How to construct a custom loss function with additional parameters？

SuSei · March 26, 2022, 2:42am

for example,

Blockquote loss = tf.reduce_mean(tf.square(heatmap_outs - gt_heatmap) * valid_mask)

If I want to calculate the loss function, in addition to y_pred and y_true, there is a valid_mask, and valid_mask is not a fixed parameter. Is there a way to achieve this by inheriting from tf.keras.losses.Loss?

Bhack · March 26, 2022, 2:54pm

github.com/keras-team/keras

Passing additional arguments to objective function

opened 02:19PM - 29 Mar 16 UTC

closed 02:24PM - 26 Apr 16 UTC

carlthome

Is it possible to pass additional arguments (**kvargs) to a custom objective fun…ction? Specifically I'd like the [mse](https://github.com/fchollet/keras/blob/master/keras/objectives.py#L6) to also maximize the difference between two different target outputs, i.e. something like: ``` Python def interference_mean_squared_error(y_true, y_pred, other_y_pred): return K.mean(K.square(y_pred - y_true) - K.square(y_true - other_y_pred) , axis=-1) ```