nnet.keras​.layer.Fla​ttenCStyle​Layer is not supported

Hello,

I've imported a NN in SaveModel fromat from TensorFlow (v2.6). It has a Keras flatten layer and when I try to generate the HDL with Deep Learning HDL Toolbox Support Package For Intel FPGA And SoC Devices I get the following message:

Error using dnnfpga.compiler.codegenfpga
Layer nnet.keras.layer.FlattenCStyleLayer is not supported. See the documentation for list of supported layers.

Error in dnnfpga.apis.Workflow/compileNetwork

Error in dnnfpga.apis.Workflow/compile

Error in generate_HDL_from_NN (line 19)
dn = hW.compile;

However, nnet.keras.layer.FlattenCStyleLayer is supported according to documentation (https://es.mathworks.com/help/deep-learning-hdl/ug/supported-networks-layers-boards-and-tools.html)

I realised that nnet.keras.layer.FlattenCStyleLayer must be followed by a Fully connected layer and it does.

These are the layers from the NN imported:

    nn.Layers =
    
    7×1 Layer array with layers:
    
     1   'input_layer'                        Image Input             28×28×1 images
     2   'flatten'                            Keras Flatten           Flatten activations into 1-D assuming C-style (row-major) order
     3   'dense'                              Fully Connected         128 fully connected layer
     4   'dense_relu'                         ReLU                    ReLU
     5   'output_layer'                       Fully Connected         10 fully connected layer
     6   'output_layer_softmax'               Softmax                 softmax
     7   'ClassificationLayer_output_layer'   Classification Output   crossentropyex with '0' and 9 other classes

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

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

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

Answer 1

[Wang Chen]

Hi Ruben,

Is it possible for you to upgrade to R2022a or higher version of MATLAB?

This issue is fixed in R2022a, please see following bug report: https://www.mathworks.com/support/bugreports/2522278?s_tid=srchtitle_nnet.keras.layer.FlattenCStyleLayer_2

Alternatively, In R2021b, If your network is a SeriesNetwork type, the nnet.keras.layer.FlattenCStyleLayer is also suppored.

Thanks,

Wang

[Rubén]

Hello @Wang Chen,

I was able to generate HDL and configure FPGA importing the NN as a SeriesNetwork according to your comment. Layers have been imported and SeriesNetwork NN has been assembled with them.

Thank you for your help.

Are non-SeriesNetworks, like DAGNetworks supported by Deep Learning toolboxes (Deep Learning HDL Toolbox Support Package as well)?

[Wang Chen]

Hi Ruben,

DAGNetworks is also supported by Deep Learning HDL Toolbox in general, but when a DAGNetwork contains "nnet.keras.layer.FlattenCStyleLayer" layer, please use R2022a or later version of the Deep Learning HDL Toolbox.

Thanks,

Wang

[Rubén]

Understood. Thank you for your support

The question comes because in deploy documentation it is only mentioned that series network can be deployed: "The Deep Learning HDL Toolbox™ Support Package for Intel FPGA and SoC Devices enables you to deploy your custom series deep learning network, and custom deep learning processor IP on the hardware."