def _reconstruct(self, sess, data_reader, max_steps):
data_reader.start()
reconstructed = StrechableNumpyArray()
out_gaps = StrechableNumpyArray()
for batch_num in range(max_steps):
try:
audio = data_reader.dataOperation(session=sess)
except StopIteration:
print("rec End of queue!", batch_num)
break
feed_dict = self._feedDict(audio, sess, False)
reconstructed_input, original = sess.run(
[self._architecture.output(),
self._architecture.target()],
feed_dict=feed_dict)
out_gaps.append(np.reshape(original, (-1)))
reconstructed.append(np.reshape(reconstructed_input, (-1)))
output_shape = self._architecture.output().shape.as_list()
output_shape[0] = -1
reconstructed = reconstructed.finalize()
reconstructed = np.reshape(reconstructed, output_shape)
out_gaps = out_gaps.finalize()
out_gaps = np.reshape(out_gaps, output_shape)
data_reader.finish()
return reconstructed, out_gaps
def reconstructAudio(self, audios, model_num=None, max_batchs=200):
with tf.Session() as sess:
if model_num is not None:
path = self.modelsPath(model_num)
else:
path = self.modelsPath(self._initial_model_num)
saver = tf.train.Saver()
saver.restore(sess, path)
print("Model restored.")
batches_count = int(len(audios) / self._batch_size)
reconstructed = StrechableNumpyArray()
original_stfts = StrechableNumpyArray()
for batch_num in range(min(batches_count, max_batchs)):
batch_data = audios[batch_num * self._batch_size:batch_num *
self._batch_size + self._batch_size]
feed_dict = {
self._model.input(): batch_data,
self._model.isTraining(): False
}
reconstructed_input, original_stft = sess.run(
[self._reconstructed_input_data, self._stft],
feed_dict=feed_dict)
original_stfts.append(np.reshape(original_stft, (-1)))
reconstructed.append(np.reshape(reconstructed_input, (-1)))
reconstructed = reconstructed.finalize()
reconstructed_stft = np.reshape(reconstructed, (-1, 37, 257))
original_stfts = original_stfts.finalize()
original_stft = np.reshape(original_stfts, (-1, 7, 257, 2))
return reconstructed_stft, original_stft
def _reconstruct(self, sess, data_reader, max_steps):
data_reader.start()
reconstructed = StrechableNumpyArray()
out_gaps = StrechableNumpyArray()
for batch_num in range(max_steps):
try:
sides, gaps = data_reader.dataOperation(session=sess)
except StopIteration:
print(batch_num)
print("rec End of queue!")
break
reconstructed_signal = sess.run(self._reconstructedSignal,
feed_dict={
self._sides: sides,
self.gap_data: gaps
})
# Nati comment: why is this step not done with reconstructed_signal?
# gap_stft = reconstructed_signal[:, 15:15+7, :]
gap_stft = tf.abs(self._stft[:, 15:15 + 7, :])
feed_dict = {
self._model.input(): reconstructed_signal,
self._model.isTraining(): False
}
reconstructed_input, original = sess.run(
[self._reconstructed_input_data, gap_stft],
feed_dict=feed_dict)
out_gaps.append(np.reshape(original, (-1)))
reconstructed.append(np.reshape(reconstructed_input, (-1)))
reconstructed = reconstructed.finalize()
reconstructed = np.reshape(reconstructed, (-1, 7, 257))
out_gaps = out_gaps.finalize()
out_gaps = np.reshape(out_gaps, (-1, 7, 257))
istft_original = sess.run(self._reconstructedAudio,
feed_dict={self._specgram: out_gaps})
istft_reconstructed = sess.run(
self._reconstructedAudio,
feed_dict={self._specgram: reconstructed})
data_reader.finish()
return istft_reconstructed, istft_original
def _reconstruct(self, sess, data_reader, max_steps):
data_reader.start()
reconstructed = StrechableNumpyArray()
out_gaps = StrechableNumpyArray()
for batch_num in range(max_steps):
try:
sides, gaps = data_reader.dataOperation(session=sess)
except StopIteration:
print(batch_num)
print("rec End of queue!")
break
reconstructed_signal = sess.run(self._reconstructedSignal,
feed_dict={
self._sides: sides,
self.gap_data: gaps
})
gap_stft = self._target_model.output()
feed_dict = {
self._model.input(): reconstructed_signal,
self._target_model.input(): reconstructed_signal,
self._model.isTraining(): False
}
reconstructed_input, original = sess.run(
[self._reconstructed_input_data, gap_stft],
feed_dict=feed_dict)
out_gaps.append(np.reshape(original, (-1)))
reconstructed.append(np.reshape(reconstructed_input, (-1)))
output_shape = self._target_model.output().shape.as_list()
output_shape[0] = -1
reconstructed = reconstructed.finalize()
reconstructed = np.reshape(reconstructed, output_shape)
out_gaps = out_gaps.finalize()
out_gaps = np.reshape(out_gaps, output_shape)
data_reader.finish()
return reconstructed, out_gaps
def reconstructAudio(self, aBatchOfSignals, model_num, max_steps=200):
with tf.Session() as sess:
path = self.modelsPath(model_num)
saver = tf.train.Saver()
saver.restore(sess, path)
print("Model restored.")
sess.run([tf.local_variables_initializer()])
reconstructed = StrechableNumpyArray()
out_gaps = StrechableNumpyArray()
input_shape = list(self._architecture.inputShape())
input_shape[0] = 0
contexts = np.empty(input_shape)
for batch_num in range(
min(int(len(aBatchOfSignals) / self._batchSize),
max_steps)):
feed_dict = self._feedDict(
aBatchOfSignals[batch_num *
self._batchSize:(batch_num + 1) *
self._batchSize], sess, False)
reconstructed_input, original, context = sess.run(
[
self._architecture.output(),
self._architecture.target(),
self._architecture.input()
],
feed_dict=feed_dict)
out_gaps.append(np.reshape(original, (-1)))
reconstructed.append(np.reshape(reconstructed_input, (-1)))
contexts = np.concatenate([contexts, context], axis=0)
output_shape = self._architecture.output().shape.as_list()
output_shape[0] = -1
reconstructed = reconstructed.finalize()
reconstructed = np.reshape(reconstructed, output_shape)
out_gaps = out_gaps.finalize()
out_gaps = np.reshape(out_gaps, output_shape)
return reconstructed, out_gaps, contexts
def _reconstruct(self, sess, data_reader, max_steps):
data_reader.start()
reconstructed = StrechableNumpyArray()
out_gaps = StrechableNumpyArray()
for batch_num in range(max_steps):
try:
sides, gaps = data_reader.dataOperation(session=sess)
except StopIteration:
print(batch_num)
print("rec End of queue!")
break
out_gaps.append(np.reshape(gaps, (-1)))
feed_dict = {self._model.input(): sides, self._model.isTraining(): False}
reconstructed.append(np.reshape(sess.run(self._reconstructed_input_data, feed_dict=feed_dict), (-1)))
reconstructed = reconstructed.finalize()
reconstructed = np.reshape(reconstructed, (-1, self._gap_length))
out_gaps = out_gaps.finalize()
out_gaps = np.reshape(out_gaps, (-1, self._gap_length))
data_reader.finish()
return reconstructed, out_gaps
def _reconstruct(self, sess, aBatchOfSignals):
reconstructed = StrechableNumpyArray()
out_gaps = StrechableNumpyArray()
input_shape = self._architecture.input().shape.as_list()
input_shape[0] = 0
contexts = np.empty(input_shape)
for batch_num in range(int(len(aBatchOfSignals)/self._batchSize)):
feed_dict = self._feedDict(aBatchOfSignals[batch_num*self._batchSize:(batch_num+1)*self._batchSize], sess, False)
reconstructed_input, original, context = sess.run([self._architecture.output(), self._architecture.target(),
self._architecture.input()],
feed_dict=feed_dict)
out_gaps.append(np.reshape(original, (-1)))
reconstructed.append(np.reshape(reconstructed_input, (-1)))
contexts = np.concatenate([contexts, context], axis=0)
output_shape = self._architecture.output().shape.as_list()
output_shape[0] = (batch_num+1) * self._batchSize
reconstructed = reconstructed.finalize()
reconstructed = np.reshape(reconstructed, output_shape)
out_gaps = out_gaps.finalize()
out_gaps = np.reshape(out_gaps, output_shape)
return reconstructed, out_gaps, contexts