Keras TensorFlow importer: can't upload weights from .h5 file using importKerasNetwork.

Hi, I have a .h5 file with a Keras TensorFlow model that was built using Sequential API. The model is carrying weights, and though Layers are being succesfully uploaded through importKerasNetwork() function, I can't seem to upload the weights with it.

What could I be doing wrong? Is there a way to debug this issue?

I tried this:

test_1=importKerasNetwork('myFile.h5')

And

test_2=importKerasNetwork('myFile.h5', 'WeightFile', 'myFile.h5')

No success whatsover.

Would it be recommendable to have the layers in an JSON file and the weights in a .h5 file?

Thanks in advance for all the help.

[Don Mathis]

What you're doing should work if the file has weights in it. If you post the file here I can investigate.

[Ajpaezm]

Thanks so much for your interest in this Don,

In this comment I have uploaded the .h5 file.

[darci taylor]

Don, I am having the same problem - I can't import my weights. If I use use importKerasNetwork, I get:

"Error using importKerasNetwork (line 93) Unable to import network because some network layers are not yet supported. To import layers and weights, call importKerasLayers with 'ImportWeights' set to true."

If I use "importKerasLayers(modelfile, 'ImportWeights', true)", I am able to see some of my layers, however there are a fair deal of layers that are "placeholder" layers, and I am not able to see the weghts at all. I get the message:

"Warning: Keras network takes vector inputs. Pass images with height=1 and channels=1. Warning: Loss function 'binary_crossentropy' is not yet supported. Warning: Unable to import some Keras layers, because they are not yet supported by the Deep Learning Toolbox. They have been replaced by placeholder layers. To find these layers, call the function findPlaceholderLayers on the returned object. "

Are the weights not importing because it's unable to identify the layers? (I have 12 layers, 8 of them have been returned as placeholders since Matlab doesn't suport 1dCNN , max, or average pooling...)

[Don Mathis]

Darci, if you're using R2018b, you can download the latest version of the keras importer. There was an update in the last month or so. 'binary_crossentropy' is supported now. The placeholder layers should contain the weights, inside the KerasConfiguration field. Unfortunately, Conv1D is not yet supported by the importer.

[José Luis Sandoval]

I have a similar problem:

>> detetor
Warning: File 'resnet50_pascal_cards_inference.h5' was saved in Keras version '2.3.1'. Import of Keras versions newer than
'2.2.4' is not yet supported. The imported model may not exactly match the model saved in the Keras file. 
Error using importKerasNetwork (line 94)
Unable to import network. Weight sharing is not yet supported.

Error in detetor (line 65)
net = importKerasNetwork(modelfile,'OutputLayerType','classification','ClassNames',classNames);

Answer 1

[Don Mathis]

It works for me when I use the latest R2018b update of the tensorflow-keras importer. What version of MATLAB are you using? And do you get an error message?

I get the attached network in MATLAB.

[Ajpaezm]

Yes, is the same outcome I have.

Doing further reading on this article there's no message or notification that tells me I successfully uploaded the weights of the model, right? I thought it Matlab told us (just like it says "SeriesNetwork with properties: Layers: [something] Weights: [Something]".

How could I know if they were indeed uploaded?

Thanks for all the help so far Don and sorry for al these questions.

[Don Mathis]

importKerasNetwork returns the network. So you can look at it:

>> n = importKerasNetwork('OVO_LSTM_model-01.h5')
n =
  SeriesNetwork with properties:
    Layers: [10×1 nnet.cnn.layer.Layer]
    
>> n.Layers
ans = 
  10x1 Layer array with layers:

     1   'SequenceInputLayer'   Sequence Input      Sequence input with 26 dimensions
     2   'lstm_1'               LSTM                LSTM with 128 hidden units
     3   'dense_1'              Fully Connected     128 fully connected layer
     4   'dense_1_relu'         ReLU                ReLU
     5   'dense_2'              Fully Connected     128 fully connected layer
     6   'dense_2_relu'         ReLU                ReLU
     7   'dense_3'              Fully Connected     32 fully connected layer
     8   'dense_3_relu'         ReLU                ReLU
     9   'dense_4'              Fully Connected     4 fully connected layer
    10   'RegressionLayer'      Regression Output   mean-squared-error

>> n.Layers(5)
ans = 
  FullyConnectedLayer with properties:

          Name: 'dense_2'
   Hyperparameters
     InputSize: 128
    OutputSize: 128
   Learnable Parameters
       Weights: [128×128 single]
          Bias: [128×1 single]
  Show all properties

[Ajpaezm]

Thanks so much for clearing this out for me Don! You've helped me tons!

:)

[José Luis Sandoval]

Any solution for this one in R2019b?

[Gopal Rao Netrakanti]

Should we revert to R2018b (doesn't make sense)? I am having the same error in 2019b.

[Daniel Graham]

me too...