Yolov3

idata · ‎07-12-2018

I try to convert yolov3-tiny to caffe. The prototxt like https://github.com/Amalle/Yolov3_caffe/blob/master/models/caffe/yolov3-tiny.prototxt

And I convert the graph using the following:


$ mvNCCompile models/caffe/yolov3-tiny.prototxt -w models/caffe/yolov3-tiny.caffemodel -s 12 -ec
mvNCCompile v02.00, Copyright @ Intel Corporation 2017
/usr/local/bin/ncsdk/Controllers/FileIO.py:65: UserWarning: You are using a large type. Consider reducing your data sizes for best performance
Blob generated

But the predicted result size was wrong.

Then I try mvNCProfile:


$ mvNCProfile models/caffe/yolov3-tiny.prototxt -w models/caffe/yolov3-tiny.caffemodel -ec
mvNCProfile v02.00, Copyright @ Intel Corporation 2017
/usr/local/bin/ncsdk/Controllers/FileIO.py:65: UserWarning: You are using a large type. Consider reducing your data sizes for best performance
Blob generated
USB: Transferring Data...
/usr/local/lib/python3.5/dist-packages/mvnc/mvncapi.py:420: DeprecationWarning: The binary mode of fromstring is deprecated, as it behaves surprisingly on unicode inputs. Use frombuffer instead
Time to Execute :  266.8  ms
USB: Myriad Execution Finished
Time to Execute :  248.16  ms
USB: Myriad Execution Finished
Traceback (most recent call last):
  File "/usr/local/bin/mvNCProfile", line 156, in <module>
    profile_net(args.network, args.inputnode, args.outputnode, args.nshaves, args.inputsize, args.weights, args.device_no, args.explicit_concat, args.ma2480, args.scheduler, args)
  File "/usr/local/bin/mvNCProfile", line 142, in profile_net
    timings, myriad_output = run_myriad(graph_file, args)
  File "/usr/local/bin/ncsdk/Controllers/MiscIO.py", line 248, in run_myriad
    myriad_output = myriad_output.reshape(sz[1], sz[2], sz[3])
ValueError: cannot reshape array of size 41308 into shape (1,1,215475)

@Tome_at_Intel

idata · ‎07-13-2018

@maqiao Looks like this Tiny Yolo v3 uses a concat as the last layer. There is a known issue where the NCSDK has issues with models having concat as the last layer. Some users have found work arounds by using dummy reshape layers https://ncsforum.movidius.com/discussion/comment/2831/#Comment_2831. Try this and let me know if this works for you.

idata · ‎07-16-2018

@Tome_at_Intel Thanks

I have try that. It can't works for me.

The last three layers as following:


layer {
  name: "conv13_perm"
  type: "Permute"
  bottom: "conv13"
  top: "conv13_perm"
  permute_param {
    order: 0
    order: 2
    order: 3
    order: 1
  }
}
layer {
  name: "conv13_flat"
  type: "Flatten"
  bottom: "conv13_perm"
  top: "conv13_flat"
  flatten_param {
    axis: 1
  }
}
layer {
  name: "output"
  type: "Concat"
  bottom: "conv13_flat"
  bottom: "conv10_flat"
  top: "output"
  concat_param {
    axis: 1
  }
}

Try this:

`

$ mvNCProfile models/caffe/yolov3-tiny.prototxt -w models/caffe/yolov3-tiny.caffemodel -s 12 -on conv13_perm -ec

mvNCProfile v02.00, Copyright @ Intel Corporation 2017

/usr/local/bin/ncsdk/Controllers/FileIO.py:65: UserWarning: You are using a large type. Consider reducing your data sizes for best performance

Blob generated

USB: Transferring Data…

Time to Execute : 216.87 ms

USB: Myriad Execution Finished

Time to Execute : 197.8 ms

USB: Myriad Execution Finished

USB: Myriad Connection Closing.

USB: Myriad Connection Closed.

Network Summary

Detailed Per Layer Profile

Bandwidth time

Name MFLOPs (MB/s) (ms)

0 data 0.0176147.1 0.006

1 scale1 149.5 1175.5 7.582

2 pool1 2.8 792.4 6.666

3 scale2 398.7 859.5 13.839

4 pool2 1.4 926.3 2.851

5 scale3 398.7 648.0 9.226

6 pool3 0.7 957.8 1.379

7 scale4 398.7 403.3 7.717

8 pool4 0.3 947.6 0.697

9 scale5 398.7 203.8 10.051

10 pool5 0.2 924.2 0.357

11 scale6 398.7 323.6 9.250

12 pool6 0.4 411.4 0.401

13 crop6 0.0 2180.1 0.088

14 scale7 1594.9 314.8 33.309

15 scale8 88.6 331.9 2.501

16 scale9 398.7 323.5 9.252

17 conv10 44.1 311.0 1.332

18 scale11 11.1 441.6 0.328

19 scale11_s0 0.0 33.8 0.010

20 upsample_p0 0.0 156.4 0.264

21 scale11_s1 0.0 38.7 0.009

……

273 scale11_s127 0.0 38.2 0.009

274 upsample_p127 0.0 158.2 0.261

275 pool_deconv 0.1 566.3 0.292

276 route 0.0 1950.4 0.085

277 route_scale5 0.0 2094.7 0.158

278 scale12 1196.2 210.3 29.215

279 conv13 88.3 197.3 2.305

280 conv10_perm 0.0 310.1 0.265

281 conv10_flat 0.0 179.6 0.458

282 conv13_perm 0.0 182.7 1.800

Total inference time 186.13

Generating Profile Report 'output_report.html'…

`

But, the following was wrong:

`

$ mvNCProfile models/caffe/yolov3-tiny.prototxt -w models/caffe/yolov3-tiny.caffemodel -s 12 -on conv13_flat -ec

/usr/local/bin/ncsdk/Controllers/FileIO.py:65: UserWarning: You are using a large type. Consider reducing your data sizes for best performance

Blob generated

USB: Transferring Data…

Segmentation fault (core dumped)

`

idata · ‎01-02-2019

I trying to use yolov3-tiny model for Movidius but i checked release note for ncsdk v2.08.01 yolov3 is not being enlisted in the list of supported models.

Does it means i cannot use yolo v3 model with movidius stick with new ncsdk?

I tested every possible way in which i can use yolov3 tiny model with movidius stick i was able to get the prototxt file and caffemodel as well.But it don't work.

Is there any possible way i can use the yolov3-tiny model with movidius stick?

idata · ‎01-02-2019

@h.vard

https://github.com/PINTO0309/OpenVINO-YoloV3