Problem Description

I am trying to write a custom loss function in TensorFlow 2.3.0. To calculate the loss, I need the y_pred parameter to be converted to a numpy array. However, I can't find a way to convert it from <class 'tensorflow.python.framework.ops.Tensor'> to numpy array, even though there seem to TensorFlow functions to do so.

Code Example

def custom_loss(y_true, y_pred):
    print(type(y_pred))
    npa = y_pred.make_ndarray()
    ...
    

if __name__ == '__main__':
    ...
    model.compile(loss=custom_loss, optimizer="adam")
    model.fit(x=train_data, y=train_data, epochs=10)

gives the error message: AttributeError: 'Tensor' object has no attribute 'make_ndarray after printing the type of the y_pred parameter: <class 'tensorflow.python.framework.ops.Tensor'>

What I have tried so far

Looking for a solution I found this seems to be a common issue and there a couple of suggestions, but they did not work for me so far:

1. " ... so just call .numpy() on the Tensor object.": How can I convert a tensor into a numpy array in TensorFlow?

so I tried:

def custom_loss(y_true, y_pred):
    npa = y_pred.numpy()
    ...

giving me AttributeError: 'Tensor' object has no attribute 'numpy'

2. "Use tensorflow.Tensor.eval() to convert a tensor to an array": How to convert a TensorFlow tensor to a NumPy array in Python

so I tried:

def custom_loss(y_true, y_pred):
    npa = y_pred.eval(session=tf.compat.v1.Session())
    ...

giving me one of the longest trace of error messages I ever have seen with the core being:

InvalidArgumentError: 2 root error(s) found.
      (0) Invalid argument: You must feed a value for placeholder tensor 'functional_1/conv2d_2/BiasAdd/ReadVariableOp/resource' with dtype resource
         [[node functional_1/conv2d_2/BiasAdd/ReadVariableOp/resource (defined at main.py:303) ]]
         [[functional_1/cropping2d/strided_slice/_1]]
      (1) Invalid argument: You must feed a value for placeholder tensor 'functional_1/conv2d_2/BiasAdd/ReadVariableOp/resource' with dtype resource
         [[node functional_1/conv2d_2/BiasAdd/ReadVariableOp/resource (defined at main.py:303) ]]

also having to call TensorFlow Compatibility Functions from Version 1.x does not feel very future-proof, so I do not like this approach too much anyhow.

3. Looking at the TensorFlow Docs there seemed to be the function I needed just waiting: tf.make_ndarray Create a numpy ndarray from a tensor.

so I tried:

def custom_loss(y_true, y_pred):
    npa = tf.make_ndarray(y_pred)
    ...

giving me AttributeError: 'Tensor' object has no attribute 'tensor_shape'

Looking at the example in the TF documentation they use this on a proto_tensor, so I tried converting to a proto first:

def custom_loss(y_true, y_pred):
    proto_tensor = tf.make_tensor_proto(y_pred)
    npa = tf.make_ndarray(proto_tensor)
    ...

but already the tf.make_tensor_proto(y_pred) raises the error: TypeError: Expected any non-tensor type, got a tensor instead.

Also trying to make a const tensor first gives the same error:

def custom_loss(y_true, y_pred):
    a = tf.constant(y_pred)
    proto_tensor = tf.make_tensor_proto(a)
    npa = tf.make_ndarray(proto_tensor)
    ...

There are many more posts around this but it seems they are all coming back to these three basic ideas. Looking forward to your suggestions!