implementation of tf.nn.weighted_cross_entropy_with_logits

Question

I am trying to add weights to classes in imbalanced data in a model with multi output layers in keras functional API

but i tried class_weight but couldn't get it to work either when my model was 1 output layer with onehot output or when dividing the out put into 50+ class each with different output layer class_weight require a dict in tf 2.80 i also tried tf.nn.weighted_cross_entropy_with_logits as a loss function but it doesnt work as a normal loss function it requires to be wrapped in a different function and all i found doesnt work for me and they doesnt explain it

some of the errors was:

Missing required positional argument
TypeError: Input 'y' of 'Mul' Op has type float32 that does not match type int64 of argument 'x'.

i use colab my data outputs:

x, y = next(valid_generator)
np.array(y).shape

(51, 32, 1)

51 array of (51,1) and batch of 32

input_shape=256
channels=1



inputs = keras.Input(shape=(input_shape, input_shape, channels), name="img")
x = layers.Conv2D(128, 3, activation="relu")(inputs)
pooling=layers.MaxPooling2D()(x)

conv=layers.Conv2D(64, 3, activation="relu")(pooling)
pooling=layers.MaxPooling2D()(conv)

conv=layers.Conv2D(64, 3, activation="relu")(pooling)
pooling=layers.MaxPooling2D()(conv)

conv=layers.Conv2D(64, 3, activation="relu")(pooling)
pooling=layers.MaxPooling2D()(conv)

flat=layers.Flatten()(pooling)
pflat=layers.Dense(256, activation="relu")(flat)
pflat = layers.Dropout(0.5)(pflat)#noraml


output_solo=[]
mmetrics={}
loss={}
class_weights={}

for i in df.iloc[:,35:-1].columns:
  out = layers.Dense(1,activation="relu",name=f'{"_".join(i.split())}_out')(pflat)
  output_solo.append(out)
  mmetrics[f'{"_".join(i.split())}_out']=tf.keras.metrics.Accuracy(name=f'Accuracy{i}out')
  loss[f'{"_".join(i.split())}_out']=weighted_binary_crossentropy(40)
  class_weights[f'{"_".join(i.split())}_out']={0: df[i].mean() , 1:1}


model = keras.Model(inputs, output_solo , name="shit")
model.summary()

keras.utils.plot_model(model, "shit.png", show_shapes=True)


model.compile(
    loss = loss ,

    metrics =mmetrics,

    optimizer = tf.keras.optimizers.Adam()

)



plateau = tf.keras.callbacks.ReduceLROnPlateau(monitor='BinaryAccuracy_normal',
                                     patience = 1,
                                     factor = 0.01,
                                     mode = 'max')


early_stopping = tf.keras.callbacks.EarlyStopping(monitor = 'BinaryAccuracy_normal',
                                        patience = 3,
                                        restore_best_weights=True,
                                        mode = 'max')



check_point=tf.keras.callbacks.ModelCheckpoint(
    filepath='/content/drive/MyDrive/save',
    monitor="BinaryAccuracy_normal",
    verbose=0,
    save_best_only=True,
    save_weights_only=True,
    mode="auto",
    save_freq="epoch")





num_epochs = 1



hist = model.fit(
    train_generator,
    epochs=num_epochs, steps_per_epoch=400,
    validation_data=valid_generator,
    validation_steps=400,
    callbacks=[check_point])