112 lines
2.6 KiB
Markdown
112 lines
2.6 KiB
Markdown
# ✅ 最终修复 - 预训练模型加载问题
|
||
|
||
---
|
||
|
||
## 🎯 问题根源
|
||
|
||
即使注释掉配置文件中的`init_cfg`，`bevfusion.py`的`init_weights`方法仍然会调用Swin Transformer的`init_weights`，导致尝试从GitHub下载预训练模型。
|
||
|
||
---
|
||
|
||
## ✅ 解决方案
|
||
|
||
### 修改 `bevfusion.py` 的 `init_weights` 方法
|
||
|
||
```python
|
||
def init_weights(self) -> None:
|
||
# ✅ 如果从checkpoint加载，跳过预训练模型初始化
|
||
# 只有当backbone配置了init_cfg时才初始化
|
||
if "camera" in self.encoders:
|
||
backbone = self.encoders["camera"]["backbone"]
|
||
# 检查是否有init_cfg配置
|
||
if hasattr(backbone, 'init_cfg') and backbone.init_cfg is not None:
|
||
backbone.init_weights()
|
||
else:
|
||
# 没有init_cfg，说明从checkpoint加载，跳过初始化
|
||
print("[BEVFusion] ⚪ Skipping camera backbone init_weights (will load from checkpoint)")
|
||
```
|
||
|
||
**逻辑**:
|
||
- ✅ 如果backbone有`init_cfg`配置 → 调用`init_weights()`加载预训练模型
|
||
- ✅ 如果backbone没有`init_cfg` → 跳过初始化，等待从checkpoint加载
|
||
|
||
---
|
||
|
||
## 📊 修复位置
|
||
|
||
```
|
||
文件: mmdet3d/models/fusion_models/bevfusion.py
|
||
行数: 第159-169行
|
||
修改: 添加init_cfg检查逻辑
|
||
```
|
||
|
||
---
|
||
|
||
## ✅ 现在的行为
|
||
|
||
### 配置文件中 (已注释)
|
||
|
||
```yaml
|
||
backbone:
|
||
type: SwinTransformer
|
||
...
|
||
# ✅ 从checkpoint加载，无需预训练模型
|
||
# init_cfg:
|
||
# type: Pretrained
|
||
# checkpoint: pretrained/swint-nuimages-pretrained.pth
|
||
```
|
||
|
||
### 代码中 (已修复)
|
||
|
||
```python
|
||
# 检查init_cfg
|
||
if hasattr(backbone, 'init_cfg') and backbone.init_cfg is not None:
|
||
backbone.init_weights() # 有配置才初始化
|
||
else:
|
||
print("Skipping init_weights") # 无配置，跳过
|
||
```
|
||
|
||
### 启动时 (使用checkpoint)
|
||
|
||
```bash
|
||
--load_from epoch_5.pth
|
||
```
|
||
|
||
**结果**:
|
||
- ✅ 不会尝试加载预训练模型
|
||
- ✅ 直接从checkpoint加载所有权重
|
||
- ✅ task_gca随机初始化
|
||
|
||
---
|
||
|
||
## 🚀 现在可以正常启动了！
|
||
|
||
```bash
|
||
docker exec -it bevfusion bash
|
||
cd /workspace/bevfusion
|
||
bash START_PHASE4A_TASK_GCA.sh
|
||
```
|
||
|
||
---
|
||
|
||
## ✅ 启动后应该看到
|
||
|
||
```
|
||
[BEVFusion] ⚪ Skipping camera backbone init_weights (will load from checkpoint)
|
||
[BEVFusion] ✨✨ Task-specific GCA mode enabled ✨✨
|
||
[object] GCA: params: 131,072
|
||
[map] GCA: params: 131,072
|
||
|
||
load checkpoint from .../epoch_5.pth
|
||
|
||
The following keys in model are not found in checkpoint:
|
||
task_gca.* (正常，随机初始化)
|
||
|
||
Epoch [1][50/xxx] ...
|
||
```
|
||
|
||
---
|
||
|
||
**🎉 预训练模型加载问题已彻底解决！现在可以正常启动训练了！**
|
||
|