convert tf.tensor to numpy

Question

I've built a custom loss function to train my model

def JSD_Tensor_loss(P,Q):
  P=tf.make_ndarray(P)
  Q=tf.make_ndarray(Q)

  M=np.divide((np.sum(P,Q)),2)
  D1=np.multiply(P,(np.log(P,M)))
  D2=np.multiply(Q,(np.log(Q,M)))

  JSD=np.divide((np.sum(D1,D2)),2)
  JSD=np.sum(JSD)
  return JSD

model.compile(optimizer='adam', loss=JSD_Tensor_loss, metrics=['accuracy'])
model.fit(x_train, x_test, epochs=EPOCHS)

Although I've converted the tensor params to numpy but I've having the following error and cant solve it AttributeError: 'Tensor' object has no attribute 'tensor_shape'

Answer 1

Your error stems from your conversion to numpy.array (first two lines).

If you have to convert a tensor (which in my opinion you shouldn't in this case), I would go with:

P, Q = P.numpy(), Q.numpy()

However, this is really not necessary here. Just replace the numpy function with the respective tf.math function (docs) and you should be good:

def JSD_Tensor_loss(P,Q):

    M=tf.math.divide((tf.math.add(P,Q)),2)
    D1=tf.math.multiply(P,(tf.math.log(P)))
    D2=tf.math.multiply(Q,(tf.math.log(Q)))

    JSD=tf.math.divide((tf.add(D1,D2)),2)
    JSD=tf.math.reduce_sum(JSD)
    return JSD

(I did not understand your call to np.log. Do you maybe use an outdated version of the function (docs))