Toybrick

标题: Caffe mobilenet-ssd 不支持转rknn [打印本页]

作者: ronyuzhang 时间: 2019-3-29 10:18
标题: Caffe mobilenet-ssd 不支持转rknn
训练了一下 https://github.com/chuanqi305/MobileNet-SSD，但是转rknn模型的时候报错，
E Unsupport input tensor type "None" of layer "output_120".
E Catch exception when building RKNN model!

作者: ronyuzhang 时间: 2019-3-29 10:21
补充错误信息：

W Unhandle status: the input shape of reshape layer mbox_conf_reshape_116 is not 4-D
W Skip blobs of conv0/bn
W Skip blobs of conv0/scale
W Skip blobs of conv1/dw/bn
W Skip blobs of conv1/dw/scale
W Skip blobs of conv1/bn
W Skip blobs of conv1/scale
W Skip blobs of conv2/dw/bn
W Skip blobs of conv2/dw/scale
W Skip blobs of conv2/bn
W Skip blobs of conv2/scale
W Skip blobs of conv3/dw/bn
W Skip blobs of conv3/dw/scale
W Skip blobs of conv3/bn
W Skip blobs of conv3/scale
W Skip blobs of conv4/dw/bn
W Skip blobs of conv4/dw/scale
W Skip blobs of conv4/bn
W Skip blobs of conv4/scale
W Skip blobs of conv5/dw/bn
W Skip blobs of conv5/dw/scale
W Skip blobs of conv5/bn
W Skip blobs of conv5/scale
W Skip blobs of conv6/dw/bn
W Skip blobs of conv6/dw/scale
W Skip blobs of conv6/bn
W Skip blobs of conv6/scale
W Skip blobs of conv7/dw/bn
W Skip blobs of conv7/dw/scale
W Skip blobs of conv7/bn
W Skip blobs of conv7/scale
W Skip blobs of conv8/dw/bn
W Skip blobs of conv8/dw/scale
W Skip blobs of conv8/bn
W Skip blobs of conv8/scale
W Skip blobs of conv9/dw/bn
W Skip blobs of conv9/dw/scale
W Skip blobs of conv9/bn
W Skip blobs of conv9/scale
W Skip blobs of conv10/dw/bn
W Skip blobs of conv10/dw/scale
W Skip blobs of conv10/bn
W Skip blobs of conv10/scale
W Skip blobs of conv11/dw/bn
W Skip blobs of conv11/dw/scale
W Skip blobs of conv11/bn
W Skip blobs of conv11/scale
W Skip blobs of conv12/dw/bn
W Skip blobs of conv12/dw/scale
W Skip blobs of conv12/bn
W Skip blobs of conv12/scale
W Skip blobs of conv13/dw/bn
W Skip blobs of conv13/dw/scale
W Skip blobs of conv13/bn
W Skip blobs of conv13/scale
W Skip blobs of conv14_1/bn
W Skip blobs of conv14_1/scale
W Skip blobs of conv14_2/bn
W Skip blobs of conv14_2/scale
W Skip blobs of conv15_1/bn
W Skip blobs of conv15_1/scale
W Skip blobs of conv15_2/bn
W Skip blobs of conv15_2/scale
W Skip blobs of conv16_1/bn
W Skip blobs of conv16_1/scale
W Skip blobs of conv16_2/bn
W Skip blobs of conv16_2/scale
W Skip blobs of conv17_1/bn
W Skip blobs of conv17_1/scale
W Skip blobs of conv17_2/bn
W Skip blobs of conv17_2/scale
done
--> Building model
E Unsuport input tensor type "None" of layer "output_120".
E Catch exception when building RKNN model!
T Traceback (most recent call last):
T File "rknn/api/rknn_base.py", line 459, in rknn.api.rknn_base.RKNNBase.build
T File "rknn/api/rknn_base.py", line 950, in rknn.api.rknn_base.RKNNBase._quantize
T File "rknn/base/RKNNlib/app/tensorzone/workspace.py", line 219, in rknn.base.RKNNlib.app.tensorzone.workspace.Workspace.load_net
T File "rknn/base/RKNNlib/app/tensorzone/graph.py", line 26, in rknn.base.RKNNlib.app.tensorzone.graph.Graph.load_net
T File "rknn/base/RKNNlib/RKNNnet.py", line 406, in rknn.base.RKNNlib.RKNNnet.RKNNNet.load
T File "rknn/base/RKNNlib/RKNNnet.py", line 410, in rknn.base.RKNNlib.RKNNnet.RKNNNet.load
T File "rknn/base/RKNNlib/RKNNnet.py", line 445, in rknn.base.RKNNlib.RKNNnet.RKNNNet.loads
T File "rknn/base/RKNNlib/layer/RKNNlayer.py", line 150, in rknn.base.RKNNlib.layer.RKNNlayer.RKNNLayer.add_input
T File "rknn/base/RKNNlib/RKNNlog.py", line 105, in rknn.base.RKNNlib.RKNNlog.RKNNLog.e
T ValueError: Unsuport input tensor type "None" of layer "output_120".

作者: hengk 时间: 2019-4-1 16:40
请问解决没有呀

作者: 771488533 时间: 2019-4-3 16:31
我也遇到这个问题，有人帮忙解决吗

作者: RK3399 时间: 2019-4-17 09:06
我也是，目前测试来看并不支持detectoutput层，但是删除该层仍然有问题，模型转换默认会删除prior box层，输出只有两个loc和conf，看他的例子应该是prior box需要用caffe预先运行出来，然后结合loc和conf自己去算。

作者: RK3399 时间: 2019-4-17 09:10
参考该贴 http://t.rock-chips.com/forum.ph ... &extra=page%3D3

作者: RK3399 时间: 2019-4-18 10:48
我解决了，哈哈，坑太多了
RK3399 Pro 的确不支持detectoutput层，需要删除，模型转换默认也会删除prior box层，输出只有两个loc和conf，prior box需要需要先用caffe预先运行出来，然后结合loc和conf自己去算。

作者: ronyuzhang 时间: 2019-4-18 15:32

RK3399 发表于 2019-4-18 10:48
我解决了，哈哈，坑太多了
RK3399 Pro 的确不支持detectoutput层，需要删除，模型转换默认也会删除prior bo ...

来个教程啊。

作者: RK3399 时间: 2019-4-18 17:04
有空再写

作者: q5671229 时间: 2019-4-19 11:42

RK3399 发表于 2019-4-18 17:04
有空再写

写个教程吧。。git上贡献一下源码？

搞不定

作者: 771488533 时间: 2019-4-19 13:52

RK3399 发表于 2019-4-18 10:48
我解决了，哈哈，坑太多了
RK3399 Pro 的确不支持detectoutput层，需要删除，模型转换默认也会删除prior bo ...

把 detectoutput层删除了那还要这模型干嘛。。。

作者: RK3399 时间: 2019-4-19 14:31

771488533 发表于 2019-4-19 13:52
把 detectoutput层删除了那还要这模型干嘛。。。

detectoutput 删除了，但是可以根据loc, conf , priorbox的运行结果，自己写detectoutput 然后输出正确检测

作者: RK3399 时间: 2019-4-19 15:16

q5671229 发表于 2019-4-19 11:42
写个教程吧。。git上贡献一下源码？搞不定

可以关注我的仓库
https://github.com/Pinnh/NPU_CaffeSSD

作者: q5671229 时间: 2019-4-19 17:11

RK3399 发表于 2019-4-19 15:16
可以关注我的仓库
https://github.com/Pinnh/NPU_CaffeSSD

谢了，已经抄了

作者: RK3399 时间: 2019-4-19 18:21

q5671229 发表于 2019-4-19 17:11
谢了，已经抄了

记得star哦

作者: lipandeng 时间: 2019-12-26 16:19

RK3399 发表于 2019-4-19 15:16
可以关注我的仓库
https://github.com/Pinnh/NPU_CaffeSSD

你好，按照这个是可以跑通的，推动过程是比较快的但在结合priorbox_flatten计算的时候很慢，是推动的3倍左右，请问有什么好的办法没有

作者: RK用户 时间: 2019-12-26 17:04

lipandeng 发表于 2019-12-26 16:19
你好，按照这个是可以跑通的，推动过程是比较快的但在结合priorbox_flatten计算的时候很慢，是推动的3倍 ...

你的是多少秒，我在板子上跑一帧的推理要130ms ，获取结果要将近4秒

作者: ylc123 时间: 2020-6-2 10:29

RK3399 发表于 2019-4-18 17:04
有空再写

大佬，写个教程呗。搞不定啊

欢迎光临 Toybrick (https://t.rock-chips.com/)