Error using dnnfpga.co​mpiler.cod​egenfpga. Layer 'block2b_add' has multiple inputs.

Hello,

I have a pretrained NN imported from TensorFlow. I have imported and replaced the only one placeholder layer and the DAGNetwork seems to be ok. However, when I try to compile the NN with Deep Learning HDL Toolbox, I get the following error:

Error using dnnfpga.compiler.codegenfpga
Layer 'block2b_add' has multiple inputs. Specify which input of layer 'block2b_add' to use.

Error in dnnfpga.apis.Workflow/compileNetwork

Error in dnnfpga.apis.Workflow/compile

Error in generate_HDL_from_NN (line 18)
    hW.compile;

block2b_add has two inputs and connections are defined.

Please find the NN in the LINK.

Could anyone help me to understand the issue?

Thank you in advance

---------------------------

Software version:

Deep Learning HDL Toolbox Support Package For Intel FPGA And SoC Devices 21.2.4

Matlab R2021b Update 6 (9.11.0.2207237)

Ubuntu 20.04

---------------------------

[Aravind Devarakonda]

Hello Ruben

While attempting to load the provided network file, MATLAB encountered a warning :

"Unable to load instances of class EfficienB0_p1_C1.kRescalingLayer246310 into a heterogeneous array. The definition of EfficienB0_p1_C1.kRescalingLayer246310 could be missing or contain an error. Default objects will be substituted."

This warning appears when layer definition is missing or contains an error. To load network, provide definitions of all layers, including custom layers. Custom layer definitions are not included in link to download network file. Provide custom layer definitions along with network file to resolve issue.

Also, refer to the Custom layers documentation page for more information.

Let me know if you have questions. Thank you for your patience.

[Rubén]

Hello @Aravind Devarakonda,

I'm sorry, but that message is not prompted while importing savedModel or NN.mat (at least with my Matlab version).

I have replaced EfficienB0_p1_C1.kRescalingLayer246310 by a scalingLayer - which has the same behaviour. However, I get the same error message while compiling.

Please, find in the following link the new NN: NN_RescalingLayer_replaced.mat

[Aravind Devarakonda]

Hi Ruben

Thank you for replacing the network originally attached.

Regarding your inquiry, I have observed that the network given has a few layers that are not yet supported in the Deep Learning HDL Toolbox and are planned to be integrated in future updates.

These layers include :

1) Scaling Layer

2) Layer Normalization Layer

3) Swish Layer and,

4) FlattenCStyleLayer immediately followed by Convolution2DLayer is also not supported.

Currently, a flatten layer must be followed by either fully connected layer(s) or a join layer (eg. depth concatenation, addition) followed by fully connected layer(s).

As a workaround, I would suggest implementing the Scaling Layer as a custom layer by following the steps available on this documentation link -

https://in.mathworks.com/help/deep-learning-hdl/ug/create-deep-learning-processor-configuration-for-custom-layers.html

Regarding FlattenCStyleLayer immediately followed by Convolution2DLayer, I suggest replacing it with a fully connected layer.

Answer 1

[Rubén]

Hello,

I tried to merge zero padding layer into group convolution layer because issue https://www.mathworks.com/support/bugreports/2661508 according to this comment https://es.mathworks.com/matlabcentral/answers/1947228-error-compiling-resnet50-deep-learning-hdl-toolbox-support-package#comment_2730974 but I get the same error.

Could you confirm this is the sme or a different issue?