ZQCNN
ZQCNN copied to clipboard
zuoqing1988
2019年2月14号公布的MTCNN版本的模型没有用
2019年2月14号公布的MTCNN版本的模型，在SampleMTCNN.cpp代码中改变了加载模型的文件路径，程序可以跑通，但是检测不出人脸，同样一张人脸照片，用原先的模型可以识别出来人脸，但是用您在2月14号公布的模型（我随便选了其中的各层进行组合）则检测不出人脸，请问我应该怎么办？
Init函数中load模型的时候有有true，1e-9，true这样的参数，改成false试试。 某些模型merge_bn之后就不行了，具体原因不明，可能是数值太大或者太小。
根据您说的，我随便试了几下：将一些true改为false，变1e-9为1e-12,仍然不行。这些参数分别是什么意思，请问我应该按照什么原则调呢？ 以下是其中一种情况的输出结果： in Layer conv1n para bias in Layer conv2_dwara bias in Layer conv2_sepra bias in Layer conv3_dwara bias in Layer conv3_sepra bias in Layer conv4_dwara bias in Layer conv4_1para bias in Layer bn4_1n para bias in Layer conv4_2para bias in Layer bn4_2n para bias in Layer conv1n para bias in Layer conv2_dwara bias in Layer conv2_sepra bias in Layer conv3_dwara bias in Layer conv3_sepra bias in Layer conv4_dwara bias in Layer conv4_sepra bias in Layer conv5_dwara bias in Layer conv5_1para bias in Layer bn5_1n para bias in Layer conv5_2para bias in Layer bn5_2n para bias in Layer conv1n para bias in Layer conv2_dwara bias in Layer conv2_sepra bias in Layer conv3_dwara bias in Layer conv3_sepra bias in Layer conv4_dwara bias in Layer conv4_sepra bias in Layer conv5_dwara bias in Layer conv5_sepra bias in Layer conv6_dwara bias in Layer conv6_1para bias in Layer bn6_1n para bias in Layer conv6_2para bias in Layer bn6_2n para bias in Layer conv1n para bias in Layer bn1own para bias in Layer conv2_dwara bias in Layer bn2_dw para bias in Layer conv2_sepra bias in Layer bn2_seppara bias in Layer conv3_dwara bias in Layer bn3_dw para bias in Layer conv3_sepra bias in Layer bn3_seppara bias in Layer conv4_dwara bias in Layer bn4_dw para bias in Layer conv4_sepra bias in Layer bn4_seppara bias in Layer conv5_dwara bias in Layer bn5_dw para bias in Layer conv5_sepra bias in Layer bn5_seppara bias in Layer conv6_dwara bias in Layer bn6_dw para bias in Layer conv6_sepra bias in Layer bn6_seppara bias in Layer conv7_dwara bias in Layer bn7_dw para bias in Layer conv7_sepra bias in Layer bn7_seppara bias in Layer conv8_dwara bias in Layer bn8_dw para bias in Layer conv8_sepra bias in Layer bn8_seppara bias in Layer conv9_dwara bias in Layer bn9_dw para bias in Layer conv9_sepra bias in Layer bn9_seppara bias in Layer conv6_3para bias in Layer bn6_3n para bias rnet = 0.6 M, onet = 2.3 M lnet = 11.5 M convert cost: 1.191 ms nms cost: 2.388 ms, (1105-->156) nms cost: 0.555 ms, (470-->74) nms cost: 0.129 ms, (208-->34) nms cost: 0.033 ms, (84-->17) nms cost: 0.008 ms, (24-->7) nms cost: 0.003 ms, (7-->2) nms cost: 0.001 ms, (1-->1) nms cost: 0.001 ms, (1-->1) nms cost: 0.781 ms first stage candidate count: 292 stage 1: cost 6.812 ms run Rnet [175] times, candidate after nms: 7 stage 2: cost 27.096 ms run Onet [0] times, candidate before nms: 0 stage 3: cost 0.002 ms run Lnet [0] times stage 4: cost 0.001 ms final found num: 0 total cost: 35.152 ms (P: 8.007 ms, R: 27.136 ms, O: 0.006 ms, L: 0.004 ms) total 3.706 s / 100 = 37.055 ms
有没有可能是代码中getblobbyname() 的参数跟模型文件里的参数没有对应，我用使用自己训练的模型的时候发现了这个问题，但是没有再关注左博发布的这些模型
我自己转换的模型也出现问题，在调用const ZQ_CNN_Tensor4D* score = pnet[0].GetBlobByName("prob1");这句代码后查看score的值竟然不是小数，里面的值要么是0，要么是1.
@qinxianyuzi @sljlp 生成的zqparams文件是可以手工修改的，json文件最后的分类和回归可能不叫"prob1"和"conv4-2"
@zuoqing1988 我尝试了json里写的名字和您提供模型的名字，好像都不行，您能帮我转下试试吗？我想解决这个问题。我训练的模型应该没有问题，因为我转成ncnn是可以的。https://pan.baidu.com/s/1WreRlXFih5sr030hBk97_A 提取码 eqtf 。万分感谢。