torch 加载他人的ckpt报错

最新推荐文章于 2025-08-11 11:11:35 发布

原创

最新推荐文章于 2025-08-11 11:11:35 发布 · 1.3k 阅读

3 ·

CC 4.0 BY-SA版权

文章标签：

#tensorflow #人工智能 #python

本文介绍了使用torch加载他人预训练模型检查点（ckpt）时常见的两个错误：一是CUDA设备与CPU之间的不匹配问题；二是状态字典加载时出现的缺失键和意外键问题。并提供了具体的解决方案。

torch 加载他人的ckpt报错

第一个错
- 原因
第二个错
- 原因

第一个错

Attempting to deserialize object on a CUDA device but torch.cuda.is_available() is False. If you are running on a CPU-only machine, please use torch.load with map_location=torch.device('cpu') to map your storages to the CPU.

原因

ckpt可能是用gpu跑的，而你用cpu，所以不匹配。需要加上参数 map_location=“cpu“ 即可。

model.load_state_dict(torch.load("net.ckpt",map_location="cpu"))

第二个错

Error(s) in loading state_dict for HITNet_KITTI:
	Missing key(s) in state_dict: "feature_extractor.down_0.0.weight", "feature_extractor.down_0.0.bias", "feature_extractor.down_1.0.weight", "feature_extractor.down_1.0.bias", "feature_extractor.down_1.2.weight", "feature_extractor.down_1.2.bias", "feature_extractor.down_2.0.weight", "feature_extractor.down_2.0.bias", "feature_extractor.down_2.2.weight", "feature_extractor.down_2.2.bias", "feature_extractor.down_3.0.weight", "feature_extractor.down_3.0.bias", "feature_extractor.down_3.2.weight", "feature_extractor.down_3.2.bias", "feature_extractor.down_4.0.weight", "feature_extractor.down_4.0.bias", "feature_extractor.down_4.2.weight", "feature_extractor.down_4.2.bias", "feature_extractor.down_4.4.weight", "feature_extractor.down_4.4.bias", "feature_extractor.down_4.6.weight", "feature_extractor.down_4.6.bias", "feature_extractor.up_3.up_conv.0.weight", "feature_extractor.up_3.up_conv.0.bias", "feature_extractor.up_3.merge_conv.0.weight", "feature_extractor.up_3.merge_conv.0.bias", "feature_extractor.up_3.merge_conv.2.weight", "feature_extractor.up_3.merge_conv.2.bias", "feature_extractor.up_3.merge_conv.4.weight", "feature_extractor.up_3.merge_conv.4.bias", "feature_