import tensorflow as tf
ke = tf.keras

import numpy as np

import glob

def preprocessing(fn):
  binary = tf.read_file(fn)
  audio = tf.contrib.ffmpeg.decode_audio(binary)
  # ... etc.
  return audio

list_of_audio_files = glob.glob('*.wav')

# Could be:
dataset = tf.data.Dataset().from_tensor_slices(list_of_audio_files)
# or:
dataset = tf.data.Dataset().from_tensor_slices((list_of_mix_files, list_of_source_files))
# or:
dataset = tf.data.Dataset().from_tensor_slices(np.random.uniform(-1, 1, [1000, 100]))
# or some other way, there are multiple (see api reference).

# Do preprocessing:
dataset = dataset.map(preprocessing)

dataset = dataset
  .batch(512) # Select batch size.
  .repeat() # Don't stop getting data when iterated through once (necessary for model.fit).
  .prefetch(100) # Prefetch to ram (there's some other command for prefetching to gpu can't remember).

model = ke.Sequential(...)
model.fit(dataset)

# You can also get the tf tensors for a batch directly from a data.dataset.
test_dataset = tf.data.Dataset().from_tensor_slices(np.random.uniform(-1, 1, [1000, 100]))
test_iterator = test_dataset.make_initializable_iterator()
next_batch = test_iterator.get_next()

# And maybe grab the model outputs for the test data.
test_output = model(next_batch)