Cannot convert a symbolic Keras input/output to a numpy array with TF Lite converter

[James B.]

I need to implement a custom convolution 2D layer which is not supported by Tensorflow and TF Lite, so I tried to define it by using this link: https://www.tensorflow.org/guide/create_op (in order to have a TF operator for the op) and this link: https://www.tensorflow.org/lite/guide/ops_custom (to have a TF Lite operator for this op). I can train the model; however, when I try to convert the model with the operator with TF Lite converter, I get this error:

TypeError: Cannot convert a symbolic Keras input/output to a numpy array. This error may indicate that you're trying to pass a symbolic value to a NumPy call, which is not supported. Or, you may be trying to pass Keras symbolic inputs/outputs to a TF API that does not register dispatching, preventing Keras from automatically converting the API call to a lambda layer in the Functional Model.

The example code is like this:

import tensorflow as tf

tf.config.run_functions_eagerly(True)

from keras.datasets import mnist

from keras.models import Model

from keras.layers import add,Input,Activation,Flatten,Dense

def convol(inp):

  conv_module = tf.load_op_library('./conv.so')

  x = conv_module.conv(inp, name="Conv")

  return x

def read_mnist(path):

    (train_x,train_y), (test_x,test_y)=mnist.load_data()

    return train_x,train_y,test_x,test_y

def tcn(train_x,train_y,test_x,test_y):

    inp=Input(shape=(28,28))

    x = convol(inp)

    x=Flatten()(x)

    x=Dense(10,activation='softmax')(x)

    model=Model(inputs=inp,outputs=x)

    model.summary()

    model.compile(loss='sparse_categorical_crossentropy', optimizer='adam', metrics=     

    ['accuracy'])

    model.fit(train_x,train_y,batch_size=100,epochs=10,validation_data=(test_x,test_y))       

    pred=model.evaluate(test_x,test_y,batch_size=100)

    print('test_loss:',pred[0],'- test_acc:',pred[1])

    train_x,train_y,test_x,test_y=read_mnist('MNIST_data')

    tcn(train_x,train_y,test_x,test_y)

    tflite_model_name = 'net'

    inp=Input(shape=(28,28))

    converter =    tf.lite.TFLiteConverter.from_concrete_functions([tf.function(convol).get_concrete_function(inp)])

    converter.allow_custom_ops = True

    tflite_model = converter.convert()

    open(tflite_model_name + '.tflite', 'wb').write(tflite_model)

I think the problem is due to the fact that TF Lite converter receives a Keras Tensor (a placeholder) rather than an array numpy, but I'm not sure how I could convert the model if I have to use a TF Lite custom op.

Thanks in advance.

[Sachin Joglekar]

Jaesung/Karim could you take a look?

[Jaesung Chung]

To workaround Keras/Num py issues, instead of the concrete function converter API, is it possible to export the model as a saved model and try the TFLiteConverter.from_saved_model API instead?

[Cristina Augello]

Thanks for the help. I have tried to export the model as a saved model: now TFLite converter says that the op isn't supported (as expected). I'll try to register it.