predictions must be <= 1. Error while training LSTM model - tensorflow

Question

So I just finished my model and wanted to start training but I think something went wrong with my metrics. First my model structure
inputs = tf.keras.Input(shape=(None,nb_features), name = 'inputs')
x = tf.keras.layers.Masking(mask_value = data.MASK_VALUE)(inputs)
x = tf.keras.layers.LSTM(hidden_units,
return_sequences = True,
dropout = dropout_rate)(x)
dense = tf.keras.layers.Dense(nb_skills, activation = 'sigmoid')
outputs = tf.keras.layers.TimeDistributed(dense, name = 'outputs')(x)
Edit:
My Model is looking like this.
Model: "DKTModel"
inputs (InputLayer) [(None, None, 80)] 0
masking (Masking) (None, None, 80) 0
lstm (LSTM) (None, None, 100) 72400
outputs (TimeDistributed) (None, None, 40) 4040
=================================================================
Total params: 76,440
Trainable params: 76,440
Non-trainable params: 0
Edit:
Here are my compile and fit function.
def compile(self, optimizer, metrics=None):
def custom_loss(y_true, y_pred):
y_true, y_pred = data.get_target(y_true, y_pred)
return tf.keras.losses.binary_crossentropy(y_true, y_pred)
super(DKTModel, self).compile(
loss = custom_loss,
optimizer = optimizer,
metrics = metrics,
experimental_run_tf_function = False)
def fit (self,
dataset,
epochs = 1,
verbose = 1,
callbacks=None,
validation_data = None,
shuffle = True,
initial_epoch = 0,
steps_per_epoch = None,
validation_steps = None,
validation_freq = 1):
return super (DKTModel, self).fit(x=dataset, epochs=epochs,verbose=verbose, callbacks = callbacks, validation_data = validation_data, shuffle = shuffle, initial_epoch = initial_epoch, steps_per_epoch = steps_per_epoch, validation_steps = validation_steps, validation_freq = validation_freq)
I get following error when running model.fit().
2 root error(s) found.(0) INVALID_ARGUMENT: assertion failed: [predictions must be <= 1] [Condition x <= y did not hold element-wise:] [x (Sum_5:0) = ] [[[19.462822][19.5533848][19.5251656]]...] [y (Cast_11/x:0) = ] [1] [[{{node assert_less_equal/Assert/AssertGuard/Assert}}]][[assert_less_equal_2/Assert/AssertGuard/pivot_f/_122/_201]](1) INVALID_ARGUMENT: assertion failed: [predictions must be <= 1] [Condition x <= y did not hold element-wise:] [x (Sum_5:0) = ] [[[19.462822][19.5533848][19.5251656]]...] [y (Cast_11/x:0) = ] [1][[{{node assert_less_equal/Assert/AssertGuard/Assert}}]]
0 successful operations.
0 derived errors ignored. [Op:__inference_train_function_7560]
Edit:
I changed the activation function from sigmoid to softmax but still got the same error. But the values are different
2 root error(s) found.(0) INVALID_ARGUMENT: assertion failed: [predictions must be <= 1] [Condition x <= y did not hold element-wise:] [x (Sum_5:0) = ] [[[0.99999994][1][1]]...] [y (Cast_11/x:0) = ] [1][{{node assert_less_equal/Assert/AssertGuard/Assert}}]]
[[broadcast_weights_2/assert_broadcastable/AssertGuard/pivot_f/_58/_101]](1) INVALID_ARGUMENT: assertion failed: [predictions must be <= 1] [Condition x <= y did not hold element-wise:] [x (Sum_5:0) = ] [[[0.99999994][1][1]]...] [y (Cast_11/x:0) = ] [1]
[[{{node assert_less_equal/Assert/AssertGuard/Assert}}]]
0 successful operations.
0 derived errors ignored. [Op:__inference_train_function_7568]
Edit:
Here are the metrics i use to compile.
student_model.compile(
optimizer = optimizer,
metrics = [
metrics.AUC(),
metrics.Precision(),
metrics.Recall()
])
Ok I just removed all metrics and it started at least but got an error which I try to solve now.
Which lets met think that some metrics are not appliable to my outputs but I dont know how to change it. Maybe someone encountered this problem before. If you need the full error I get as well as the last part I showed above, let me know im thankful for every help.
Thanks

Answer 1

I see that you have a nb_skills number of outputs. If this number is >2 I think that the problem is that you are using a sigmoid as activation function.
Try substituting it with a softmax.
Remember that for multi-class classification you should use a softmax activation function. The softmax function normalizes a set of N (nb_skills) real numbers into a probability distribution such that they sum up to 1. You can think of the sigmoid as a softmax with N=2 classes.
Update:
From your updated code I see now that you also used a binary_crossentropy, and again this is not fit for a multi-classification problem. So instead of:
tf.keras.losses.binary_crossentropy(y_true, y_pred)
try using:
tf.keras.losses.sparse_categorical_crossentropy(y_true, y_pred)