Custom loss function NOT found

I’ve defined several custom loss functions:

dice_loss = dice_loss(num_classes=args.num_classes,name='dice_loss')
tversky_loss = tversky_loss(beta=0.7,num_classes = args.num_classes,name='tversky_loss)

and my build model function is like this:

def build_model(hp):
        model = models.segmentationModel(output_channels=args.num_classes)
        model.build(input_shape=((1,256,256,3)))
        model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=hp.Choice('learning_rate', [1e-2, 1e-3, 1e-4])),
                    loss = hp.Choice('loss', ['dice_loss', 'tversky_loss']),
                    metrics=[tf.keras.metrics.SparseCategoricalAccuracy(name='sparse_categorical_accuracy', dtype=None),UpdatedMeanIoU(name='Mean_IOU',num_classes=args.num_classes)])
        return model

When I execute this code, I get this error:

ValueError: Unknown loss function:dice_loss

[mandar2812]

@dhaneshr I think you might have to try something similar to what is mentioned here when loading custom objects.

Keras would not know from the strings dice_loss and tversky_loss that you mean your custom implementations.

Something like this might help:

def build_model(hp):
  custom_objects = {"tversky_loss": ..., "dice_loss": ...}
  with keras.utils.custom_object_scope(custom_objects):
      ## build model code here

[dhaneshr]

Hi again Mandar,

I ended up implementing the 2nd option which seems to be working so far. Here is my code:

    acc_metric = tf.keras.metrics.SparseCategoricalAccuracy(name='xent_acc', dtype=None)
    miou_metric = UpdatedMeanIoU(name='miou',num_classes=args.num_classes)
    

    def build_model(hp):
        model = models.keras_model(output_channels=args.num_classes,dropout_rate=hp.Choice('dropout_rate',[0.10,0.25,0.5,0.7]))
        model.build(input_shape=((1,256,256,3)))
        loss_hyp_param = hp.Choice('loss', ['dice_loss', 'tversky_loss']) 
        beta = hp.Choice('beta', [0.1,0.25,0.5,0.75])
        if loss_hyp_param == "dice_loss":
            loss_fn = dice_loss(num_classes=args.num_classes,name='dice_loss')
        else:
            loss_fn = tversky_loss(beta=beta,num_classes = args.num_classes,name='tversky_loss')

        model.compile(optimizer=hp.Choice('optimizer', ['Nadam','Adam']),
                    loss = loss_fn,
                    metrics=[miou_metric,acc_metric])
        return model


    hp = HyperParameters()

    tuner = RandomSearch(
        build_model,
        objective = kerastuner.Objective('val_miou', direction="max"),
        max_trials=60,
        executions_per_trial = 1,
        overwrite=True,
        directory='test_dir')
   
    tuner.search_space_summary()

    tuner.search(train_dataset,
                epochs=25,
                validation_data=val_dataset)
    tuner.results_summary()