I am working with old TensorFlow (2.7) and Java Tensorflow but the answer may still help others tracing similar issues. My model produces the error when saved to the 2.7 save() format: CustomMaskWarning: Custom mask layers require a config and must override get_config. When loading, the custom mask layer must be passed to the custom_objects argument. Originally I imagined this came from the metric and loss function, since the model cannot save and load these, presumably because of some sort of serialization issue. def masked_mae(y_true, y_pred): # Mask NaN values, replace by 0 y_true = tf.where(tf.math.is_nan(y_true), y_pred, y_true) # Calculate absolute differences absolute_differences = tf.abs(y_true - y_pred) # Compute the mean, ignoring potential NaN values (if any remain after replacement) mae = tf.reduce_mean(absolute_differences) return mae I don't know how to deliver custom_objects in Java, and I'm done training, so I tried compiling the model with loss=None and metrics=None and saving it that way. The model can be loaded with this change but the error has not gone away. Is there a way to trace what layer is causing this and why? Is it just a nuisance? There are no obvious uses of masks. Continue reading...
