Intel® Distribution of OpenVINO™ Toolkit
Community assistance about the Intel® Distribution of OpenVINO™ toolkit, OpenCV, and all aspects of computer vision-related on Intel® platforms.
6392 Discussions

mvNCCompile error code 5: Toolkit Error: Stage Details Not Supported: Pow

idata
Employee
1,025 Views

I am trying to compile a Tensorflow graph for movidius and getting this error, are Pow operations not supported by the sdk?

 

It seems like this is NOT supported. ANy help will be appreciated!
0 Kudos
13 Replies
idata
Employee
730 Views

@leninpatra I don't see Pow listed in TensorFlowParser.py in the NCSDK Controllers folder. Can you provide a link to this model?

0 Kudos
idata
Employee
730 Views

Hi Tome-

 

We have done some more work on Movidius and we found that following operations are missing-

 

These are our requests - we want Movidius to support -

 

1) m x n convolutions:

 

So looks like- it only supports square convolution operation, i.e, conv2d( x, [n,n], …), where n is the width and height of an image for instance. But we have non-square data and we want to apply the conv2d on it but Movidius doesn't work with that.

 

2) Bug: basic operations (like scalar multiplication) don't work on the input tensor ! Seems like it is crashing.

 

3) tf.nn.moments

 

4) subtraction

 

Can you please help?

 

Thanks

 

Lenin
0 Kudos
idata
Employee
730 Views

@leninpatra, nice to see you on the forum. You are in good hands with @Tome_at_Intel!

0 Kudos
idata
Employee
730 Views

@AshwinVijayakumar Thanks. @leninpatra Which version of the NCSDK are you using?

 

     

  • The NCSDK does indeed have support for non-square convolutions; i.e. the Inception model features non-square convolutions (1x3, 3x1, 1x7, 7x1, etc).
  •  

  • tf.nn.moments is not supported as of NCSDK 2.05. This feature hasn't been implemented yet and I cannot provide a roadmap at the moment.
  •  

  • Multiplication is only implemented for Tensors at the moment. If any scaling needs to be performed on the input, it should be performed outside of the model in the deployment application.
  •  

  • Looking at the NCSDK 2.05 TensorFlow parser, Subtraction seems to be implemented.
  •  

 

Hope this helps.

0 Kudos
idata
Employee
730 Views

Hi Tome-

 

Basically i have a 128x1 convolution but when i ran it on the movidius, it interpreted it as a 128x128 and that caused the resulting dimension to underflow..and it is really crashing..

 

--------This is the error we are seeing---------------

 

File "/usr/local/lib/python3.5/dist-packages/tensorflow/python/framework/tensor_shape.py", line 39, in init

 

raise ValueError("Dimension %d must be >= 0" % self._value)

 

ValueError: Dimension -7 must be >= 0

 

So- can current movidius SDK support long 1d // or long slim convolutions ~ of the like of 128x1 or 128x4, etc etc?

 

Thanks for the help-

 

lenin
0 Kudos
idata
Employee
730 Views

Hi Tome-

 

Any updates on my request?

 

Thanks

 

Lenin
0 Kudos
idata
Employee
730 Views

@leninpatra We definitely support some 1d convolutions as seen in the inception networks. I don't know if we support convolutions as large as 128x1 or 128x4. If you would can provide the network, I'd like to take a look at your issue to pinpoint the issue.

 

Additionally how are you determining that the convolution is being interpreted as 128x128? Thanks!

0 Kudos
idata
Employee
730 Views

Hi Tome-

 

It seems like as long as the convolution is < 120x1, Then Movidius can support it.

 

can you please reach out to your internal ML SW development team and see movidius can support 128x1 convolution?

 

If you look at the error dimension-7: essentially Movidius SW framework is converting the 128x1 to a square convolution. SInce it is large, it is crashing..

 

Appreciate the help to resolve this problem!

 

-Lenin
0 Kudos
idata
Employee
730 Views

@leninpatra Thanks for reporting this. I'm trying to re-create what you are seeing, so that we can examine the issue on our setup. What NCSDK version are you using and if possible, can you include a link to the model you're using? Thanks.

0 Kudos
idata
Employee
730 Views

Tome- we are using NCSDK 2.05 version.

 

Can i send you the model privately to your Intel email address?

 

Thanks

 

Lenin
0 Kudos
idata
Employee
730 Views

Tome

 

I have created a model for you-

 

this network is supposed to do a 128x1 convolution on a 512x32 input, resulting in a 385x32 output

 

however, the movidius model crashes because it tries to do a 128x128 convolution on the 512x32 input, resulting in a 385x(-95) output.

 

-Lenin

0 Kudos
idata
Employee
730 Views

@leninpatra You can send me a private message and include the link to the model. Thanks.

0 Kudos
idata
Employee
730 Views

@Tome_at_Intel - Pinged you privately with the model!

0 Kudos
Reply