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 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 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
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)
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)
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)
from architecture.contextEncoderLSTMArchitecture import ContextEncoderLSTMArchitecture
from architecture.parameters.lstmContextInpaintingParameters import LstmContextInpaintingParameters
from system.lstmSystem import LSTMSystem
import os
from system.preAndPostProcessor import PreAndPostProcessor
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
sessionsName = "test_ContextLSTM_"
params = LstmContextInpaintingParameters(lstmSize=512,
batchSize=64,
signalLength=5120,
inputChannels=4,
gapLength=1024,
fftWindowLength=512,
fftHopSize=128)
contextArchitecture = ContextEncoderLSTMArchitecture(params)
aPreProcessor = PreAndPostProcessor(params)
aContextEncoderSystem = LSTMSystem(contextArchitecture, aPreProcessor, params,
sessionsName)
aContextEncoderSystem.train(
"../variationalAutoEncoder/fake_w5120_g1024_h512.tfrecords",
"../variationalAutoEncoder/fake_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_nsynth4096_bs256_fft256_"
params = LstmContextInpaintingParameters(lstmSize=512,
batchSize=256,
signalLength=12288,
inputChannels=4,
gapLength=4096,
fftWindowLength=256,
fftHopSize=64)
contextArchitecture = RealImagContextEncoderLSTMArchitecture(params)
aPreProcessor = PreAndPostProcessor(params)
aContextEncoderSystem = RealImagContextLSTMSystem(contextArchitecture,
aPreProcessor, params,
sessionsName)
aContextEncoderSystem.train(
"../../aDataset/nsynth_train_w12288_g4096_h512.tfrecords",
"../../aDataset/nsynth_valid_w12288_g4096_h512.tfrecords", 1e-3)