What to set 'ClassNames' to when using importKerasNetwork() to import network and weights for a regression model?

Hi,

I have trained a regression model in Keras/Tensorflow and exported the network achitecure and weights to .json and .h5 files. I would like to import my regression model to Matlab (R2018a) and looks like importKerasNetwork() does just that. However, it keeps throwing the following error: "Reference to non-existent field 'class_name'." I don't pass any 'Classes' (R2018b) or 'ClassNames' (R2018a) becasue I'm doing regression not classification. This is the how I call the function:

net = importKerasNetwork('model_architecture.json','WeightFile','my_model_weights.h5', 'OutputLayerType','regression');

The documentation (R2018b as well as R2018a) doesn't state what to pass in 'Classes' when you are dealing with a regression problem. I tried passing an empty array but got the same error. Any ideas how to solve this problem?

Thanks!

Answer 1

[Don Mathis]

There were a few updates to that support package after the initial release. Try downloading it and installing it again through the Add-Ons icon in MATLAB. Also if you can upload your model files here I can take a look.

[MM]

Installing the updates did the trick. Thanks Don!

[MM]

Hi again,

So things were working fine until today when the same error started to pop up again. I installed the most recent updates (same file as the one I installed back in December). Unfortunately, it didn't resolve the issue this time. I'm still getting:

Warning: Model file already specifies the output layer type. 'OutputLayerType' will
be ignored. 
> In nnet.internal.cnn.keras.importKerasNetwork>iUpdateTrainingConfig (line 60)
  In nnet.internal.cnn.keras.importKerasNetwork (line 22)
  In importKerasNetwork (line 84) 
Error using importKerasNetwork (line 86)
Reference to non-existent field 'class_name'.

Any idea, what could have gone wrong?

[Don Mathis]

I would need to see the network. Are you still using R2018a?

[MM]

Thank you Don for the prompt response! Yes I'm still using R2018a. When you say 'see the network', do you mean the .h5 files?

I attached the .h5 files.

[Don Mathis]

That keras network utilizes a "breaking change" that the Keras developers made to their file format in October 2018, and which unfortunately is not supported by the R2018a version of the Keras importer. The R2018b importer can import it, and I've attached the imported network here in case it is any help to you.

[MM]

Thank you so much Don! Having the network.mat file will definetly get me going until I upgrade to R2018b.

Thanks again!

[MM]

Just had R2018b installed and re-installed the Keras add-on. Still no luck. Same error:

Error using importKerasNetwork (line 93)

Reference to non-existent field 'class_name'

Any idea what could be wrong?

[Don Mathis]

I just checked and support for this Keras change will be available in the February update of the Keras importer support package. That update is scheduled to go out by the end of Feb (tomorrow!)

[MM]

Wonderful. Please let me know when the update is ready.

Thanks!

[Don Mathis]

Try downloading it now.

[MM]

It worked! Thanks a lot Don!

[Yousaf Rahman]

I am having the same issue, I am using the latest version of the importer (downloaded 4/2/2019), but I am using MATLAB 2017b. I had trouble installing the importer in 2018b, and have a Mathworks Support case ongoing for that.

[MohamedElHefnawy]

I am having the same issue. I am using R2018a and installed the latest version of the package (March 13 2020).

cnn_model_demo = importKerasNetwork('model.json','WeightFile','cnn_model.h5', 'OutputLayerType','classification')

Error using importKerasNetwork (line 86)

Reference to non-existent field 'class_name'.

Answer 2

[Sriramkrishnan Muralikrishnan]

Hi,

I have the same problem with importing

net = importKerasNetwork(modelfile,'OutputLayerType','regression')

Error using importKerasNetwork (line 77)

Reference to non-existent field 'class_name'

I am using R2017b and just installed the add on for Keras. Could you please let me know where I can find the update which might resolve this problem.

Thanks

[MM]

Click on the little drop-down arrow on the Add-Ons icon --> Check for updates --> Add-on fetaures. This usually worked for me.

[MohamedElHefnawy]

I am having the same issue. I am using R2018a and installed the latest version of the package (March 13 2020).

cnn_model_demo = importKerasNetwork('model.json','WeightFile','cnn_model.h5', 'OutputLayerType','classification')

Error using importKerasNetwork (line 86)

Reference to non-existent field 'class_name'.

Answer 3

[Ben Wetherill]

Hello. I'm also having this issue. I recently downloaded the tool for 2017b via the offline installation method. After some digging and use of HDFViewer I can see the issue.

My NN is built as follows:

...
model = Sequential()
model.add(Dense(200, input_shape=(input_size,), activation='relu'))
model.add(Dense(160, activation='relu'))
model.add(Dense(output_size, activation='sigmoid'))

model.compile(loss='binary_crossentropy', optimizer='rmsprop', metrics=['accuracy'])

history = model.fit(input_data_train, output_data_train, validation_data = (input_data_test,output_data_test),
                    epochs=100, batch_size=64)
                    
output_data_pred = model.predict(input_data_test)

model.save('unvalidated_input_data.h5')

Calling importKerasLayers() from the Matlab prompt leads to the following line in KerasLayerInsideSequentialModel.m throwing the error.

this.ClassName = Struct.class_name;

When I look at the contents of the struct at this point, I'm dealing with an object called 'name' with a value of 'sequential_9'. This was easy to find in HDFViewer. See below for the object dump.

{"class_name": "Sequential", "config": {"name": "sequential_9", "layers": [{"class_name": "Dense", "config": {"name": "dense_24", "trainable": true, "batch_input_shape": [null, 248], "dtype": "float32", "units": 62, "activation": "relu", "use_bias": true, "kernel_initializer": {"class_name": "VarianceScaling", "config": {"scale": 1.0, "mode": "fan_avg", "distribution": "uniform", "seed": null}}, "bias_initializer": {"class_name": "Zeros", "config": {}}, "kernel_regularizer": null, "bias_regularizer": null, "activity_regularizer": null, "kernel_constraint": null, "bias_constraint": null}}, {"class_name": "Dense", "config": {"name": "dense_25", "trainable": true, "units": 108, "activation": "sigmoid", "use_bias": true, "kernel_initializer": {"class_name": "VarianceScaling", "config": {"scale": 1.0, "mode": "fan_avg", "distribution": "uniform", "seed": null}}, "bias_initializer": {"class_name": "Zeros", "config": {}}, "kernel_regularizer": null, "bias_regularizer": null, "activity_regularizer": null, "kernel_constraint": null, "bias_constraint": null}}]}}

So the cause is clear, but I don't know if its my version of the importer or the way I've setup the model in Keras. From this thread others have had a similar issue but have fixed it with a tool update. My version should include this update. I'm using Keras version 2.2.4.

Any thoughts?

[Don Mathis]

I'm pretty sure this is due to a breaking change made by Keras in version 2.2.3 that changed their representation of Sequential models. That change was made in late 2018. The earliest Keras importer version that supports the breaking change was R2018b. You could try to save the Keras model in a version earlier than 2.2.3.

[Ben Wetherill]

Yup, that fixed it. Well, I say fixed it - now I have an issue in that the binary_crossentropy loss function isn't implemented in the 2017b version of the importer and my IT department hasn't released 2018b yet. :/ Yay for Mordac, The Preventer of Information Services. ;)