CODEWITHSUNDEEP
pythontensorflow2.0
flobotics robotics
flobotics robotics

Reputation: 216

What is a replacement for tf.losses.absolute_difference

My question is about in TF2.0. There is no tf.losses.absolute_difference() function and also there is no tf.losses.Reduction.MEAN attribute.

What should I use instead? Is there a list of deleted TF functions in TF2 and perhaps their replacement.

This is TF1.x code which does not run with TF2: 

result = tf.losses.absolute_difference(a,b,reduction=tf.losses.Reduction.MEAN)

Upvotes: 2

Views: 1383

Answers (1)

Vlad
Vlad

Reputation: 8595

You still can access this function via tf.compat.v1:

import tensorflow as tf

labels = tf.constant([[0, 1], [1, 0], [0, 1]])
predictions = tf.constant([[0, 1], [0, 1], [1, 0]])

res = tf.compat.v1.losses.absolute_difference(labels,
                                              predictions,
                                              reduction=tf.compat.v1.losses.Reduction.MEAN)
print(res.numpy()) # 0.6666667

Or you could implement it yourself:

import tensorflow as tf
from tensorflow.python.keras.utils import losses_utils

def absolute_difference(labels, predictions, weights=1.0, reduction='mean'):
    if reduction == 'mean':
        reduction_fn = tf.reduce_mean
    elif reduction == 'sum':
        reduction_fn = tf.reduce_sum
    else:
        # You could add more reductions
        pass
    labels = tf.cast(labels, tf.float32)
    predictions = tf.cast(predictions, tf.float32)
    losses = tf.abs(tf.subtract(predictions, labels))
    weights = tf.cast(tf.convert_to_tensor(weights), tf.float32)
    res = losses_utils.compute_weighted_loss(losses,
                                             weights,
                                             reduction=tf.keras.losses.Reduction.NONE)

    return reduction_fn(res, axis=None)

res = absolute_difference(labels, predictions)
print(res.numpy()) # 0.6666667

Upvotes: 3

Related Questions