Caution in Training Steps

[Andrew-Ma-2001]

I notice that in the training steps given in the file Train Model-All Words.ipynb in line 12 training dataset part

# create the datasets for training
batch_size = 32

train_dataset = Dataset.from_tensor_slices(
    (X_train, Y_train)
).repeat(
    count=-1
).shuffle(
    len(X_train)
).batch(
    batch_size
)

validation_dataset = Dataset.from_tensor_slices((X_validate, Y_validate)).batch(X_validate.shape[0]//10)

test_dataset = Dataset.from_tensor_slices((X_test, Y_test)).batch(len(X_test))

We can see that for train_dataset it have an option of Dataset.repeat(count=-1) which repeats the dataset infiniely while in training. I want to mention that for those who use custom audio dataset and apply the author code should take full consideration of using this option. For dataset that are umbalanced it may cause training hard to converge or over-fitting model.