Loss function testing

Question

I just want to define a loss function to test.

I used as example the euclidean distance:

def euc_dist_keras(y_true, y_pred):
    return K.sqrt(K.sum(K.square(y_true - y_pred), axis=-1, keepdims=True))

Since I have to obtain from the net list of couples (x,y), I want to test this outside the NN.

So I used:

y_true = [[0., 1.], [0., 0.]]
y_pred = [[1., 1.], [1., 0.]]

With just:

edk = euc_dist_keras(y_true, y_pred)

I obtained the error: TypeError: unsupported operand type(s) for -: 'list' and 'list'

So I used:

y_true_array = np.array(y_true)
y_pred_array = np.array(y_pred)
edk = euc_dist_keras(y_true_array, y_pred_array)

But obtained:

Tensor("Sqrt:0", shape=(2, 1), dtype=float64)

Instead of expected output value: 1

How to obtain the desired value? The same euc_dist_keras, used in:

model.compile(loss=euc_dist_keras, optimizer=opt)

will work exactly in the same way I'm testing it?

Thanks!

Added:

with tf.Session() as sess:  print(edk.eval())

I obtained: [[1.] [1.]]

I expected: 1.

Maybe I make some mistake in the def? Or the mean of all samples is made just when I use it when compile the model?

you need to add K.mean in the loss – Marco Cerliani Sep 10 '20 at 06:58 — Marco Cerliani, Sep 10 '20 at 06:58

score 0 · Accepted Answer · answered Sep 10 '20 at 06:58

0

It sounds like this is TensorFlow v1 code. If so, you must run the operations within a "session" in order to evaluate them. See SO post How to print the value of a Tensor object in TensorFlow? and the TF v1 documentation on sessions:
https://www.tensorflow.org/api_docs/python/tf/compat/v1/Session

answered Sep 10 '20 at 06:58

Pascal Getreuer

In each case: with tf.Session() as sess: print(edk.eval()) seems to almost work. It gives [[1.][1.]]. I expected 1. – user3318590 Sep 10 '20 at 07:18

1 Answers1