cancel
Showing results for 
Search instead for 
Did you mean: 
Highlighted
Participant
Participant
335 Views
Registered: ‎04-06-2018

[DNNC] Error while compiling a DNN with dilated convolution layers

Hi,

While compiling my DNN, I get the following failed check:

[BACKEND][Check Failed: dilation <= input_height][/tmp/DNNDK_Pipeline_dnnc/dnnc/submodules/asicv2com/src/Operator/OperatorConv.cpp:77][DATA_OUTRANGE][Data value is out of range!]

Indeed, my DNN contains multiple dilated convolutions with a dilation rate superior to the input_height. However, the dilated convolution is performed only on the width dimension. The DNNDK User Guide states for convolution: "Support kernel-w and kernel-h values ranging from 1 to 8 in any combination."

So, is it indended or is it a bug?

Thank you,

Neutrinoxy

0 Kudos
5 Replies
Highlighted
Xilinx Employee
Xilinx Employee
301 Views
Registered: ‎03-27-2013

Re: [DNNC] Error while compiling a DNN with dilated convolution layers

Hi @neutrinoxy ,

 

I can record the message and try to contact our dev team in next week's meeting.

But if you can help me to reproduce the issue on a simple project it would be more efficient to locate the root cause.

And I have tried to write a simple example about deploy custom model to MPSoC platform.

Is it possible that you can help to modify the model here to add your layer and configurations : https://github.com/gewuek/flower_classification_dnndk_v2/blob/master/x86/train_model.py

and send it back to me?

I can try to reproduce the issue on my side.

Best Regards,
Jason
-----------------------------------------------------------------------------------------------
Please mark the Answer as "Accept as solution" if the information provided is helpful.

Give Kudos to a post which you think is helpful and reply oriented.
-----------------------------------------------------------------------------------------------
0 Kudos
Highlighted
Participant
Participant
245 Views
Registered: ‎04-06-2018

Re: [DNNC] Error while compiling a DNN with dilated convolution layers

Hi @jasonwu,

In order to reproduce the issue, you should only need to add a conv layer like this:

keras.layers.Conv2D(32, (1,5), padding="same", activation='relu', dilation_rate=2)

With an input with the shape (None, 1, Y, Z).

DNNC will then check that the height (1) is less than the dilation rate (2 or more), and trigger an error.

I hope you'll be able to raise this problem during your next meeting.

Best regards,

Neutrinoxy

0 Kudos
Highlighted
Xilinx Employee
Xilinx Employee
214 Views
Registered: ‎03-27-2013

Re: [DNNC] Error while compiling a DNN with dilated convolution layers

Hi @neutrinoxy ,

 

I can reproduce this issue with the infromation you provided.

[BACKEND][Check Failed: dilation <= input_height][/tmp/DNNDK_Pipeline_dnnc/dnnc/submodules/asicv2com/src/Operator/OperatorConv.cpp:77][DATA_OUTRANGE][Data value is out of range!] 

Now my Convnet looks a little wild here:

_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d (Conv2D) (None, 128, 128, 32) 896
_________________________________________________________________
max_pooling2d (MaxPooling2D) (None, 16, 64, 32) 0
_________________________________________________________________
conv2d_1 (Conv2D) (None, 16, 64, 64) 18496
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 4, 32, 64) 0
_________________________________________________________________
conv2d_2 (Conv2D) (None, 4, 32, 64) 36928
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 1, 16, 64) 0
_________________________________________________________________
conv2d_3 (Conv2D) (None, 1, 16, 128) 24704
_________________________________________________________________
flatten (Flatten) (None, 2048) 0
_________________________________________________________________
dense (Dense) (None, 100) 204900
_________________________________________________________________
dense_1 (Dense) (None, 5) 505
=================================================================

I would try to check with our dev team if this is an expected behavor or just a unnecessary barrier.

But that may take some time.

Best Regards,
Jason
-----------------------------------------------------------------------------------------------
Please mark the Answer as "Accept as solution" if the information provided is helpful.

Give Kudos to a post which you think is helpful and reply oriented.
-----------------------------------------------------------------------------------------------
0 Kudos
Participant
Participant
163 Views
Registered: ‎04-06-2018

Re: [DNNC] Error while compiling a DNN with dilated convolution layers

Hi @jasonwu,

Do you have an update on this? It would be nice to know if the issue comes from the check or if the framework doesn't support non-square dilated convolutions. I'm stuck right now and I cannot move forward because of the source code being unavailable.

Thank you,

Neutrinoxy

0 Kudos
Highlighted
Xilinx Employee
Xilinx Employee
149 Views
Registered: ‎03-27-2013

Re: [DNNC] Error while compiling a DNN with dilated convolution layers

Hi @neutrinoxy ,

 

I do report the issue to the contact of the dev team as soon as I can reproduce the issue on my side. But I am afraid that they ar working on the new release of Vitis AI. So the response from them would be delayed.

Best Regards,
Jason
-----------------------------------------------------------------------------------------------
Please mark the Answer as "Accept as solution" if the information provided is helpful.

Give Kudos to a post which you think is helpful and reply oriented.
-----------------------------------------------------------------------------------------------
0 Kudos