def test03TheStftKnowsWhatSignalItShouldTakeForTheSTFTOfTheGap(self):
fake_batch_of_signal = np.array([np.arange(self.signal_length)])
produced_signal = self.anStftForTheInpaintingSetting._removeExtraSidesForSTFTOfGap(
fake_batch_of_signal)
gap_begins = (self.signal_length - self.gap_length) // 2
gap_ends = gap_begins + self.gap_length
padding = self.fft_window_length - self.fft_hop_size
np.testing.assert_almost_equal(
fake_batch_of_signal[:, gap_begins - padding:gap_ends + padding],
produced_signal)
fft_window_length = 128
fft_hop_size = 32
anStftForTheInpaintingSetting = PreAndPostProcessor(
signalLength=self.signal_length,
gapLength=self.gap_length,
fftWindowLength=fft_window_length,
fftHopSize=fft_hop_size)
produced_signal = anStftForTheInpaintingSetting._removeExtraSidesForSTFTOfGap(
fake_batch_of_signal)
padding = fft_window_length - fft_hop_size
np.testing.assert_almost_equal(
fake_batch_of_signal[:, gap_begins - padding:gap_ends + padding],
produced_signal)
def setUp(self):
self.signal_length = 5120
self.gap_length = 1024
self.fft_window_length = 512
self.fft_hop_size = 128
self.params = LstmContextInpaintingParameters(lstmSize=512, batchSize=64, signalLength=self.signal_length,
gapLength=self.gap_length,
fftWindowLength=self.fft_window_length, fftHopSize=self.fft_hop_size)
self.anStftForTheInpaintingSetting = PreAndPostProcessor(self.params)
def setUp(self):
self.signal_length = 5120
self.gap_length = 1024
self.fft_window_length = 512
self.fft_hop_size = 128
self.anStftForTheInpaintingSetting = PreAndPostProcessor(
signalLength=self.signal_length,
gapLength=self.gap_length,
fftWindowLength=self.fft_window_length,
fftHopSize=self.fft_hop_size)
def test02TheStftKnowsHowMuchPaddingItShouldApply(self):
self.assertEquals(self.anStftForTheInpaintingSetting.padding(), self.fft_window_length-self.fft_hop_size)
fft_window_length = 1024
fft_hop_size = 128
params = LstmContextInpaintingParameters(lstmSize=512, batchSize=64, signalLength=self.signal_length,
gapLength=self.gap_length,
fftWindowLength=fft_window_length, fftHopSize=fft_hop_size)
anStftForTheInpaintingSetting = PreAndPostProcessor(params)
self.assertEquals(anStftForTheInpaintingSetting.padding(), fft_window_length - fft_hop_size)
fft_window_length = 1024
fft_hop_size = 256
params = LstmContextInpaintingParameters(lstmSize=512, batchSize=64, signalLength=self.signal_length,
gapLength=self.gap_length,
fftWindowLength=fft_window_length, fftHopSize=fft_hop_size)
anStftForTheInpaintingSetting = PreAndPostProcessor(params)
self.assertEquals(anStftForTheInpaintingSetting.padding(), fft_window_length - fft_hop_size)
def test02TheStftKnowsHowMuchPaddingItShouldApply(self):
self.assertEquals(self.anStftForTheInpaintingSetting.padding(),
self.fft_window_length - self.fft_hop_size)
fft_window_length = 1024
fft_hop_size = 128
anStftForTheInpaintingSetting = PreAndPostProcessor(
signalLength=self.signal_length,
gapLength=self.gap_length,
fftWindowLength=fft_window_length,
fftHopSize=fft_hop_size)
self.assertEquals(anStftForTheInpaintingSetting.padding(),
fft_window_length - fft_hop_size)
fft_window_length = 1024
fft_hop_size = 256
anStftForTheInpaintingSetting = PreAndPostProcessor(
signalLength=self.signal_length,
gapLength=self.gap_length,
fftWindowLength=fft_window_length,
fftHopSize=fft_hop_size)
self.assertEquals(anStftForTheInpaintingSetting.padding(),
fft_window_length - fft_hop_size)
__author__ = 'Andres'
tf.reset_default_graph()
train_filename = '../test_w5120_g1024_h512_ex63501.tfrecords'
valid_filename = '../test_w5120_g1024_h512_ex63501.tfrecords'
signal_length = 5120
gap_length = 1024
batch_size = 256
fft_window_length = 512
fft_hop_size = 128
aTargetModel = EmptyTfGraph(shapeOfInput=(batch_size, signal_length), name="Target Model")
anStftForTheInpaintingSetting = PreAndPostProcessor(signalLength=signal_length,
gapLength=gap_length,
fftWindowLength=fft_window_length,
fftHopSize=fft_hop_size)
anStftForTheInpaintingSetting.addStftForGapTo(aTargetModel)
aTargetModel.divideComplexOutputIntoMagAndPhase() # (256, 11, 257, 2)
aModel = EmptyTfGraph(shapeOfInput=(batch_size, signal_length), name="context encoder")
anStftForTheInpaintingSetting.addStftForTheContextTo(aModel)
aModel.divideComplexOutputIntoMagAndPhase()
aModel.addReshape((batch_size, 16, 257, 4))
with tf.variable_scope("Encoder"):
filter_shapes = [(7, 89), (3, 17), (2, 6), (1, 5), (1, 3)]
input_channels = [4, 32, 64, 128, 128]
output_channels = [32, 64, 128, 128, 200]
strides = [[1, 2, 2, 1], [1, 2, 3, 1], [1, 2, 3, 1], [1, 1, 2, 1], [1, 1, 1, 1]]
class TestStftForTheContextEncoder(TestCase):
def setUp(self):
self.signal_length = 5120
self.gap_length = 1024
self.fft_window_length = 512
self.fft_hop_size = 128
self.anStftForTheInpaintingSetting = PreAndPostProcessor(
signalLength=self.signal_length,
gapLength=self.gap_length,
fftWindowLength=self.fft_window_length,
fftHopSize=self.fft_hop_size)
def test01TheStftTakesTheInpaintingParametersAsInput(self):
self.assertEquals(self.anStftForTheInpaintingSetting.signalLength(),
self.signal_length)
self.assertEquals(self.anStftForTheInpaintingSetting.gapLength(),
self.gap_length)
self.assertEquals(self.anStftForTheInpaintingSetting.fftWindowLenght(),
self.fft_window_length)
self.assertEquals(self.anStftForTheInpaintingSetting.fftHopSize(),
self.fft_hop_size)
def test02TheStftKnowsHowMuchPaddingItShouldApply(self):
self.assertEquals(self.anStftForTheInpaintingSetting.padding(),
self.fft_window_length - self.fft_hop_size)
fft_window_length = 1024
fft_hop_size = 128
anStftForTheInpaintingSetting = PreAndPostProcessor(
signalLength=self.signal_length,
gapLength=self.gap_length,
fftWindowLength=fft_window_length,
fftHopSize=fft_hop_size)
self.assertEquals(anStftForTheInpaintingSetting.padding(),
fft_window_length - fft_hop_size)
fft_window_length = 1024
fft_hop_size = 256
anStftForTheInpaintingSetting = PreAndPostProcessor(
signalLength=self.signal_length,
gapLength=self.gap_length,
fftWindowLength=fft_window_length,
fftHopSize=fft_hop_size)
self.assertEquals(anStftForTheInpaintingSetting.padding(),
fft_window_length - fft_hop_size)
def test03TheStftKnowsWhatSignalItShouldTakeForTheSTFTOfTheGap(self):
fake_batch_of_signal = np.array([np.arange(self.signal_length)])
produced_signal = self.anStftForTheInpaintingSetting._removeExtraSidesForSTFTOfGap(
fake_batch_of_signal)
gap_begins = (self.signal_length - self.gap_length) // 2
gap_ends = gap_begins + self.gap_length
padding = self.fft_window_length - self.fft_hop_size
np.testing.assert_almost_equal(
fake_batch_of_signal[:, gap_begins - padding:gap_ends + padding],
produced_signal)
fft_window_length = 128
fft_hop_size = 32
anStftForTheInpaintingSetting = PreAndPostProcessor(
signalLength=self.signal_length,
gapLength=self.gap_length,
fftWindowLength=fft_window_length,
fftHopSize=fft_hop_size)
produced_signal = anStftForTheInpaintingSetting._removeExtraSidesForSTFTOfGap(
fake_batch_of_signal)
padding = fft_window_length - fft_hop_size
np.testing.assert_almost_equal(
fake_batch_of_signal[:, gap_begins - padding:gap_ends + padding],
produced_signal)
def test04TheStftProducesAnSTFTOfTheExpectedShapeForTheGap(self):
batch_size = 32
aBatchOfSignals = tf.placeholder(tf.float32,
shape=(batch_size,
self.signal_length),
name='input_data')
aStft = self.anStftForTheInpaintingSetting.stftForGapOf(
aBatchOfSignals)
framesOnGap = ((
(self.gap_length + self.anStftForTheInpaintingSetting.padding() *
2) - self.fft_window_length) / self.fft_hop_size) + 1
binsPerFrame = self.fft_window_length // 2 + 1
realAndImagChannels = 2
self.assertEquals(aStft.shape.as_list(),
[32, framesOnGap, binsPerFrame, realAndImagChannels])
def test05TheStftRemovesTheGapCorrectly(self):
fake_batch_of_signal = np.array([np.arange(self.signal_length)])
produced_signal = self.anStftForTheInpaintingSetting._removeGap(
fake_batch_of_signal)
gap_begins = (self.signal_length - self.gap_length) // 2
gap_ends = gap_begins + self.gap_length
left_side = fake_batch_of_signal[:, :gap_begins]
right_side = fake_batch_of_signal[:, gap_ends:]
signal_without_gap = tf.stack((left_side, right_side), axis=1)
with tf.Session() as sess:
produced_signal, signal_without_gap = sess.run(
[produced_signal, signal_without_gap])
np.testing.assert_almost_equal(signal_without_gap, produced_signal)
def test06TheStftAddsTheCorrectPaddingToTheSides(self):
side_length = (self.signal_length - self.gap_length) // 2
left_side = np.array([np.arange(side_length, dtype=np.float32)])
right_side = np.array([np.arange(side_length, dtype=np.float32)])
fake_batch_of_sides = tf.stack((left_side, right_side), axis=1)
produced_signal = self.anStftForTheInpaintingSetting._addPaddingForStftOfContext(
fake_batch_of_sides)
with tf.Session() as sess:
produced_signal = sess.run(produced_signal)
left_side_padded = np.concatenate(
(left_side,
np.zeros((1, self.fft_window_length - self.fft_hop_size))),
axis=1)
right_side_padded = np.concatenate((np.zeros(
(1, self.fft_window_length - self.fft_hop_size)), right_side),
axis=1)
new_signal = np.stack([left_side_padded, right_side_padded], axis=1)
np.testing.assert_almost_equal(new_signal, produced_signal)
def test07TheStftOfTheContextHasTheExpectedShape(self):
batch_size = 32
aBatchOfSignals = tf.placeholder(tf.float32,
shape=(batch_size,
self.signal_length),
name='input_data')
aStft = self.anStftForTheInpaintingSetting.stftForTheContextOf(
aBatchOfSignals)
side_length = (self.signal_length - self.gap_length) // 2
framesOnSides = (
(side_length + self.anStftForTheInpaintingSetting.padding() -
self.fft_window_length) / self.fft_hop_size) + 1
binsPerFrame = self.fft_window_length // 2 + 1
realAndImagChannels = 2
beforeAndAfterChannels = 2
self.assertEquals(aStft.shape.as_list(), [
32, framesOnSides, binsPerFrame,
realAndImagChannels * beforeAndAfterChannels
])
def test08TheStftProducesTheCorrectShapeWhenDoingTheInverseStftOnTheGap(
self):
batch_size = 32
framesOnGap = ((
(self.gap_length + self.anStftForTheInpaintingSetting.padding() *
2) - self.fft_window_length) / self.fft_hop_size) + 1
binsPerFrame = self.fft_window_length // 2 + 1
batchOfGapStft = tf.zeros((batch_size, framesOnGap, binsPerFrame),
dtype=tf.complex64)
batchOfGaps = self.anStftForTheInpaintingSetting.inverseStftOfGap(
batchOfGapStft)
with tf.Session() as sess:
batchOfGaps = sess.run(batchOfGaps)
self.assertEquals(batchOfGaps.shape, (batch_size, self.gap_length))
def test09TheStftProducesTheCorrectShapeWhenDoingTheInverseStftOnTheFullSignal(
self):
batch_size = 32
frameCount = ((self.signal_length - self.fft_window_length) /
self.fft_hop_size) + 1
binsPerFrame = self.fft_window_length // 2 + 1
batchOfSignalStft = tf.zeros((batch_size, frameCount, binsPerFrame),
dtype=tf.complex64)
batchOfSignals = self.anStftForTheInpaintingSetting.inverseStftOfSignal(
batchOfSignalStft)
with tf.Session() as sess:
batchOfGaps = sess.run(batchOfSignals)
self.assertEquals(batchOfGaps.shape, (batch_size, self.signal_length))
import pickle
from architecture.contextEncoderArchitecture import ContextEncoderArchitecture
from system.contextEncoderSystem import ContextEncoderSystem
from system.preAndPostProcessor import PreAndPostProcessor
architecturesParametersFile = "Papers_Context_Encoder_parameters.pkl"
sessionsName = "Papers_Context_Encoder"
with open(architecturesParametersFile, 'rb') as savedFile:
Context_Encoder_parameters = pickle.load(savedFile)
aContextEncoderArchitecture = ContextEncoderArchitecture(
*Context_Encoder_parameters['Architecture Params'])
aPreProcessor = PreAndPostProcessor(
*Context_Encoder_parameters['PreProcessor Params'])
aContextEncoderSystem = ContextEncoderSystem(
aContextEncoderArchitecture, Context_Encoder_parameters['batchSize'],
aPreProcessor, sessionsName)
aContextEncoderSystem.train("nsynth_train_w5120_g1024_h512.tfrecords",
"nsynth_valid_w5120_g1024_h512.tfrecords", 1e-3)
from architecture.parameters.lstmContextInpaintingParameters import LstmContextInpaintingParameters
from architecture.realImagContextEncoderLSTMArchitecture import RealImagContextEncoderLSTMArchitecture
import os
from system.preAndPostProcessor import PreAndPostProcessor
from system.realImagContextLSTMSystem import RealImagContextLSTMSystem
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
sessionsName = "test_nsynth2048_bs256_fft512_"
params = LstmContextInpaintingParameters(lstmSize=512,
batchSize=256,
signalLength=6144,
inputChannels=4,
gapLength=2048,
fftWindowLength=512,
fftHopSize=128)
contextArchitecture = RealImagContextEncoderLSTMArchitecture(params)
aPreProcessor = PreAndPostProcessor(params)
aContextEncoderSystem = RealImagContextLSTMSystem(contextArchitecture,
aPreProcessor, params,
sessionsName)
aContextEncoderSystem.train(
"../../aDataset/nsynth_train_w6144_g2048_h512.tfrecords",
"../../aDataset/nsynth_valid_w6144_g2048_h512.tfrecords", 1e-4)
import pickle
from architecture.contextEncoderArchitecture import ContextEncoderArchitecture
from system.contextEncoderSystem import ContextEncoderSystem
from system.preAndPostProcessor import PreAndPostProcessor
architecturesParametersFile = "complex_network_parameters.pkl"
sessionsName = "complex_network"
with open(architecturesParametersFile, 'rb') as savedFile:
Context_Encoder_parameters = pickle.load(savedFile)
aContextEncoderArchitecture = ContextEncoderArchitecture(*Context_Encoder_parameters.architectureParameters())
aPreProcessor = PreAndPostProcessor(*Context_Encoder_parameters.preProcessorParameters())
aContextEncoderSystem = ContextEncoderSystem(aContextEncoderArchitecture, Context_Encoder_parameters.batchSize(),
aPreProcessor, sessionsName)
aContextEncoderSystem.train("nsynth_train_w5120_g1024_h512.tfrecords", "nsynth_valid_w5120_g1024_h512.tfrecords", 1e-3)