deepellipse-circleid / code /convnext_point_feature.py
qinxunhui's picture
Add small artifacts (code, outputs, configs, LICENSE)
519ab98 verified
Raw History Blame Contribute Delete
9.52 kB
#!/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,
)