解决PointNet++语义分割模型中类别修改导致的断言错误与标签处理

/opt/conda/conda-bld/pytorch_1614378098133/work/aten/src/THCUNN/ClassNLLCriterion.cu:108: cunn_ClassNLLCriterion_updateOutput_kernel: block: [0,0,0], thread: [10,0,0] Assertion `t >= 0 && t < n_classes` failed.

import numpy as np
import torch

# 假设你已经加载了数据集，并且可以访问到所有样本的真实标签
# 这里用一个示例列表代替实际的数据集标签
# 错误的示例：标签不是从0开始且不连续
# all_dataset_labels = [1, 5, 10, 1, 5, 10, 1, 5, 10]
# 正确的示例：标签从0开始且连续，对应3个类别
# all_dataset_labels = [0, 1, 2, 0, 1, 2, 0, 1, 2]
# 另一个错误的示例：如果num_classes=17，但数据集中有标签17
all_dataset_labels = [0, 1, ..., 16, 17, 0, 1, ...] # 假设数据集中存在标签17

# 假设模型定义的类别数量
num_classes_in_model = 17

# 收集数据集中所有唯一的标签
unique_labels_in_dataset = np.unique(all_dataset_labels)

print(f"模型配置的类别数量 (num_classes): {num_classes_in_model}")
print(f"数据集中发现的唯一标签: {unique_labels_in_dataset}")
print(f"数据集中唯一标签的数量: {len(unique_labels_in_dataset)}")

# 检查标签是否符合要求
if len(unique_labels_in_dataset) != num_classes_in_model:
    print("警告：数据集中唯一标签的数量与模型配置的num_classes不匹配！")
elif not (min(unique_labels_in_dataset) == 0 and max(unique_labels_in_dataset) == num_classes_in_model - 1):
    print(f"警告：数据集标签未进行零索引或未完全顺序化。")
    print(f"期望标签范围：0 到 {num_classes_in_model - 1}")
    print(f"实际标签范围：{min(unique_labels_in_dataset)} 到 {max(unique_labels_in_dataset)}")
else:
    print("数据集标签与模型配置的num_classes一致，且已进行零索引和顺序化。")

import torch
from torch.utils.data import Dataset, DataLoader
import numpy as np

# 假设你的原始数据集标签是这样的
# 例如，原始数据集中有3个类别，但它们的ID是10, 20, 30
original_raw_labels = [10, 20, 30, 10, 20, 30, 10, 20, 30]

# 1. 确定所有原始唯一标签
unique_original_labels = sorted(list(np.unique(original_raw_labels)))
print(f"原始数据集中的唯一标签: {unique_original_labels}")

# 2. 创建映射字典
# 假设我们有 len(unique_original_labels) 个类别
num_classes_for_model = len(unique_original_labels)
label_mapping = {
    original_id: new_id
    for new_id, original_id in enumerate(unique_original_labels)
}
print(f"标签映射字典: {label_mapping}")
print(f"模型期望的类别数量 (num_classes): {num_classes_for_model}")

class CustomSegmentationDataset(Dataset):
    def __init__(self, raw_labels, label_map, num_classes):
        self.raw_labels = raw_labels
        self.label_map = label_map
        self.num_classes = num_classes

    def __len__(self):
        return len(self.raw_labels)

    def __getitem__(self, idx):
        original_label = self.raw_labels[idx]
        # 3. 应用映射
        mapped_label = self.label_map.get(original_label, -1) # 如果遇到未知标签，可以抛出错误

        if mapped_label == -1:
            raise ValueError(f"Encountered unmapped label: {original_label}")

        # 确保映射后的标签在 [0, num_classes-1] 范围内
        if not (0 <= mapped_label < self.num_classes):
            raise ValueError(f"Mapped label {mapped_label} out of expected range [0, {self.num_classes-1}]")

        # 在实际应用中，这里还会加载点云数据等
        # 假设这里只返回一个虚拟的点云数据和映射后的标签
        point_cloud_data = torch.randn(1024, 3) # 示例点云数据
        return point_cloud_data, torch.tensor(mapped_label, dtype=torch.long)

# 实例化数据集和数据加载器
dataset = CustomSegmentationDataset(original_raw_labels, label_mapping, num_classes_for_model)
dataloader = DataLoader(dataset, batch_size=2, shuffle=True)

# 模拟训练循环，检查标签是否正确
print("\n模拟数据加载和标签检查：")
for i, (points, labels) in enumerate(dataloader):
    print(f"Batch {i+1}:")
    print(f"  标签 (mapped labels): {labels}")
    print(f"  标签最小值: {labels.min().item()}, 标签最大值: {labels.max().item()}")

    # 再次检查标签是否在正确范围内
    if not (labels.min().item() >= 0 and labels.max().item() < num_classes_for_model):
        raise AssertionError("Mapped labels are still out of range!")

    if i >= 2: # 仅演示几个批次
        break

print("\n标签重映射成功，所有标签都在预期范围内。")