Compare commits
2 Commits
| Author | SHA1 | Date | |
|---|---|---|---|
|
538f987a10 | ||
|
|
2c2798fe04 |
97
README.md
97
README.md
@@ -14,8 +14,8 @@ conda activate yolo
|
||||
# 2. 安装 PyTorch (CUDA 12.4)
|
||||
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124
|
||||
|
||||
# 3. 安装 Ultralytics
|
||||
pip install ultralytics
|
||||
# 3. 安装 Ultralytics 和 常用工具
|
||||
pip install ultralytics tqdm opencv-python pyyaml
|
||||
|
||||
# 4. 其它依赖
|
||||
pip install opencv-python matplotlib albumentations
|
||||
@@ -29,36 +29,43 @@ watch -n 1 nvidia-smi
|
||||
|
||||
---
|
||||
|
||||
## 2. 数据集 (V2 大规模合并版)
|
||||
## 2. 数据集 (V2 大规模合并版 - 22 类)
|
||||
|
||||
### 2.1 数据集组成
|
||||
|
||||
我们合并了多个来源的高质量分割数据集，总数据量约 **12,400 张**：
|
||||
我们合并了多个来源的高质量分割数据集，数据源经过精心清洗和重映射：
|
||||
|
||||
| 数据集来源 | 描述 | 作用 |
|
||||
|------|------|------|
|
||||
| **MIT Indoor** | 室内场景分类 | 提供丰富背景和基础物体 |
|
||||
| **Indoor Blind** | 早期筛选的导盲数据 | 基础核心数据 |
|
||||
| **Stair Seg** | 楼梯分割 | 增强对楼梯的识别 (Map to `stairs`) |
|
||||
| **MIT Indoor** | 室内场景分类 | 核心补充：提供 `appliance`, `tableware`, `furniture` 等日常物品 |
|
||||
| **Indoor Blind** | 早期筛选的导盲数据 | 基础核心数据 (Walkable areas) |
|
||||
| **Washroom** | 卫浴分割 | 覆盖 `toilet`, `sink` 等卫生间关键设施 |
|
||||
| **Spoon Fork Chopstick**| 餐具特写 | 极大增强 `tableware` 中细小餐具的识别 (筷子/勺子/叉子) |
|
||||
| **Stair Seg** | 楼梯分割 | 增强对楼梯的识别 |
|
||||
| **Stair Chair Couch**| 楼梯/椅子分割 | 补充高质量椅子与楼梯数据 |
|
||||
| **Estima AI** | 室内房间分割 | 映射为 **`floor`**，极大增强可行走区域识别 |
|
||||
|
||||
### 2.2 类别定义 (20 类)
|
||||
**已移除**:
|
||||
- **Estima AI**: 因全部为平面设计图而非实景图，已移除。
|
||||
|
||||
统一后的 20 个导盲核心类别：
|
||||
### 2.2 类别定义 (22 类)
|
||||
|
||||
| ID | 类别 | 说明 | ID | 类别 | 说明 |
|
||||
统一后的 22 个导盲核心类别：
|
||||
|
||||
| ID | 类别 | 中文 | ID | 类别 | 中文 |
|
||||
|----|------|------|----|------|------|
|
||||
| 0 | `floor` | 可行走地面 | 10 | `stairs` | 楼梯 |
|
||||
| 1 | `corridor` | 走廊/通道 | 11 | `wall` | 墙壁 |
|
||||
| 2 | `sidewalk` | 人行道 | 12 | `window` | 窗户 |
|
||||
| 3 | `chair` | 椅子 | 13 | `cabinet` | 柜子 |
|
||||
| 4 | `table` | 桌子 | 14 | `trash_can`| 垃圾桶 |
|
||||
| 5 | `sofa_bed` | 沙发/床 | 15 | `person` | 行人 |
|
||||
| 6 | `door` | 门 | 16 | `cup_bottle`| 杯子/瓶子 |
|
||||
| 7 | `elevator` | 电梯 | 17 | `bag` | 包 |
|
||||
| 8 | `plant` | 植物 | 18 | `electronics`| 电子电器 |
|
||||
| 9 | `clock` | 时钟/挂钟 | 19 | `obstacle` | 通用障碍物 |
|
||||
| 0 | `floor` | 地面 | 11 | `wall` | 墙壁 |
|
||||
| 1 | `corridor` | 走廊 | 12 | `window` | 窗户 |
|
||||
| 2 | `sidewalk` | 人行道 | 13 | `cabinet` | 柜子 |
|
||||
| 3 | `chair` | 椅子 | 14 | `trash_can`| 垃圾桶 |
|
||||
| 4 | `table` | 桌子 | 15 | `person` | 行人 |
|
||||
| 5 | `sofa_bed` | 沙发/床 | 16 | `bag` | 包/背包 |
|
||||
| 6 | `door` | 门 | 17 | `electronics`| 电子电器 |
|
||||
| 7 | `elevator` | 电梯 | 18 | `plant` | 植物 |
|
||||
| 8 | `stairs` | 楼梯 | 19 | `obstacle` | 通用障碍 |
|
||||
| 9 | `appliance` | 家电 | 20 | `toilet` | 卫生间/马桶 |
|
||||
| 10 | `sink` | 洗手台 | 21 | `tableware` | 餐具/物品 |
|
||||
|
||||
> **注**: `tableware` (ID 21) 包含杯子、碗、盘子、勺子、筷子、瓶子等。
|
||||
|
||||
### 2.3 目录结构 (服务器端)
|
||||
|
||||
@@ -68,15 +75,12 @@ watch -n 1 nvidia-smi
|
||||
/home/rongye/ProgramFiles/Yolo/
|
||||
├── yolo11l-seg.pt # 预训练权重 (标准模型)
|
||||
├── train_merged.py # 🔥 主力训练脚本 (使用合并数据)
|
||||
├── train.py # 🛡️ 备份训练脚本 (使用旧数据)
|
||||
├── datasets/
|
||||
│ ├── blind_guidance_merged/ # 🔥 V2 主力数据集 (12k images)
|
||||
│ │ ├── data.yaml # 配置 (20类)
|
||||
│ │ ├── train/
|
||||
│ │ ├── valid/
|
||||
│ │ └── test/
|
||||
│ └── indoor_blind/ # 🛡️ 备份数据集 (1.8k images)
|
||||
│ └── data.yaml # 配置 (14类)
|
||||
│ └── blind_guidance_merged/ # 🔥 V2 主力数据集
|
||||
│ ├── data.yaml # 配置 (22类, 路径必须正确)
|
||||
│ ├── train/
|
||||
│ ├── valid/
|
||||
│ └── test/
|
||||
└── blind_guide_project/ # 训练日志输出
|
||||
```
|
||||
|
||||
@@ -86,15 +90,17 @@ watch -n 1 nvidia-smi
|
||||
|
||||
### 3.1 主力配置 `blind_guidance_merged/data.yaml`
|
||||
|
||||
**关键点**：`path` 必须是服务器上的绝对路径。
|
||||
|
||||
```yaml
|
||||
# 路径必须为绝对路径
|
||||
# 路径必须为服务器绝对路径
|
||||
path: /home/rongye/ProgramFiles/Yolo/datasets/blind_guidance_merged
|
||||
train: train/images
|
||||
val: valid/images
|
||||
test: test/images
|
||||
|
||||
nc: 20
|
||||
names: [floor, corridor, sidewalk, chair, table, sofa_bed, door, elevator, stairs, wall, window, cabinet, trash_can, person, cup_bottle, bag, electronics, plant, clock, obstacle]
|
||||
nc: 22
|
||||
names: [floor, corridor, sidewalk, chair, table, sofa_bed, door, elevator, stairs, wall, window, cabinet, trash_can, person, bag, electronics, plant, obstacle, appliance, toilet, sink, tableware]
|
||||
```
|
||||
|
||||
### 3.2 训练脚本 `train_merged.py`
|
||||
@@ -117,40 +123,34 @@ results = model.train(
|
||||
optimizer="AdamW", # 优化器
|
||||
close_mosaic=15, # 最后15轮关闭增强
|
||||
project="blind_guide_project",
|
||||
name="yolo11l_blind_v2" # V2 版本
|
||||
name="yolo11l_blind_v2", # V2 版本
|
||||
amp=False # ⚠️ 关键：必须关闭混合精度以防止 Loss NaN/Inf
|
||||
)
|
||||
```
|
||||
|
||||
> **注意**: 如果训练出现 `Loss NaN/Inf`，请确保设置 `amp=False`。混合精度虽然省显存，但在某些特定数据集分布下会导致梯度溢出。
|
||||
|
||||
---
|
||||
|
||||
## 4. 开始训练
|
||||
## 4. 开始训练流程
|
||||
|
||||
### 步骤 1: 上传数据
|
||||
将本地生成的 `blind_guidance_merged` 文件夹完整上传到服务器 `/home/rongye/ProgramFiles/Yolo/datasets/` 目录。
|
||||
|
||||
> **⚠️ 特别提醒**: 上传后，请务必检查服务器上的 `data.yaml` 中的 `path` 字段是否为 `/home/rongye/ProgramFiles/Yolo/datasets/blind_guidance_merged`。如果还是 Windows 路径，请手动修改或上传正确的版本。
|
||||
|
||||
### 步骤 2: 运行训练
|
||||
```bash
|
||||
conda activate yolo
|
||||
cd /home/rongye/ProgramFiles/Yolo
|
||||
python train_merged.py
|
||||
```
|
||||
|
||||
### 步骤 3: 监控
|
||||
使用 `watch -n 1 nvidia-smi` 查看显存占用，确保训练正常开始。
|
||||
使用 `watch -n 1 nvidia-smi` 查看显存占用。
|
||||
|
||||
---
|
||||
|
||||
## 5. 备份方案 (旧数据)
|
||||
|
||||
如果需要回退到旧版本训练（仅 14 类，数据量较小）：
|
||||
|
||||
1. 使用备份脚本：`train.py`
|
||||
2. 数据路径：`/home/rongye/ProgramFiles/Yolo/datasets/indoor_blind/data.yaml`
|
||||
3. 运行：`python train.py`
|
||||
|
||||
---
|
||||
|
||||
## 6. 模型导出 (TensorRT)
|
||||
## 5. 模型导出 (TensorRT)
|
||||
|
||||
训练完成后 (V2)，最佳模型位于 `blind_guide_project/yolo11l_blind_v2/weights/best.pt`。
|
||||
|
||||
@@ -158,6 +158,7 @@ python train_merged.py
|
||||
```python
|
||||
from ultralytics import YOLO
|
||||
model = YOLO("blind_guide_project/yolo11l_blind_v2/weights/best.pt")
|
||||
# 导出为 TensorRT engine, 半精度 fp16, 动态尺寸固定为 480 (或 640 根据推理端需求)
|
||||
model.export(format="engine", imgsz=480, half=True, device=0)
|
||||
```
|
||||
|
||||
目标文件名：`yolo11l-seg-indoor.engine` (导出后重命名)
|
||||
|
||||
@@ -13,13 +13,15 @@ names:
|
||||
- cabinet
|
||||
- trash_can
|
||||
- person
|
||||
- cup_bottle
|
||||
- bag
|
||||
- electronics
|
||||
- plant
|
||||
- clock
|
||||
- obstacle
|
||||
nc: 20
|
||||
- appliance
|
||||
- toilet
|
||||
- sink
|
||||
- tableware
|
||||
nc: 22
|
||||
path: /home/rongye/ProgramFiles/Yolo/datasets/blind_guidance_merged
|
||||
test: test/images
|
||||
train: train/images
|
||||
|
||||
30
datasets/floor/data.yaml
Normal file
30
datasets/floor/data.yaml
Normal file
@@ -0,0 +1,30 @@
|
||||
train: train/images
|
||||
val: valid/images
|
||||
test: test/images
|
||||
|
||||
nc: 22
|
||||
names:
|
||||
- floor
|
||||
- corridor
|
||||
- sidewalk
|
||||
- chair
|
||||
- table
|
||||
- sofa_bed
|
||||
- door
|
||||
- elevator
|
||||
- stairs
|
||||
- wall
|
||||
- window
|
||||
- cabinet
|
||||
- trash_can
|
||||
- person
|
||||
- bag
|
||||
- electronics
|
||||
- plant
|
||||
- obstacle
|
||||
- appliance
|
||||
- toilet
|
||||
- sink
|
||||
- tableware
|
||||
|
||||
path: /home/rongye/ProgramFiles/Yolo/datasets/floor
|
||||
11
floor_finetune.md
Normal file
11
floor_finetune.md
Normal file
@@ -0,0 +1,11 @@
|
||||
```bash
|
||||
yolo segment train \
|
||||
model=yolo11l-seg-indoor.pt \
|
||||
data=datasets/floor/data.yaml \
|
||||
epochs=10 \
|
||||
freeze=20 \
|
||||
amp=False \
|
||||
device=1 \
|
||||
project=/home/rongye/ProgramFiles/Yolo/blind_guide_project \
|
||||
name=floor_finetune
|
||||
```
|
||||
Reference in New Issue
Block a user