def test_model_train(sess, params):
sample_rate = 44100
time_length = 1
n_points = sample_rate * time_length
t = np.arange(0, n_points) / sample_rate
freq = 440
sine_wave = np.sin(t * 2 * np.pi * t * freq).reshape((-1, 1))
conditioning = compute_features(sine_wave, params)
sine_wave = sine_wave[np.newaxis, :, :]
conditioning = conditioning[np.newaxis, :, :]
sine_wave_encoded = mu_law_encode(sine_wave, params.quantisation)
train_spec = model_fn(
{'waveform': tf.constant(sine_wave_encoded),
'conditioning': tf.constant(conditioning)},
tf.estimator.ModeKeys.TRAIN,
params
)
with tf.variable_scope('', reuse=tf.AUTO_REUSE):
predict_spec = model_fn(
{'waveform': tf.constant(sine_wave_encoded),
'conditioning': tf.constant(conditioning)},
tf.estimator.ModeKeys.PREDICT,
params
)
sess.run(tf.global_variables_initializer())
for i in range(2000):
train_value, loss_value = sess.run([train_spec.train_op,
train_spec.loss])
print(f'{i} - {loss_value}')
if loss_value < 0.8:
break
else:
raise AssertionError(
f'Training did not converge. Final loss: {loss_value}'
)
predictions = sess.run(predict_spec.predictions)
actual = mu_law_decode(predictions[0, :-1], params.quantisation)
expected = sine_wave[0, params.receptive_field:, 0]
accuracy = np.mean(np.abs(actual - expected))
# Channels are quantised as per mu_law_encode, assert that average error
# is less that the quantisation level, i.e, 1.
assert accuracy < 1.0
def test_model_shape_predict(sess, params):
features = make_inputs(params, params.receptive_field)
model = model_fn(
features,
tf.estimator.ModeKeys.PREDICT,
params
)
sess.run(tf.global_variables_initializer())
predictions = sess.run(model.predictions)
assert predictions.shape[1] == 1
def test_conditioning_shape_eval(sess, params):
params.conditioning = True
features = make_inputs(params, params.receptive_field + 1)
model = model_fn(
features,
tf.estimator.ModeKeys.EVAL,
params
)
sess.run(tf.global_variables_initializer())
predictions = sess.run(model.predictions)
assert predictions.shape[1] == 1
def loop(i, waveform_):
features = {
'waveform': waveform_[tf.newaxis, -slice_size:, :],
'conditioning': conditioning_tf[i:i + slice_size, :]
}
estimator_spec = model_fn(features, tf.estimator.ModeKeys.PREDICT,
params)
next_point = estimator_spec.predictions
waveform_ = tf.concat([waveform_, next_point[0, -1:, tf.newaxis]],
axis=0)
return [tf.add(i, 1), waveform_]