RuntimeError: CUDA error: no kernel image is available for execution on the device

[lsj1111]

当我按照上述环境跑代码时，出现了这个错误，请问作者有遇到过吗，怎么解决的

[yuhongtian17]

检查您的显卡型号、"nvidia-smi" cuda版本、"nvcc -V" cuda版本和PyTorch版本是否匹配（例如RTX 4090要求CUDA 11.6及以上）。如果您难以确定，也可以将上述信息提供给我们来帮助您检查。

[lsj1111]

但是当我把CUDA和torch版本都提升之后，感觉就无法安装您源码对应的MMCV了，会导致运行时有很多的bug，因为MMCV2.0版本有太多的更新

[yuhongtian17]

我们的代码版本是基于mmrotate-0.3.3/0.3.4编写的。也就是说只要mmrotate-0.3.3/0.3.4的例如Rotated Faster RCNN r50示例代码能跑通的话我们的代码也能跑通；官方代码也不能跑通的话请到mmrotate仓库下报告issue。应当根据您的显卡型号选择合适的（而非最新的）CUDA和PyTorch版本。

[lsj1111]

当我将maevit的预训练权重加载进来时一切都很正常，但是当我将您的预训练权重加载进来时，似乎不能成功加载导致效果很差，代码如下所示：
pretrained = r'D:\Tool\Datasets\pretrained_weight\Spatial-Transform-Decoupling-Models\pretrained\mae_pretrain_vit_base_full.pth'

angle_version = 'le90'
norm_cfg = dict(type='LN', requires_grad=True)
model = dict(
type='RotatedimTED',
#pretrained=pretrained,
proposals_dim=6,
backbone=dict(
type='VisionTransformer',
init_cfg=dict(type='Pretrained', checkpoint=pretrained),
img_size=224,

[yuhongtian17]

这些描述不足够定位错误。是否可以提供更详细的配置文件和报错信息呢？

[lsj1111]

我总结了一下问题，你提供的可下载的预训练权重中包含多个文件：第一个是MAE预训练的VIT权重-> mae_pretrain_vit_base_full.pth，第二个是你引入了STD模块后训练了12个epoch的权重 ->epoch_12.pth，但是现在有一个很奇怪的事情，当我拿第一个权重训练时得到的MAP值居然是远大于第二个的，理论上来说不是应该第二个权重训练出来的效果会更好吗，所以我很疑惑，还请解答。

[lsj1111]

当我加载你的epoch_12权重时，出现了下述报错信息：
missing keys in source state_dict: fc_cls.weight, fc_cls.bias, decoder_blocks.2.layer_reg.norms.0.weight, decoder_blocks.2.layer_reg.norms.0.bias, decoder_blocks.2.layer_reg.norms.1.weight, decoder_blocks.2.layer_reg.norm
s.1.bias, decoder_blocks.2.layer_reg.norms.2.weight, decoder_blocks.2.layer_reg.norms.2.bias, decoder_blocks.2.layer_reg.convs.0.weight, decoder_blocks.2.layer_reg.convs.0.bias, decoder_blocks.2.layer_reg.convs.1.weight,
decoder_blocks.2.layer_reg.convs.1.bias, decoder_blocks.2.layer_reg.convs.2.weight, decoder_blocks.2.layer_reg.convs.2.bias, decoder_blocks.2.layer_reg.norm_reg.weight, decoder_blocks.2.layer_reg.norm_reg.bias, decoder_bl
ocks.2.layer_reg.fc_reg.weight, decoder_blocks.2.layer_reg.fc_reg.bias, decoder_blocks.3.layer_reg.norms.0.weight, decoder_blocks.3.layer_reg.norms.0.bias, decoder_blocks.3.layer_reg.norms.1.weight, decoder_blocks.3.layer
_reg.norms.1.bias, decoder_blocks.3.layer_reg.norms.2.weight, decoder_blocks.3.layer_reg.norms.2.bias, decoder_blocks.3.layer_reg.convs.0.weight, decoder_blocks.3.layer_reg.convs.0.bias, decoder_blocks.3.layer_reg.convs.1
.weight, decoder_blocks.3.layer_reg.convs.1.bias, decoder_blocks.3.layer_reg.convs.2.weight, decoder_blocks.3.layer_reg.convs.2.bias, decoder_blocks.3.layer_reg.norm_reg.weight, decoder_blocks.3.layer_reg.norm_reg.bias, decoder_blocks.3.layer_reg.fc_reg.weight, decoder_blocks.3.layer_reg.fc_reg.bias, decoder_blocks.4.layer_reg.norms.0.weight, decoder_blocks.4.layer_reg.norms.0.bias, decoder_blocks.4.layer_reg.norms.1.weight, decoder_block
s.4.layer_reg.norms.1.bias, decoder_blocks.4.layer_reg.convs.0.weight, decoder_blocks.4.layer_reg.convs.0.bias, decoder_blocks.4.layer_reg.convs.1.weight, decoder_blocks.4.layer_reg.convs.1.bias, decoder_blocks.4.layer_re
g.norm_reg.weight, decoder_blocks.4.layer_reg.norm_reg.bias, decoder_blocks.4.layer_reg.fc_reg.weight, decoder_blocks.4.layer_reg.fc_reg.bias, decoder_blocks.5.layer_reg.norms.0.weight, decoder_blocks.5.layer_reg.norms.0.
bias, decoder_blocks.5.layer_reg.norms.1.weight, decoder_blocks.5.layer_reg.norms.1.bias, decoder_blocks.5.layer_reg.convs.0.weight, decoder_blocks.5.layer_reg.convs.0.bias, decoder_blocks.5.layer_reg.convs.1.weight, decoder_blocks.5.layer_reg.convs.1.bias, decoder_blocks.5.layer_reg.norm_reg.weight, decoder_blocks.5.layer_reg.norm_reg.bias, decoder_blocks.5.layer_reg.fc_reg.weight, decoder_blocks.5.layer_reg.fc_reg.bias, decoder_blocks.6
.layer_reg.norms.0.weight, decoder_blocks.6.layer_reg.norms.0.bias, decoder_blocks.6.layer_reg.convs.0.weight, decoder_blocks.6.layer_reg.convs.0.bias, decoder_blocks.6.layer_reg.norm_reg.weight, decoder_blocks.6.layer_re
g.norm_reg.bias, decoder_blocks.6.layer_reg.fc_reg.weight, decoder_blocks.6.layer_reg.fc_reg.bias, decoder_blocks.7.layer_reg.norms.0.weight, decoder_blocks.7.layer_reg.norms.0.bias, decoder_blocks.7.layer_reg.convs.0.wei
ght, decoder_blocks.7.layer_reg.convs.0.bias, decoder_blocks.7.layer_reg.norm_reg.weight, decoder_blocks.7.layer_reg.norm_reg.bias, decoder_blocks.7.layer_reg.fc_reg.weight, decoder_blocks.7.layer_reg.fc_reg.bias, decoder_box_norm.weight, decoder_box_norm.bias
我不知道问题出在了哪里。

[yuhongtian17]

是这样的，因为epoch_12.pth是按检测器所定义的变量名作为权重存储字典的key的，它与mae pretrained的权重存储字典的key不一致。而检测器定义的init_weight函数是按mae pretrained的字典key去查找对应value和加载权重的。因此epoch_12.pth不能直接作为检测器的预训练模型从头开始全监督预训练，否则就会出现未能成功加载权重、相当于随机初始化的情形；如果需要权重微调，则应按test时的整个模型load_state_dict方式加载权重然后开始训练。

[quxianjiuguo]

Can I use a single 4090 in this github to train my own dataset？

[yuhongtian17]

See issue #7

[WenLinLliu]

Can I use a single 4090 in this github to train my own dataset？

yes，you can，no problem