LV.1

FrostLee

79积分0赞

7 帖子 43 回复 0 收藏

TA的动态

TA的帖子

TA的回复

yolov5模型基于CNStream框架使用CNCV前处理报错 我的回复：#3fcam回复[图片]可以看库的后缀，另外你的driver是4.6 ？[图片]或者cncv和toolkit都可以看deb包的名字，有显示版本号展开镜像的环境搭建应该不是通过deb安装的，都是so库直接拷贝来用的，我只能贴一下CNStream程序依赖库的信息：linux-vdso.so.1 (0x0000ffff9ae6a000) libcnstream_va.so.6.2.0 => /root/CNStream_target/lib/libcnstream_va.so.6.2.0 (0x0000ffff9a8b2000) libcnstream_core.so.6.2.0 => /root/CNStream_target/lib/libcnstream_core.so.6.2.0 (0x0000ffff9a69f000) libeasydk.so.3 => /root/CNStream_target/lib/libeasydk.so.3 (0x0000ffff9a592000) libopencv_highgui.so.3.4 => /root/CNStream_target/lib/libopencv_highgui.so.3.4 (0x0000ffff9a579000) libopencv_features2d.so.3.4 => /root/CNStream_target/lib/libopencv_features2d.so.3.4 (0x0000ffff9a4e5000 ) libgflags.so.2.2 => /root/CNStream_target/lib/libgflags.so.2.2 (0x0000ffff9a4c4000) libglog.so.0 => /root/CNStream_target/lib/libglog.so.0 (0x0000ffff9a477000) libpthread.so.0 => /lib/aarch64-linux-gnu/libpthread.so.0 (0x0000ffff9a44b000) libdl.so.2 => /lib/aarch64-linux-gnu/libdl.so.2 (0x0000ffff9a436000) libopencv_flann.so.3.4 => /root/CNStream_target/lib/libopencv_flann.so.3.4 (0x0000ffff9a3d3000) libcnrt.so => /root/CNStream_target/lib/libcnrt.so (0x0000ffff99bcf000) libcndrv.so => /root/CNStream_target/lib/libcndrv.so (0x0000ffff99acc000) libcncodec.so => /root/CNStream_target/lib/libcncodec.so (0x0000ffff99a4b000) libion.so => /root/CNStream_target/lib/libion.so (0x0000ffff99a39000) libcncv.so.0.4.602 => /root/CNStream_target/lib/libcncv.so.0.4.602 (0x0000ffff97e85000) libopencv_videoio.so.3.4 => /root/CNStream_target/lib/libopencv_videoio.so.3.4 (0x0000ffff97e45000) libopencv_imgcodecs.so.3.4 => /root/CNStream_target/lib/libopencv_imgcodecs.so.3.4 (0x0000ffff97d24000) libopencv_imgproc.so.3.4 => /root/CNStream_target/lib/libopencv_imgproc.so.3.4 (0x0000ffff979ac000) libopencv_core.so.3.4 => /root/CNStream_target/lib/libopencv_core.so.3.4 (0x0000ffff97714000) libavcodec.so.58 => /root/CNStream_target/lib/libavcodec.so.58 (0x0000ffff96549000) libavformat.so.58 => /root/CNStream_target/lib/libavformat.so.58 (0x0000ffff96331000) libavutil.so.56 => /root/CNStream_target/lib/libavutil.so.56 (0x0000ffff962bb000) libswscale.so.5 => /root/CNStream_target/lib/libswscale.so.5 (0x0000ffff9624e000) libswresample.so.3 => /root/CNStream_target/lib/libswresample.so.3 (0x0000ffff9623c000) libliveMedia.so.81 => /root/CNStream_target/lib/libliveMedia.so.81 (0x0000ffff96107000) libUsageEnvironment.so.3 => /root/CNStream_target/lib/libUsageEnvironment.so.3 (0x0000ffff960f4000) libBasicUsageEnvironment.so.1 => /root/CNStream_target/lib/libBasicUsageEnvironment.so.1 (0x0000ffff960d a000) libgroupsock.so.8 => /root/CNStream_target/lib/libgroupsock.so.8 (0x0000ffff960ae000) libSDL2-2.0.so.0 => /root/CNStream_target/lib/libSDL2-2.0.so.0 (0x0000ffff95f45000) librdkafka.so.1 => /root/CNStream_target/lib/librdkafka.so.1 (0x0000ffff95cc1000) librt.so.1 => /lib/aarch64-linux-gnu/librt.so.1 (0x0000ffff95caa000) libstdc++.so.6 => /usr/lib/aarch64-linux-gnu/libstdc++.so.6 (0x0000ffff95b16000) libm.so.6 => /lib/aarch64-linux-gnu/libm.so.6 (0x0000ffff95a5d000) libgcc_s.so.1 => /lib/aarch64-linux-gnu/libgcc_s.so.1 (0x0000ffff95a37000) libc.so.6 => /lib/aarch64-linux-gnu/libc.so.6 (0x0000ffff958dd000) /lib/ld-linux-aarch64.so.1 (0x0000ffff9ae3f000) libjpu.so => /root/CNStream_target/lib/libjpu.so (0x0000ffff958b6000) 0

关于CNStream解码速率问题 我的回复：#2三叶虫回复您好，因已超期，您侧未更新回复信息，此帖超时关闭，后续遇到问题，请您发起新帖，版主会及时分析处理。好的，谢谢。 0

yolov5模型基于CNStream框架使用CNCV前处理报错 我的回复：#1fcam回复请补充 cnstream cntoolkit cncv 版本信息CNSTREAM VERSION:v6.2.0 其他两个版本咋看，跑sample也没有打印。 0

yolov5模型量化问题 我的回复：#6三叶虫回复请问您现在问题解决了吗？解决了解决了。 0

寒武纪支持pytorch量化感知训练后的模型吗？ 我的回复：#1JCBCam回复220只能接受普通的fp32 pytorch模型好，谢谢。 0

寒武纪支持pytorch量化感知训练后的模型吗？ 我的回复：#2fcam回复您好，MLU220不支持QAT训练后的量化模型。好的，谢谢。 0

yolov5模型量化问题 我的回复：#4fcam回复报错代码位置只是在load模型（CPU）阶段，和MLU没有任何关系。从提供的log信息看，没有发现model和checkpoint size不匹配，这个比较奇怪。如果怀疑和MLU相关，可以注释掉量化代码和其他MLU相关代码，先排查单纯的模型load是否有问题。展开你说得对，还没调用MLU的接口，这里错误是因为使用非量化模型加载了量化参数导致的错误。 0

yolov5模型量化问题 我的回复：量化代码如下：def quantification(imgfile, mlu220 = False): with torch.no_grad(): ckpt = torch.load(opt.weight, map_location=torch.device('cpu')) # load checkpoint state_dict = ckpt['model'] if 'model' in ckpt else ckpt model = Model(img_sizes = opt.img_size, conf_thres = 0.2, nms_thres = 0.4, maxBoxNum=1024, cfg=opt.cfg, ch=3, mlu_postprocess= opt.mlu_postprocess).to(torch.device('cpu')) # create model.load_state_dict(state_dict, strict=False) # load model.eval().float().fuse() model = model.to(torch.device('cpu')) model = mlu_quantize.quantize_dynamic_mlu(model, { 'iteration':1000,'mean': mean, 'std': std, 'data_scale':1.0, 'firstconv': mlu220,'perchannel':True, 'use_avg':False }, dtype='int8', gen_quant=True ) calibration(model, "./coco-eval",20) torch.save(model.state_dict(), save_path) 0

转换fasestdet模型报错 我的回复：#14含含回复1. 关于.data， .data是早期torch的用法，.data和.detach()框架内部的运算是不一样的2. 生成的离线模型会包括运行指令，会比.pt 大一些展开感谢感谢。 0

转换fasestdet模型报错 我的回复：#12含含回复可以尝试修改下代码将.data 更换成.detach：diff --git a/module/shufflenetv2.py b/module/shufflenetv2.pyindex 89d0eb7..80cd875 100644--- a/module/shufflenetv2.py+++ b/module/shufflenetv2.py@@ -54,7 +54,8 @@ class ShuffleV2Block(nn.Module): return torch.cat((self.branch_proj(x_proj), self.branch_main(x)), 1) def channel_shuffle(self, x):- batchsize, num_channels, height, width = x.data.size()+ #batchsize, num_channels, height, width = x.data.size()+ batchsize, num_channels, height, width = x.detach().size() assert (num_channels % 4 == 0) x = x.reshape(batchsize * num_channels // 2, 2, height * width) x = x.permute(1, 0, 2)展开按你说的修改，或者把“.data”去掉就能转换成功了，这是为什么？另外，转换完的.cambricon为啥比原来的.pt文件还大？ 0

上一页 1 2 3 下一页