = None, device: Optional[torch.device] = None, ) -> torch.Tensor: if len(rgb_bgr_list) != len(points_xy_list): raise ValueError("rgb_bgr_list 与 points_xy_list 长度须一致") if not rgb_bgr_list: raise ValueError("输入列表为空") device = device or _default_device() if extractor is None: extractor = ConvNeXtFeatureExtractor(checkpoint_path=checkpoint_path, device=device) extractor.to(device) extractor.eval() images = _batch_rgb_bgr_to_tensor(rgb_bgr_list, device) _, _, h, w = images.shape kps = torch.stack([ torch.from_numpy(pts.astype(np.float32)) for pts in points_xy_list ], dim=0).to(device) kps[..., 0] = kps[..., 0].clamp(0, w - 1) kps[..., 1] = kps[..., 1].clamp(0, h - 1) return extractor(images, kps) def run_batch_test( image_dir: Union[str, Path] = "data/images", n_samples: int = 3, seed: int = 42, img_size: int = DEFAULT_IMG_SIZE, checkpoint_path: Optional[PathLike] = None, out_dir: Union[str, Path] = "outputs/convnext_point_feat_test", ) -> None: root = Path(image_dir) paths = sorted(root.glob("*.png")) + sorted(root.glob("*.jpg")) if not paths: raise FileNotFoundError(f"目录下无图像: {image_dir}") random.seed(seed) picks = random.sample(paths, min(n_samples, len(paths))) device = _default_device() print(f"[device] {device}") rgb_list: List[np.ndarray] = [] kps_list: List[np.ndarray] = [] for p in picks: orig_bgr = cv2.imread(str(p), cv2.IMREAD_COLOR) if orig_bgr is None: continue result = extract_ellipse_keypoints_from_bgr(orig_bgr, img_size=img_size) if result is None: print(f" [skip] {p.name}") continue rgb, kps, _ = result rgb_list.append(rgb) kps_list.append(kps) print(f" [ok] {p.name}: {kps.shape[0]} pts") if not rgb_list: raise RuntimeError("无可用样本") extractor = ConvNeXtFeatureExtractor(checkpoint_path=checkpoint_path, device=device) extractor.to(device) extractor.eval() features = extract_convnext_point_features_batch( rgb_list, kps_list, extractor=extractor, device=device, ) print(f"[batch] features {tuple(features.shape)} expected dim {CONVNEXT_DEEP_RAW_DIM}") out_root = Path(out_dir) out_root.mkdir(parents=True, exist_ok=True) np.save(out_root / "batch_deep_features.npy", features.cpu().numpy()) if __name__ == "__main__": parser = argparse.ArgumentParser(description="ConvNeXt 多点深度特征测试") parser.add_argument("--image-dir", default="data/images") parser.add_argument("--n", type=int, default=3) parser.add_argument("--seed", type=int, default=42) parser.add_argument("--img-size", type=int, default=DEFAULT_IMG_SIZE) parser.add_argument("--checkpoint", default=None) parser.add_argument("--out-dir", default="outputs/convnext_point_feat_test") args = parser.parse_args() run_batch_test( image_dir=args.image_dir, n_samples=args.n, seed=args.seed, img_size=args.img_size, checkpoint_path=args.checkpoint, out_dir=args.out_dir, )