loss: NaN when training ucr timeseries set

Question

I'm trying to change the input of this model,

https://keras.io/examples/timeseries/timeseries_classification_from_scratch/

the model architecture is as follow:

def make_model(input_shape):
    input_layer = keras.layers.Input(input_shape)

    conv1 = keras.layers.Conv1D(filters=64, kernel_size=3, padding="same")(input_layer)
    conv1 = keras.layers.BatchNormalization()(conv1)
    conv1 = keras.layers.ReLU()(conv1)

    conv2 = keras.layers.Conv1D(filters=64, kernel_size=3, padding="same")(conv1)
    conv2 = keras.layers.BatchNormalization()(conv2)
    conv2 = keras.layers.ReLU()(conv2)

    conv3 = keras.layers.Conv1D(filters=64, kernel_size=3, padding="same")(conv2)
    conv3 = keras.layers.BatchNormalization()(conv3)
    conv3 = keras.layers.ReLU()(conv3)

    gap = keras.layers.GlobalAveragePooling1D()(conv3)

    output_layer = keras.layers.Dense(num_classes, activation="softmax")(gap)

    return keras.models.Model(inputs=input_layer, outputs=output_layer)

model = make_model(input_shape=x_train.shape[1:])

the original model take the input as FordA dataset from the UCR/UEA archive. I just change the dataset to another dataset of UCR/UEA archive.

It has no error, but there is a NaN loss problem as follow:

Epoch 1/500
3/3 [==============================] - 2s 104ms/step - loss: nan - sparse_categorical_accuracy: 0.0000e+00 - val_loss: nan - val_sparse_categorical_accuracy: 0.0000e+00 - lr: 0.0010
Epoch 2/500
3/3 [==============================] - 0s 23ms/step - loss: nan - sparse_categorical_accuracy: 0.0000e+00 - val_loss: nan - val_sparse_categorical_accuracy: 0.0000e+00 - lr: 0.0010
Epoch 3/500
3/3 [==============================] - 0s 24ms/step - loss: nan - sparse_categorical_accuracy: 0.0000e+00 - val_loss: nan - val_sparse_categorical_accuracy: 0.0000e+00 - lr: 0.0010

I think it is related to input's length. I don't know where the length of the input affects.

Does anyone know why?