Toybrick

标题: senet154 toolkit1.3.0 segmentation fault [打印本页]

作者: jy_dll    时间: 2020-2-27 18:40
标题: senet154 toolkit1.3.0 segmentation fault
本帖最后由 jy_dll 于 2020-2-27 18:52 编辑

基于rknn-toolkit-1.3.0 尝试转换pytorch senet154, 用load_pytorch 导入*.pt, 导入成功,但是build失败,log如附件所示,senet154如附件所示,麻烦帮忙看看,非常感谢~
env:torch1.2.0, torchvision0.4.0, tensorflow1.14
不开量化,不预编译
------------------------------------------------------------------------
D Fill zero bias for convolution_at_input14.1_810 (convolution)
D Fill zero bias for convolution_at_input7.1_812 (convolution)
D Fill zero bias for convolution_at_input26.1_813 (convolution)
D Fill zero bias for convolution_at_input11.1_818 (convolution)
D Fill zero bias for convolution_at_input4.1_819 (convolution)
D Fill zero bias for convolution_at_input1.1_822 (convolution)
D Optimizing network with conv2d_big_kernel_size_transform
done
--> Building model..
I Generate input meta
I Load input meta
I Generate input meta
D import clients finished
I Load net...
Segmentation fault (core dumped)






作者: jefferyzhang    时间: 2020-2-28 08:52
已上报问题,NPU部门有回复我会回给你
作者: jefferyzhang    时间: 2020-2-28 17:40
NPU部门回复:

env: torch1.2.0, torchvision0.4.0, tensorflow1.14
我这边试了一下,并未出现段错误,float的pytorch的推理结果和rknn的推理结果是一样的。
请按照如下步骤排查:
1.生成pt是否正确,是否忘记调用net.eval()?
net = senet.senet154()
net.eval()
dummy_input = torch.randn([1, 3, 224, 224])
trace_model = torch.jit.trace(net, dummy_input)
trace_model.save(pt_save_path)
2. 固件或驱动是否是最新版本?
3. 使用rknn_toolkit v1.3.1 beta3:
链接: https://pan.baidu.com/s/1Kn2FGAdF_j3CMLNEsC3OPw 提取码: rcds
作者: wangkelin    时间: 2021-9-3 10:57
se模块不是不支持吗?自定义算子不是只支持tf吗,为什么还能加载.pt文件




欢迎光临 Toybrick (https://t.rock-chips.com/) Powered by Discuz! X3.3