What's the threshold value of binary_accuracy in keras Metrices is used to predicted one sample as positive and negative cases? is that threshold value 0.5? how to adjust it? I want to set the threshold value 0.80, if the predicted value is 0.79, then it is considered a negative sample，otherwise，if the predicted value is 0.81, then it is considered a positive sample.
-
Please, before making questions here, you need to certify the best way to do it (as well as improving the English). So, check How to create a Minimal, Complete, and Verifiable example and update it properly, ok?– Diogo SJan 14, 2017 at 16:01
Add a comment
|
3 Answers
binary_accuracy don't have threshold param but you can easily define one yourself.
import keras
from keras import backend as K
def threshold_binary_accuracy(y_true, y_pred):
threshold = 0.80
if K.backend() == 'tensorflow':
return K.mean(K.equal(y_true, K.tf.cast(K.lesser(y_pred,threshold), y_true.dtype)))
else:
return K.mean(K.equal(y_true, K.lesser(y_pred,threshold)))
a_pred = K.variable([.1, .2, .6, .79, .8, .9])
a_true = K.variable([0., 0., 0., 0., 1., 1.])
print K.eval(keras.metrics.binary_accuracy(a_true, a_pred))
print K.eval(threshold_binary_accuracy(a_true, a_pred))
Now you can use it as metrics=[threshold_binary_accuracy]
answered Jan 15, 2017 at 3:38
-
So, if one were to set
threshold=0.5, they should get the same value as if they just used the default"accuracy"metric, right? I'm getting wildly different values when I do this... Aug 28, 2017 at 3:37 -
def binary_accuracy(y_true, y_pred, threshold=0.5): threshold = math_ops.cast(threshold, y_pred.dtype) y_pred = math_ops.cast(y_pred > threshold, y_pred.dtype) return K.mean(math_ops.equal(y_true, y_pred), axis=-1)Seems standard in Tensorflow but not in Keras. This function works exactly like the normal binary accuracy in Keras. Encountered the same problem as you, there is definitely something wrong with the suggested method.– AxlMar 1, 2019 at 9:17 -
To answer the initial question, keras uses the round function to assign classes, so the threshold is 0.5.
https://github.com/fchollet/keras/blob/master/keras/metrics.py
def binary_accuracy(y_true, y_pred):
return K.mean(K.equal(y_true, K.round(y_pred)))
Even faster, if you use tf 2 you can use: tf.keras.metrics.BinaryAccuracy() which has an internal threshold argument that you can set
answered Mar 20, 2020 at 15:37
