Download code/exp6_data_loader.py from qinxunhui/deepellipse-circleid: direct link, hf CLI and curl.
- Browser
- Download file 43.7 kB
-
https://huggingface.co/qinxunhui/deepellipse-circleid/resolve/main/code/exp6_data_loader.py
- Command line
-
hf download hf://qinxunhui/deepellipse-circleid/code/exp6_data_loader.py
-
curl -L -o exp6_data_loader.py https://huggingface.co/qinxunhui/deepellipse-circleid/resolve/main/code/exp6_data_loader.py
43.7 kB
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
@Time : 2026/3/17 11:04
@File : feature_data_loader.py.py
@Description :
DeepEllipse 数据加载模块
功能:
- 加载 train/val/test CSV 文件
- 椭圆特征打包为 FeaTensor 列向量,反解后供 section41/42 / deep_elipse_model 使用
- 数据增强:二值化/骨架化 + 旋转 + 仿射变换
- 标签编码:writer_id ↔ label_id 自动转换
"""
import copy
import os
import random
from concurrent.futures import ProcessPoolExecutor, as_completed
from itertools import islice
import cv2
import numpy
import numpy as np
import pandas as pd
import torch
from matplotlib import pyplot as plt
from torch.utils.data import Dataset, DataLoader, Sampler
from torchvision import transforms
from PIL import Image
from sklearn.preprocessing import LabelEncoder
import json
from typing import Optional, Union, List, Tuple, Dict, Sequence
import logging
from augmentation import CircleAugmentation
from elipse_feature import (
DEFAULT_N_THETA,
DEFAULT_N_SEGMENTS,
SegmentGeometry,
FeaTensor,
FeaPackLayout,
FeaFields,
unpack_fea_fields,
extract_ellipse_features,
compute_cwt_scalogram,
compute_fft_amplitude_global,
compute_fft_amplitude_segmented,
)
from image_precess import preprocess_writer_image, resize_image
# from augmentation import CircleAugmentation
from section41 import build_u_man
def _collate_fea_tensor(batch_fea: Sequence[FeaTensor]) -> FeaTensor:
return FeaTensor.stack(batch_fea)
def circle_id_collate_fn(batch):
"""DataLoader collate:FeaTensor 保持紧凑 batch;input_all 可含可选 cwt_scalo。"""
input_alls, labels, pen_ids, writers, image_ids = zip(*batch)
first = input_alls[0]
if len(first) == 2:
ske_list = [ia[0] for ia in input_alls]
img_list = [ia[1] for ia in input_alls]
return (
(None, torch.stack(ske_list, dim=0), torch.stack(img_list, dim=0), None),
torch.tensor(labels),
torch.tensor(pen_ids),
list(writers),
list(image_ids),
)
fea_list = [ia[0] for ia in input_alls]
ske_list = [ia[1] for ia in input_alls]
img_list = [ia[2] for ia in input_alls]
cwt_list = [ia[3] if len(ia) > 3 else None for ia in input_alls]
cwt_batch = None
if cwt_list and all(c is not None for c in cwt_list):
cwt_batch = torch.stack(cwt_list, dim=0)
return (
(
_collate_fea_tensor(fea_list),
torch.stack(ske_list, dim=0),
torch.stack(img_list, dim=0),
cwt_batch,
),
torch.tensor(labels),
torch.tensor(pen_ids),
list(writers),
list(image_ids),
)
class Config:
"""全局配置参数"""
_IS_WINDOWS = os.name == "nt"
# 路径配置(按操作系统自动切换)
if _IS_WINDOWS:
IMAGE_ROOT = r"/path/to/circle_id\data"
DATA_SPLIT_ROOT = r"/path/to/circle_id\data"
TRAIN_CSV = os.path.join(DATA_SPLIT_ROOT, "train.csv")
VAL_CSV = os.path.join(DATA_SPLIT_ROOT, "train.csv")
TEST_CSV = os.path.join(DATA_SPLIT_ROOT, "train.csv")
FEA_CACHE_ROOT = os.path.join(r"/path/to/circle_id", "fea_cache")
else:
IMAGE_ROOT = "/path/to/circle_id/images/"
DATA_SPLIT_ROOT = "/path/to/circle_id/data_split/"
TRAIN_CSV = os.path.join(DATA_SPLIT_ROOT, "train_train3.csv") #"train_train3_our_train.csv" #train_train3.csv
VAL_CSV = os.path.join(DATA_SPLIT_ROOT, "train_val3.csv")
TEST_CSV = os.path.join(DATA_SPLIT_ROOT, "test_with_pen.csv")
FEA_CACHE_ROOT = "/path/to/circle_id_fea_cache" #"/path/to/circle_id/fea_cache/"
# 图像预处理配置
TARGET_SIZE = 100 # 图像目标尺寸
# 数据增强配置
USE_AUGMENT = True
ROTATION_DEGREES = 10 # 随机旋转角度范围 (±10°)
AFFINE_TRANSLATE = 0.03 # 仿射平移比例 (±3%)
AFFINE_SCALE = (0.97, 1.03) # 仿射缩放比例
# 归一化配置 (适配骨架图 [0,255]→[0,1])
NORM_MEAN = [0.5] # 单通道均值
NORM_STD = [0.5] # 单通道标准差
# 日志配置
VERBOSE_DEBUG = False
DEBUG_PREPROCESS = False
DEBUG_SAMPLE_RATIO = 0.01
DEBUG_MIN_SAMPLES = 4
# 椭圆特征配置
N_THETA = DEFAULT_N_THETA # 360
N_SEGMENTS = DEFAULT_N_SEGMENTS # 10
S_L = N_THETA // N_SEGMENTS # 36
FREQ_DIM = 64 # §4.2 输出频率维度
DEEP_FEATURE_DIM = 448 # ResNet深度特征维度
# fea_tensor 预计算缓存(FEA_CACHE_ROOT 见上方路径配置)
PRECOMPUTE_CWT = False
FEA_AUG_VARIANTS = 30 # 每张图预计算增强变体数
USE_PRECOMPUTED_FEA = True
PRECOMPUTE_NUM_WORKERS = 0 # 0 = 自动使用全部 CPU 核心
_NORM_MEAN_SKEL = np.array(0.1, dtype=np.float32).reshape(1, 1, -1)
_NORM_STD_SKEL = np.array(0.3, dtype=np.float32).reshape(1, 1, -1)
_NORM_MEAN_IMG = np.array([0.485, 0.456, 0.406], dtype=np.float32).reshape(1, 1, -1)
_NORM_STD_IMG = np.array([0.229, 0.224, 0.225], dtype=np.float32).reshape(1, 1, -1)
def _fea_cache_path(cache_root: str, image_id: str) -> str:
safe_id = str(image_id).replace(os.sep, "_").replace("/", "_")
return os.path.join(cache_root, f"{safe_id}.pt")
def check_fea_cache_coverage(
csv_path: str,
cache_root: str,
) -> Tuple[int, int, List[str]]:
"""检查 CSV 中样本的 fea 缓存覆盖率。返回 (found, total, missing_sample)。"""
if not os.path.exists(csv_path):
return 0, 0, []
df = pd.read_csv(csv_path)
total = len(df)
found = 0
missing: List[str] = []
for image_id in df["image_id"].astype(str):
if os.path.exists(_fea_cache_path(cache_root, image_id)):
found += 1
elif len(missing) < 5:
missing.append(image_id)
return found, total, missing
def fea_cache_has_cwt(cache_root: str, split: str = "train") -> bool:
"""抽样检查 fea 缓存是否含 cwt_scalo(需 precompute 时加 --precompute-cwt)。"""
split_dir = os.path.join(cache_root, split)
if not os.path.isdir(split_dir):
return False
for name in os.listdir(split_dir):
if not name.endswith(".pt"):
continue
try:
cache = torch.load(
os.path.join(split_dir, name), map_location="cpu", weights_only=False
)
return "cwt_scalo" in cache
except Exception:
continue
return False
def compute_ske_img_tensors(image_bgr: np.ndarray) -> Tuple[torch.Tensor, torch.Tensor]:
"""从 BGR 图像快速计算 ske_tensor / img_tensor(不做椭圆特征提取)。"""
ske = preprocess_writer_image(image_bgr, use_binarization=True, use_skeleton=True)
ske_rgb = cv2.cvtColor(ske, cv2.COLOR_GRAY2RGB) / 255.0
ske_norm = (ske_rgb - _NORM_MEAN_SKEL) / _NORM_STD_SKEL
ske_tensor = torch.FloatTensor(ske_norm).permute(2, 0, 1)
image = cv2.cvtColor(image_bgr, cv2.COLOR_BGR2RGB)
img_norm = (image / 255.0 - _NORM_MEAN_IMG) / _NORM_STD_IMG
img_tensor = torch.FloatTensor(img_norm).permute(2, 0, 1)
return ske_tensor, img_tensor
def compute_fea_tensor(image_bgr: np.ndarray) -> Optional[FeaTensor]:
"""从 BGR 图像提取椭圆特征并打包为 FeaTensor(耗时操作)。"""
ske = preprocess_writer_image(image_bgr, use_binarization=True, use_skeleton=True)
bundle = extract_ellipse_features(ske, half_norm=50)
if bundle is None:
return None
valid_count = np.sum(bundle.sequence.interp_flag == 0)
if valid_count < 120:
return None
u_man = build_u_man(bundle.sequence)
amp_global = compute_fft_amplitude_global(u_man[:, :4])
geom = SegmentGeometry(
n_theta=Config.N_THETA,
n_segments=Config.N_SEGMENTS,
s_l=Config.S_L,
half=Config.S_L // 2,
l_win=2 * Config.S_L,
)
amp_segment = compute_fft_amplitude_segmented(u_man[:, :4], geom)
return FeaTensor.from_tensors(
g_global=torch.from_numpy(bundle.global_feat.as_vector()),
u_man=torch.from_numpy(u_man),
amp_global=torch.from_numpy(amp_global.reshape(-1).astype(np.float32)),
amp_segment=torch.from_numpy(amp_segment.reshape(-1).astype(np.float32)),
)
def prepare_model_input_dataloader(image_bgr, only_image=False):
"""
提取椭圆特征并打包为 FeaTensor 列向量。
反解用法(section41 / section42 / deep_elipse_model):
fea = fea_tensor.fields # FeaFields
fea.g_global # (6,)
fea.u_man # (360, 8)
fea.s_res / fea.valid # 由 u_man 派生
fea.amp_global / fea.amp_segment # FFT 幅值,§4.2 输入
"""
ske_tensor, img_tensor = compute_ske_img_tensors(image_bgr)
if only_image is True:
return ske_tensor, img_tensor
fea_tensor = compute_fea_tensor(image_bgr)
if fea_tensor is None:
return None
return fea_tensor, ske_tensor, img_tensor
_worker_augmenter: Optional[CircleAugmentation] = None
def _resolve_num_workers(num_workers: int) -> int:
if num_workers <= 0:
return os.cpu_count() or 4
return num_workers
def _init_precompute_worker() -> None:
"""每个子进程限制 BLAS/OpenCV 线程,避免多进程过度抢占 CPU。"""
os.environ.setdefault("OMP_NUM_THREADS", "1")
os.environ.setdefault("MKL_NUM_THREADS", "1")
os.environ.setdefault("OPENBLAS_NUM_THREADS", "1")
torch.set_num_threads(1)
try:
cv2.setNumThreads(1)
except Exception:
pass
def _get_worker_augmenter(use_augment: bool) -> Optional[CircleAugmentation]:
global _worker_augmenter
if not use_augment:
return None
if _worker_augmenter is None:
_worker_augmenter = CircleAugmentation()
return _worker_augmenter
def _precompute_fea_one(task: Dict) -> str:
"""
处理单张图的 fea 预计算(供多进程 worker 调用)。
Returns:
"ok" | "partial" | "fail"
"""
image_id = task["image_id"]
image_path = task["image_path"]
cache_path = task["cache_path"]
target_size = task["target_size"]
num_variants = task["num_variants"]
use_augment = task["use_augment"]
try:
image = cv2.imread(image_path)
if image is None:
return "fail"
augmenter = _get_worker_augmenter(use_augment)
fea_rows: List[torch.Tensor] = []
image_rows: List[np.ndarray] = []
cwt_rows: List[torch.Tensor] = []
precompute_cwt = task.get("precompute_cwt", False)
for _ in range(num_variants):
aug_image = resize_image(image.copy())
if augmenter is not None:
aug_image = augmenter.aug_image(aug_image)
aug_image = resize_image(aug_image, target_size=target_size)
fea_tensor = compute_fea_tensor(aug_image)
if fea_tensor is None:
continue
fea_rows.append(fea_tensor.data.reshape(-1).float())
image_rows.append(aug_image.copy())
if precompute_cwt:
u = fea_tensor.fields.u_man.numpy()
cwt_rows.append(
torch.from_numpy(compute_cwt_scalogram(u[:, :4]).reshape(-1))
)
if not fea_rows:
return "fail"
os.makedirs(os.path.dirname(cache_path) or ".", exist_ok=True)
payload = {
"image_id": image_id,
"fea_data": torch.stack(fea_rows, dim=0),
"images": torch.from_numpy(np.stack(image_rows, axis=0)),
}
if precompute_cwt and cwt_rows:
payload["cwt_scalo"] = torch.stack(cwt_rows, dim=0)
torch.save(payload, cache_path)
if len(fea_rows) < num_variants:
return "partial"
return "ok"
except Exception:
return "fail"
def precompute_fea_cache(
csv_path: str,
image_root: str = Config.IMAGE_ROOT,
cache_root: str = Config.FEA_CACHE_ROOT,
target_size: int = Config.TARGET_SIZE,
num_variants: int = Config.FEA_AUG_VARIANTS,
use_augment: bool = True,
overwrite: bool = False,
num_workers: int = Config.PRECOMPUTE_NUM_WORKERS,
precompute_cwt: bool = False,
verbose: bool = True,
) -> Dict[str, int]:
"""
离线预计算 fea_tensor 并缓存。
每张图随机增强 num_variants 次,每次计算 fea_tensor 并保存对应增强后的图像
(供加载时实时计算 ske/img,保证与 fea 来自同一增强结果)。
缓存文件:{cache_root}/{image_id}.pt
fea_data: (K, L) float32
images: (K, H, W, 3) uint8 BGR
cwt_scalo: (K, 4*40*360) float32 — 可选,precompute_cwt=True
"""
os.makedirs(cache_root, exist_ok=True)
df = pd.read_csv(csv_path)
stats = {"ok": 0, "skip_exists": 0, "fail": 0, "partial": 0}
tasks: List[Dict] = []
for _, row in df.iterrows():
image_id = str(row["image_id"])
cache_path = _fea_cache_path(cache_root, image_id)
if not overwrite and os.path.exists(cache_path):
stats["skip_exists"] += 1
continue
tasks.append(
{
"image_id": image_id,
"image_path": os.path.join(image_root, row["image_path"]),
"cache_path": cache_path,
"target_size": target_size,
"num_variants": num_variants,
"use_augment": use_augment,
"precompute_cwt": precompute_cwt,
}
)
n_workers = _resolve_num_workers(num_workers)
if verbose:
print(f"预计算任务: {len(tasks)} 张待处理, {stats['skip_exists']} 张已跳过, workers={n_workers}")
if not tasks:
if verbose:
print(f"✅ fea 预计算完成 @ {cache_root}: {stats}")
return stats
def _consume_results(result_iter):
for status in result_iter:
stats[status] = stats.get(status, 0) + 1
if n_workers <= 1:
iterator = tasks
if verbose:
try:
from tqdm import tqdm
iterator = tqdm(tasks, desc="precompute fea")
except ImportError:
pass
_consume_results(_precompute_fea_one(task) for task in iterator)
else:
with ProcessPoolExecutor(
max_workers=n_workers,
initializer=_init_precompute_worker,
) as executor:
futures = [executor.submit(_precompute_fea_one, task) for task in tasks]
result_iter = (f.result() for f in as_completed(futures))
if verbose:
try:
from tqdm import tqdm
result_iter = tqdm(result_iter, total=len(futures), desc="precompute fea")
except ImportError:
pass
_consume_results(result_iter)
if verbose:
print(f"✅ fea 预计算完成 @ {cache_root}: {stats}")
return stats
# ================= 核心数据集类 =================
class CircleIDDataset(Dataset):
"""
CircleID 数据集类
支持:
- 加载 CSV 元数据 (image_id, image_path, writer_id, pen_id)
- 预计算特征加载 或 实时特征计算
- 数据增强 (仅实时计算模式)
- 标签编码 (writer_id → label_id)
"""
def __init__(
self,
csv_path: str,
image_root: str = Config.IMAGE_ROOT,
target_size: int = Config.TARGET_SIZE,
transform: Optional[transforms.Compose] = None,
label_encoder: Optional[LabelEncoder] = None,
fit_encoder: bool = False,
sample_ratio: float = 1.0,
min_samples: int = 0,
is_test: bool = False,
enable_augment: bool = True,
use_precomputed_fea: bool = Config.USE_PRECOMPUTED_FEA,
fea_cache_root: str = Config.FEA_CACHE_ROOT,
img_only: bool = False,
verbose: bool = Config.VERBOSE_DEBUG
):
"""
初始化数据集
Args:
csv_path: CSV 文件路径
image_root: 图像根目录
transform: torchvision 变换 (用于图像模式)
label_encoder: LabelEncoder 实例
fit_encoder: 是否拟合 encoder (仅训练集)
sample_ratio: 数据采样比例 (用于调试)
min_samples: 最小采样数量
is_test: 是否为测试集 (无标签)
use_binarization: 是否二值化
use_skeleton: 是否骨架化
enable_augment: 是否随机增强(训练 True,验证/测试 False)
use_precomputed_fea: 是否从缓存加载预计算的 fea_tensor
fea_cache_root: fea_tensor 缓存目录
img_only: True 时仅返回 (ske, img),跳过椭圆 fea 提取
verbose: 是否输出调试信息
"""
self.csv_path = csv_path
self.image_root = image_root
self.target_size = target_size
self.transform = transform
self.is_test = is_test
self.img_only = img_only
self.use_precomputed_fea = use_precomputed_fea
self.fea_cache_root = fea_cache_root
self.enable_augment = enable_augment
# 在线模式:训练集实时随机增强;验证/测试不增强
self.use_online_augment = enable_augment and not use_precomputed_fea
# 预计算模式:训练集随机选取变体;验证/测试固定用第 0 个(无增强)
self.use_random_fea_variant = enable_augment and use_precomputed_fea
self.verbose = verbose
self.error_idx = set()
# 1. 加载 CSV
self.df = pd.read_csv(csv_path)
# 2. 数据采样 (调试用)
if sample_ratio < 1.0 or min_samples > 0:
original_len = len(self.df)
sample_size = max(int(original_len * sample_ratio), min_samples)
sample_size = min(sample_size, original_len)
if sample_size < original_len and verbose:
print(f"📊 数据采样:{original_len} → {sample_size} 张图片")
self.df = self.df.sample(n=sample_size, random_state=42).reset_index(drop=True)
# debug
# self.df = self.df.sample(n=100, random_state=42).reset_index(drop=True)
# 3. 标签编码
self.label_encoder = label_encoder
self._setup_labels(fit_encoder)
# 5. 数据增强器(仅训练集在线模式)
if self.use_online_augment:
self.augmenter = CircleAugmentation()
else:
self.augmenter = None
# 6. 打印统计信息
if verbose:
aug_tag = "随机增强" if self.enable_augment else "无增强"
print(f"✅ 数据集加载完成:{len(self.df)} 张图片 ({aug_tag})" +
(f", {self.df['writer_id'].nunique()} 个 Writer" if not is_test else " (测试集)"))
def _setup_labels(self, fit_encoder: bool):
"""设置标签编码"""
if self.is_test:
# 测试集:无标签
self.num_classes = 0
self.df['label'] = -1
# self.df['pen_id'] = -1
else:
if fit_encoder:
# 训练集:拟合 encoder
self.label_encoder = LabelEncoder()
self.df['label'] = self.label_encoder.fit_transform(self.df['writer_id'])
self.num_classes = len(self.label_encoder.classes_)
if self.verbose:
print(f"🎯 训练集 - Writer 类别数:{self.num_classes}")
elif self.label_encoder is not None:
# 验证集:使用已有 encoder
self.df['label'] = self.df['writer_id'].apply(
lambda x: self.label_encoder.transform([x])[0]
if x in self.label_encoder.classes_ else -1
)
self.num_classes = len(self.label_encoder.classes_)
else:
raise ValueError("验证集需要提供 label_encoder")
def __len__(self) -> int:
return len(self.df)
def _load_from_fea_cache(self, image_id: str) -> Optional[Tuple[FeaTensor, torch.Tensor, torch.Tensor]]:
"""从缓存加载 fea;训练集随机选变体,验证/测试固定用第 0 个。"""
cache_path = _fea_cache_path(self.fea_cache_root, image_id)
if not os.path.exists(cache_path):
return None
cache = torch.load(cache_path, map_location="cpu", weights_only=False)
fea_data = cache["fea_data"]
images = cache["images"]
n_variants = fea_data.shape[0]
if n_variants == 0:
return None
if self.use_random_fea_variant:
variant_idx = random.randint(0, n_variants - 1)
else:
variant_idx = 0
fea_tensor = FeaTensor(data=fea_data[variant_idx])
image_bgr = images[variant_idx].numpy()
if image_bgr.dtype != np.uint8:
image_bgr = image_bgr.astype(np.uint8)
ske_tensor, img_tensor = compute_ske_img_tensors(image_bgr)
cwt_tensor = None
if "cwt_scalo" in cache:
flat = cache["cwt_scalo"][variant_idx]
cwt_tensor = flat.view(4, 40, Config.N_THETA).float()
if cwt_tensor is not None:
return fea_tensor, ske_tensor, img_tensor, cwt_tensor
return fea_tensor, ske_tensor, img_tensor
def __getitem__(self, idx: int):
"""
获取单个样本
Returns:
feature: (40, 120) 特征张量 [C, L]
label: writer_id 编码 (-1 表示测试集)
pen_id: 笔的标识
writer_id: 原始 writer_id 字符串
"""
if idx in self.error_idx:
new_index = np.random.randint(0, len(self.df))
return self.__getitem__(new_index)
row = self.df.iloc[idx]
pen_id = row['pen_id']
# row['image_path'] = "images/00295.png" # debug
image_path = os.path.join(self.image_root, row['image_path'])
writer_id = str(row['writer_id']) if 'writer_id' in row else ''
label = row['label'] if not self.is_test else -1
image_id = str(row['image_id'])
if self.use_precomputed_fea:
cached = self._load_from_fea_cache(image_id)
if cached is not None:
if self.img_only:
_, ske_tensor, img_tensor = cached[:3]
return (ske_tensor, img_tensor), label, pen_id, writer_id, image_id
return cached, label, pen_id, writer_id, image_id
image = cv2.imread(image_path) # BGR
if image is None:
self.error_idx.add(idx)
new_index = np.random.randint(0, len(self.df))
return self.__getitem__(new_index)
try:
if self.augmenter is not None:
image = resize_image(image)
image = self.augmenter.aug_image(image)
# image = self.augmenter.aug_image_offical(image)
image = resize_image(image)
# 与 §4.1 / ResNet 输入一致:保持长宽比缩放并 pad 到 target_size×target_size
image = resize_image(image, target_size=self.target_size)
input_all = prepare_model_input_dataloader(
image, only_image=self.img_only,
)
if input_all is None and not self.img_only:
self.error_idx.add(idx)
new_index = np.random.randint(0, len(self.df))
return self.__getitem__(new_index)
else:
return input_all, label, pen_id, writer_id, image_id
except Exception as e:
print(f"\n❌ [ERROR] idx={idx}, image_id={image_id}, writer_id={writer_id}")
print(f" Exception: {type(e).__name__}: {e}")
import traceback
traceback.print_exc()
self.error_idx.add(idx)
new_index = np.random.randint(0, len(self.df))
# cv2.imwrite(f"/path/to/circle_id/circle_output/polar_order/temp/fail_elipse_{image_id}.png", image)
return self.__getitem__(new_index)
def get_label_encoder(self) -> Optional[LabelEncoder]:
"""获取标签编码器 (用于测试集解码)"""
return self.label_encoder
def get_writer_by_label(self, label: int) -> Optional[str]:
"""根据 label_id 获取 writer_id"""
if self.label_encoder is not None and 0 <= label < len(self.label_encoder.classes_):
return self.label_encoder.classes_[label]
return None
def _build_writer_label_pools(
labels: Sequence[int],
*,
exclude_label: int = -1,
) -> Tuple[Dict[int, List[int]], List[int]]:
label_to_indices: Dict[int, List[int]] = {}
for idx, lab in enumerate(labels):
li = int(lab)
if li == exclude_label:
continue
label_to_indices.setdefault(li, []).append(idx)
return label_to_indices, sorted(label_to_indices.keys())
def parse_pk_mix(pk_mix: str) -> List[Tuple[int, int]]:
"""Parse '4,128;8,64' into [(4,128), (8,64)]; all configs must share P*K."""
configs: List[Tuple[int, int]] = []
for part in pk_mix.split(";"):
part = part.strip()
if not part:
continue
chunks = [c.strip() for c in part.split(",") if c.strip()]
if len(chunks) != 2:
raise ValueError(
f"Invalid pk_mix segment {part!r}; expected 'P,K' (e.g. 4,128)"
)
p, k = int(chunks[0]), int(chunks[1])
if p <= 0 or k <= 0:
raise ValueError(f"PK values must be positive, got P={p}, K={k}")
configs.append((p, k))
if not configs:
raise ValueError(f"pk_mix is empty: {pk_mix!r}")
batch_sizes = {p * k for p, k in configs}
if len(batch_sizes) != 1:
raise ValueError(f"All PK configs must have the same P*K, got {configs}")
return configs
def _sample_pk_batch(
label_to_indices: Dict[int, List[int]],
writer_labels: Sequence[int],
*,
p: int,
k: int,
) -> List[int]:
chosen = random.sample(list(writer_labels), p)
batch: List[int] = []
for lab in chosen:
pool = label_to_indices[lab]
if len(pool) >= k:
batch.extend(random.sample(pool, k))
else:
batch.extend(random.choices(pool, k=k))
return batch
class WriterPKBatchSampler(Sampler):
"""Yields batch index lists: P writers × K images; excludes ignore_label."""
def __init__(
self,
labels: Sequence[int],
*,
p: int,
k: int,
exclude_label: int = -1,
num_batches: Optional[int] = None,
):
if p <= 0 or k <= 0:
raise ValueError(f"p and k must be positive, got p={p}, k={k}")
self.p = int(p)
self.k = int(k)
self.batch_size = self.p * self.k
self.label_to_indices, self.labels = _build_writer_label_pools(
labels, exclude_label=exclude_label
)
if len(self.labels) < self.p:
raise ValueError(
f"WriterPKBatchSampler: need at least {self.p} writers, "
f"got {len(self.labels)}"
)
total = sum(len(v) for v in self.label_to_indices.values())
self._num_batches = int(
num_batches if num_batches is not None else max(1, total // self.batch_size)
)
self.last_pk: Optional[Tuple[int, int]] = (self.p, self.k)
def __len__(self) -> int:
return self._num_batches
def __iter__(self):
for _ in range(self._num_batches):
self.last_pk = (self.p, self.k)
yield _sample_pk_batch(
self.label_to_indices, self.labels, p=self.p, k=self.k
)
class WriterPKMixBatchSampler(Sampler):
"""Each batch randomly picks one (P,K) from pk_configs; all must share P*K."""
def __init__(
self,
labels: Sequence[int],
*,
pk_configs: Sequence[Tuple[int, int]],
exclude_label: int = -1,
num_batches: Optional[int] = None,
):
self.pk_configs = [(int(p), int(k)) for p, k in pk_configs]
batch_sizes = {p * k for p, k in self.pk_configs}
if len(batch_sizes) != 1:
raise ValueError(
f"WriterPKMixBatchSampler: all configs need same P*K, got {self.pk_configs}"
)
self.batch_size = batch_sizes.pop()
self.max_p = max(p for p, _ in self.pk_configs)
self.label_to_indices, self.labels = _build_writer_label_pools(
labels, exclude_label=exclude_label
)
if len(self.labels) < self.max_p:
raise ValueError(
f"WriterPKMixBatchSampler: need at least {self.max_p} writers, "
f"got {len(self.labels)}"
)
total = sum(len(v) for v in self.label_to_indices.values())
self._num_batches = int(
num_batches if num_batches is not None else max(1, total // self.batch_size)
)
self.last_pk: Optional[Tuple[int, int]] = None
def __len__(self) -> int:
return self._num_batches
def __iter__(self):
for _ in range(self._num_batches):
p, k = random.choice(self.pk_configs)
self.last_pk = (p, k)
yield _sample_pk_batch(
self.label_to_indices, self.labels, p=p, k=k
)
def _dataloader_extra_kwargs(
num_workers: int,
prefetch_factor: Optional[int],
) -> dict:
extra: dict = {}
if num_workers > 0 and prefetch_factor is not None:
extra["prefetch_factor"] = int(prefetch_factor)
extra["persistent_workers"] = False
return extra
# ================= 数据加载器工厂 =================
def create_dataloaders(
batch_size: int = 32,
num_workers: int = 32,
pin_memory: bool = True,
sample_ratio: float = 1.0,
min_samples: int = 0,
use_precomputed: bool = True,
fea_cache_root: Optional[str] = None,
img_only: bool = False,
verbose: bool = True,
pk_p: Optional[int] = None,
pk_k: Optional[int] = None,
pk_mix: Optional[str] = None,
pk_exclude_label: int = -1,
eval_batch_size: Optional[int] = None,
prefetch_factor: Optional[int] = None,
) -> Tuple[DataLoader, DataLoader, Optional[DataLoader], LabelEncoder]:
"""
创建训练/验证/测试数据加载器。
增强策略:训练集随机增强;验证集、测试集不增强。
Returns:
train_loader, val_loader, test_loader, label_encoder
"""
cache_root = fea_cache_root or Config.FEA_CACHE_ROOT
if use_precomputed and verbose:
split_specs = [
("train", Config.TRAIN_CSV),
("val", Config.VAL_CSV),
]
if os.path.exists(Config.TEST_CSV):
split_specs.append(("test", Config.TEST_CSV))
precompute_hint_printed = False
for split_name, csv_path in split_specs:
found, total, missing = check_fea_cache_coverage(
csv_path, os.path.join(cache_root, split_name)
)
pct = 100.0 * found / max(total, 1)
print(f"fea 缓存 [{split_name}]: {found}/{total} ({pct:.1f}%)")
if found < total:
print(f" 警告: {split_name} 缺 {total - found} 个缓存", end="")
if missing:
print(f",示例 id: {missing[:3]}")
else:
print()
if not precompute_hint_printed:
print(
" 补全命令: python exp6_data_loader.py "
"--precompute-fea --split val_test --precompute-cwt"
)
precompute_hint_printed = True
# 1. 创建训练集 (拟合 label encoder,随机增强)
train_dataset = CircleIDDataset(
csv_path=Config.TRAIN_CSV,
image_root=Config.IMAGE_ROOT,
fit_encoder=True,
sample_ratio=sample_ratio,
min_samples=min_samples,
is_test=False,
enable_augment=True,
use_precomputed_fea=use_precomputed,
fea_cache_root=os.path.join(cache_root, "train"),
img_only=img_only,
verbose=verbose
)
label_encoder = train_dataset.label_encoder
# 2. 创建验证集 (使用已有 encoder,不增强)
val_dataset = CircleIDDataset(
csv_path=Config.VAL_CSV,
image_root=Config.IMAGE_ROOT,
label_encoder=label_encoder,
fit_encoder=False,
is_test=False,
enable_augment=False,
use_precomputed_fea=use_precomputed,
fea_cache_root=os.path.join(cache_root, "val"),
img_only=img_only,
verbose=verbose
)
eval_bs = int(eval_batch_size) if eval_batch_size is not None else int(batch_size)
worker_kw = _dataloader_extra_kwargs(num_workers, prefetch_factor)
# 3. 创建测试集 (可选,不增强)
if os.path.exists(Config.TEST_CSV):
test_dataset = CircleIDDataset(
csv_path=Config.TEST_CSV,
image_root=Config.IMAGE_ROOT,
label_encoder=label_encoder,
is_test=True,
enable_augment=False,
use_precomputed_fea=use_precomputed,
fea_cache_root=os.path.join(cache_root, "test"),
img_only=img_only,
verbose=verbose
)
test_loader = DataLoader(
test_dataset,
batch_size=eval_bs,
shuffle=False,
num_workers=num_workers,
pin_memory=pin_memory,
collate_fn=circle_id_collate_fn,
**worker_kw,
)
else:
test_loader = None
if verbose:
print(f"⚠️ 测试集文件不存在: {Config.TEST_CSV}")
# 4. 创建 DataLoader
train_batch_sampler = None
if pk_mix:
pk_configs = parse_pk_mix(pk_mix)
train_labels = train_dataset.df["label"].astype(int).tolist()
train_batch_sampler = WriterPKMixBatchSampler(
train_labels,
pk_configs=pk_configs,
exclude_label=pk_exclude_label,
)
if verbose:
cfg_str = ", ".join(f"P{p}K{k}" for p, k in pk_configs)
print(
f" Train PK mix batch_sampler: [{cfg_str}], "
f"batch={train_batch_sampler.batch_size}, "
f"writers={len(train_batch_sampler.labels)}, "
f"batches/epoch={len(train_batch_sampler)}"
)
elif pk_p is not None and pk_k is not None:
train_labels = train_dataset.df["label"].astype(int).tolist()
train_batch_sampler = WriterPKBatchSampler(
train_labels, p=int(pk_p), k=int(pk_k), exclude_label=pk_exclude_label
)
if verbose:
print(
f" Train PK batch_sampler: P={pk_p}, K={pk_k}, "
f"batch={int(pk_p) * int(pk_k)}, writers={len(train_batch_sampler.labels)}, "
f"batches/epoch={len(train_batch_sampler)}"
)
if train_batch_sampler is not None:
train_loader = DataLoader(
train_dataset,
batch_sampler=train_batch_sampler,
num_workers=num_workers,
pin_memory=pin_memory,
collate_fn=circle_id_collate_fn,
**worker_kw,
)
else:
train_loader = DataLoader(
train_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=num_workers,
pin_memory=pin_memory,
drop_last=True,
collate_fn=circle_id_collate_fn,
**worker_kw,
)
val_loader = DataLoader(
val_dataset,
batch_size=eval_bs,
shuffle=False,
num_workers=num_workers,
pin_memory=pin_memory,
collate_fn=circle_id_collate_fn,
**worker_kw,
)
if verbose:
print(f"📦 DataLoader 创建完成:")
if use_precomputed:
print(f" fea 缓存: {cache_root}/{{train,val,test}}")
print(f" 训练: 随机选取增强变体 | 验证/测试: 固定原图变体")
print(f" Train: {len(train_dataset)} samples, {len(train_loader)} batches (增强)")
print(f" Val: {len(val_dataset)} samples, {len(val_loader)} batches (无增强)")
if test_loader:
print(f" Test: {len(test_dataset)} samples, {len(test_loader)} batches (无增强)")
return train_loader, val_loader, test_loader, label_encoder
def merge_csv():
import pandas as pd
# 1. 定义文件路径
# 请替换为你实际的文件路径
csv_path_1 = "/path/to/circle_id/data_split/our_train.csv"
csv_path_2 = "/path/to/circle_id/data_split/train_train3.csv"
output_path = "/path/to/circle_id/data_split/train_train3_our_train.csv"
# 2. 读取两个 CSV 文件
df1 = pd.read_csv(csv_path_1)
df2 = pd.read_csv(csv_path_2)
# 3. 合并数据 (上下拼接)
# ignore_index=True 会重新生成连续的索引 (0, 1, 2...),避免索引重复
merged_df = pd.concat([df1, df2], ignore_index=True)
# 4. 保存为新的 CSV 文件
# index=False 表示保存时不包含行索引列
merged_df.to_csv(output_path, index=False)
print(f"合并完成!新文件已保存至: {output_path}")
print(f"原始数据行数: {len(df1)} + {len(df2)} = {len(merged_df)}")
def build_test_loader(
csv_path: str,
*,
class_names: Sequence[str],
fea_cache_root: Optional[str] = None,
image_root: Optional[str] = None,
batch_size: int = 128,
num_workers: int = 8,
pin_memory: bool = True,
) -> Tuple[DataLoader, LabelEncoder]:
"""Build a *test-only* loader (no train/val datasets, no augmentation).
Released inference needs the test split alone, so unlike
:func:`create_dataloaders` this never materialises the training or
validation datasets -- and therefore never touches the (large) training
feature cache. The label encoder is taken from the released class-name
list instead of being fitted on the training CSV.
Returns (test_loader, label_encoder).
"""
encoder = LabelEncoder()
encoder.classes_ = pd.Index([str(c) for c in class_names])
cache_root = fea_cache_root or Config.FEA_CACHE_ROOT
dataset = CircleIDDataset(
csv_path=csv_path,
image_root=image_root or Config.IMAGE_ROOT,
label_encoder=encoder,
fit_encoder=False,
is_test=True,
enable_augment=False,
use_precomputed_fea=True,
fea_cache_root=os.path.join(cache_root, "test"),
verbose=False,
)
loader = DataLoader(
dataset,
batch_size=batch_size,
shuffle=False,
num_workers=num_workers,
pin_memory=pin_memory,
collate_fn=circle_id_collate_fn,
)
return loader, encoder
# ================= 测试代码 =================
if __name__ == "__main__":
import argparse
parser = argparse.ArgumentParser(description="CircleID data loader")
parser.add_argument(
"--precompute-fea",
default=True,
help="离线预计算 fea_tensor 缓存",
)
parser.add_argument(
"--split",
choices=["train", "val", "test", "val_test", "all"],
default="all",
help="预计算划分:val_test=仅 val+test(train 已有缓存时用);all=train+val+test",
)
parser.add_argument(
"--num-variants",
type=int,
default=Config.FEA_AUG_VARIANTS,
help="每张图随机增强并预计算的变体数",
)
parser.add_argument(
"--cache-root",
type=str,
default=Config.FEA_CACHE_ROOT,
help="fea_tensor 缓存目录",
)
parser.add_argument(
"--precompute-cwt",
action="store_true",
help="fea 预计算时同时缓存 cwt_scalo (4×40×360)",
)
parser.add_argument(
"--train-augment",
default=True,
help="训练集做随机增强",
)
parser.add_argument(
"--num-workers",
type=int,
default=Config.PRECOMPUTE_NUM_WORKERS,
help="预计算进程数,0 表示自动使用全部 CPU 核心",
)
parser.add_argument(
"--overwrite",
action="store_true",
help="覆盖已有缓存",
)
args = parser.parse_args()
if args.precompute_fea:
split_csv = {
"train": Config.TRAIN_CSV,
"val": Config.VAL_CSV,
"test": Config.TEST_CSV,
}
splits = list(split_csv.keys()) if args.split == "all" else (
["val", "test"] if args.split == "val_test" else [args.split]
)
for split in splits:
csv_path = split_csv[split]
if not os.path.exists(csv_path):
print(f"⚠️ 跳过 {split}: {csv_path} 不存在")
continue
split_use_augment = (split == "train") and args.train_augment
split_num_variants = args.num_variants if split == "train" else 1
split_cache_root = os.path.join(args.cache_root, split)
print(f"\n=== 预计算 {split} (augment={split_use_augment}, variants={split_num_variants}, workers={args.num_workers or 'auto'}) ===")
precompute_fea_cache(
csv_path=csv_path,
cache_root=split_cache_root,
num_variants=split_num_variants,
use_augment=split_use_augment,
overwrite=args.overwrite,
num_workers=args.num_workers,
precompute_cwt=args.precompute_cwt,
)
raise SystemExit(0)
# 简单测试
print("测试 DeepEllipse data_loader")
# 1. 测试单样本加载
dataset = CircleIDDataset(
csv_path=Config.TRAIN_CSV,
sample_ratio=1,
fit_encoder=True,
enable_augment=True,
verbose=True
)
input_all, label, pen_id, writer_id, image_id = dataset[0]
fea_tensor, ske_tensor, img_tensor, *_ = input_all
fea = fea_tensor.fields
print(f"样本:")
print(f" data column: {fea_tensor.data.shape} len={FeaPackLayout.len_core()}")
print(f" g_global: {fea.g_global.shape} u_man: {fea.u_man.shape}")
print(f" s_res: {fea.s_res.shape} valid: {fea.valid.shape}")
print(f" amp_global: {fea.amp_global.shape} amp_segment: {fea.amp_segment.shape}")
print(f" ske: {ske_tensor.shape}, img: {img_tensor.shape}")
print(f" Label: {label}, pen_id: {pen_id}, Writer: {writer_id}")
# 2. 测试 DataLoader
train_loader, val_loader, test_loader, encoder = create_dataloaders(
num_workers=1,
batch_size=4,
sample_ratio=0.001,
verbose=True
)
loader = test_loader if test_loader is not None else train_loader
loader_name = "test" if test_loader is not None else "train"
for batch_idx, batch in enumerate(islice(loader, 10)):
input_all, label, pen_id, writer_id, image_id = batch
fea_tensor, ske_tensor, img_tensor = input_all
fea = fea_tensor.fields
print(f"[{loader_name} batch {batch_idx}] data={fea_tensor.data.shape}, "
f"u_man={fea.u_man.shape}, amp_seg={fea.amp_segment.shape}")
print(f" Label: {label}, pen_id: {pen_id}, Writer: {writer_id}")
print("所有测试通过!")
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
@Time : 2026/3/17 11:04
@File : feature_data_loader.py.py
@Description :
DeepEllipse 数据加载模块
功能:
- 加载 train/val/test CSV 文件
- 椭圆特征打包为 FeaTensor 列向量,反解后供 section41/42 / deep_elipse_model 使用
- 数据增强:二值化/骨架化 + 旋转 + 仿射变换
- 标签编码:writer_id ↔ label_id 自动转换
"""
import copy
import os
import random
from concurrent.futures import ProcessPoolExecutor, as_completed
from itertools import islice
import cv2
import numpy
import numpy as np
import pandas as pd
import torch
from matplotlib import pyplot as plt
from torch.utils.data import Dataset, DataLoader, Sampler
from torchvision import transforms
from PIL import Image
from sklearn.preprocessing import LabelEncoder
import json
from typing import Optional, Union, List, Tuple, Dict, Sequence
import logging
from augmentation import CircleAugmentation
from elipse_feature import (
DEFAULT_N_THETA,
DEFAULT_N_SEGMENTS,
SegmentGeometry,
FeaTensor,
FeaPackLayout,
FeaFields,
unpack_fea_fields,
extract_ellipse_features,
compute_cwt_scalogram,
compute_fft_amplitude_global,
compute_fft_amplitude_segmented,
)
from image_precess import preprocess_writer_image, resize_image
# from augmentation import CircleAugmentation
from section41 import build_u_man
def _collate_fea_tensor(batch_fea: Sequence[FeaTensor]) -> FeaTensor:
return FeaTensor.stack(batch_fea)
def circle_id_collate_fn(batch):
"""DataLoader collate:FeaTensor 保持紧凑 batch;input_all 可含可选 cwt_scalo。"""
input_alls, labels, pen_ids, writers, image_ids = zip(*batch)
first = input_alls[0]
if len(first) == 2:
ske_list = [ia[0] for ia in input_alls]
img_list = [ia[1] for ia in input_alls]
return (
(None, torch.stack(ske_list, dim=0), torch.stack(img_list, dim=0), None),
torch.tensor(labels),
torch.tensor(pen_ids),
list(writers),
list(image_ids),
)
fea_list = [ia[0] for ia in input_alls]
ske_list = [ia[1] for ia in input_alls]
img_list = [ia[2] for ia in input_alls]
cwt_list = [ia[3] if len(ia) > 3 else None for ia in input_alls]
cwt_batch = None
if cwt_list and all(c is not None for c in cwt_list):
cwt_batch = torch.stack(cwt_list, dim=0)
return (
(
_collate_fea_tensor(fea_list),
torch.stack(ske_list, dim=0),
torch.stack(img_list, dim=0),
cwt_batch,
),
torch.tensor(labels),
torch.tensor(pen_ids),
list(writers),
list(image_ids),
)
class Config:
"""全局配置参数"""
_IS_WINDOWS = os.name == "nt"
# 路径配置(按操作系统自动切换)
if _IS_WINDOWS:
IMAGE_ROOT = r"/path/to/circle_id\data"
DATA_SPLIT_ROOT = r"/path/to/circle_id\data"
TRAIN_CSV = os.path.join(DATA_SPLIT_ROOT, "train.csv")
VAL_CSV = os.path.join(DATA_SPLIT_ROOT, "train.csv")