kendryte / nncase

Open deep learning compiler stack for Kendryte AI accelerators ✨
Apache License 2.0
730 stars 181 forks source link

转换ConvNext预训练模型时,提前退出,无报错信息 #1200

Closed ssyj911 closed 4 months ago

ssyj911 commented 4 months ago

Describe the bug 将ConvNext预训练模型转kmodel时,nncase 编译器输出:"2.Optimize target independent..." 后退出。无任何其他提示

To Reproduce 预训练模型可以在以下地址获取: https://dl.fbaipublicfiles.com/convnext/convnext_tiny_1k_224_ema.pth

Origin model and code 按照 nncase 提供的 编译float32 onnx模型 代码进行 kmodel 进行转换

Environment (please complete the following information):

Additional context 尝试将nncase安装包中的 nncase-1.9.0/examples/yolox/model/yolox_nano_224.onnx 进行转换,未发现问题

curioyang commented 4 months ago

@ssyj911 请直接提供onnx模型

ssyj911 commented 4 months ago

@ssyj911 请直接提供onnx模型

这个太大了,上传到百度云盘啦:链接: https://pan.baidu.com/s/1MEIbKCnJama3k1AjjSn4uw?pwd=ymgf 提取码: ymgf

curioyang commented 4 months ago

@ssyj911 编译的问题已经修复了,但是这个模型里有大量的CPU算子layernorm,在实际推理的时候速度很慢,且510 在设计上存在一定的精度损失。 如果在510上部署实际项目,建议更换模型 或者可以尝试在230上部署,int16量化,230对部分CPU算子是有V扩展优化的,且不太容易出现精度损失

ssyj911 commented 4 months ago

非常感谢您的帮助

---原始邮件--- 发件人: "Curio @.> 发送时间: 2024年5月14日(周二) 上午9:35 收件人: @.>; 抄送: @.**@.>; 主题: Re: [kendryte/nncase] 转换ConvNext预训练模型时,提前退出,无报错信息 (Issue #1200)

@ssyj911 编译的问题已经修复了,但是这个模型里有大量的CPU算子layernorm,在实际推理的时候速度很慢,且510 在设计上存在一定的精度损失。 如果在510上部署实际项目,建议更换模型 或者可以尝试在230上部署,int16量化,230对部分CPU算子是有V扩展优化的,且不太容易出现精度损失

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you were mentioned.Message ID: @.***>