search

zuoqing1988 / ZQCNN

一款推理框架，同时有很多有用的demo，觉得好用请点星啊
MIT License
2.22k stars 508 forks source link

感谢分享，我想请教一个问题，请问人脸检测精选的mxnet模型如何成功的转换成ncnn的模型的啊？ #64

Open qaz734913414 opened 5 years ago

qaz734913414 commented 5 years ago

感谢分享，我想请教一个问题，请问人脸检测精选的mxnet模型如何成功的转换成ncnn的模型的啊？

zuoqing1988 commented 5 years ago

手工删掉最后无用的几层，然后再根据实际情况确定param最前面的layer数量和blob数量

qaz734913414 commented 5 years ago

7767517 17 18 Input data 0 1 data 0=16 1=16 2=3 Convolution conv1 1 1 data conv1 0=16 1=3 11=3 3=2 13=2 5=1 6=432 PReLU prelu1 1 1 conv1 prelu1 0=16 ConvolutionDepthWise conv2_dw 1 1 prelu1 conv2_dw 0=16 1=3 11=3 3=2 13=2 5=1 6=144 7=16 PReLU prelu2_dw 1 1 conv2_dw prelu2_dw 0=16 Convolution conv2_sep 1 1 prelu2_dw conv2_sep 0=24 1=1 11=1 5=1 6=384 PReLU prelu2 1 1 conv2_sep prelu2 0=24 ConvolutionDepthWise conv3_dw 1 1 prelu2 conv3_dw 0=24 1=3 11=3 5=1 6=216 7=24 PReLU prelu3_dw 1 1 conv3_dw prelu3_dw 0=24 Split splitncnn_0 1 2 prelu3_dw prelu3_dw_splitncnn_0 prelu3_dw_splitncnn_1 Convolution conv4_1 1 1 prelu3_dw_splitncnn_1 conv4_1 0=2 1=1 11=1 5=1 6=48 BatchNorm bn4_1 1 1 conv4_1 bn4_1 0=2 Reshape conv4_1_reshape 1 1 bn4_1 conv4_1_reshape 0=2 Softmax cls_prob 1 1 conv4_1_reshape cls_prob 0=0 Convolution conv4_2 1 1 prelu3_dw_splitncnn_0 conv4_2 0=4 1=1 11=1 5=1 6=96 BatchNorm bn4_2 1 1 conv4_2 bn4_2 0=4 Reshape conv4_2_reshape 1 1 bn4_2 conv4_2_reshape 0=4

可以取出conv4_2的数据，但是cls_prob取不出数据，是不是不是这样改的啊？

zuoqing1988 commented 5 years ago

可以先尝试我改好的，再对比对比就都会了 https://github.com/zuoqing1988/ZQCNN-MTCNN-vs-libfacedetection/tree/master/NCNN/model

qaz734913414 commented 5 years ago

谢谢指导，我去试一下

qaz734913414 commented 5 years ago

convert cost: 2.972 ms nms cost: 0.043 ms, (4-->0) nms cost: 0.063 ms, (2-->0) nms cost: 0.030 ms, (11-->1) nms cost: 0.015 ms, (7-->0) nms cost: 0.019 ms, (18-->2) nms cost: 0.013 ms, (24-->4) nms cost: 0.017 ms, (31-->5) nms cost: 0.008 ms, (14-->3) nms cost: 0.004 ms, (6-->1) nms cost: 0.000 ms, (0-->0) nms cost: 0.001 ms, (0-->0) nms cost: 0.006 ms first stage candidate count: 16 stage 1: cost 75.057 ms run Rnet [7] times, candidate after nms: 1 stage 2: cost 7.097 ms run Onet [1] times, candidate after nms: 1 stage 3: cost 7.073 ms final found num: 1 total cost: 92.706 ms (P: 78.079 ms, R: 7.322 ms, O: 7.304 ms) total 12.409 s / 100 = 124.089 ms

我尝试将https://github.com/zuoqing1988/ZQCNN-MTCNN-vs-libfacedetection 项目中NCNN跑了一下,编译成32位开启了openMp,没有开指令优化，速度只有124ms,设备配置是Intel(R) Pentium(R) CPU G2020 @ 2.90GHz image 请问还可以优化速度不？

zuoqing1988 commented 5 years ago

ncnn本身就没有x86的优化，你得用arm跑

qaz734913414 commented 5 years ago

原来如此，看来需要移植到移动端再来看效果了，谢谢您的解答。