def test_VAE_saveState(self):
mb_size = 49 # because dummyDataset98.npz is a 98 data size
epoch_nb = 11
datasetName = 'dummyDataset98.npz'
datasetDir = './dummyDataset/'
testDataset = NPZ_Dataset(datasetName, datasetDir, 'Spectrums')
train_loader = torch.utils.data.DataLoader(testDataset,
batch_size=mb_size,
shuffle=True)
# define vae structure
X_dim = 1024
Z_dim = 6
IOh_dims_Enc = [X_dim, 401, Z_dim]
IOh_dims_Dec = [Z_dim, 399, X_dim]
NL_types_Enc = ['relu6']
NL_types_Dec = ['relu6', 'sigmoid']
vae = VAE(X_dim, Z_dim, IOh_dims_Enc, IOh_dims_Dec, NL_types_Enc,
NL_types_Dec, mb_size)
vae.trainVAE(train_loader, epoch_nb)
# save it
if vae.trained:
vae.save(datasetName, datasetDir)
self.assertTrue(vae.created and vae.trained and vae.saved)
def vae_train():
# load train
x_train = load_dataset('adi')
# load test normal and anomaly
test_anomaly = load_dataset('vans')
test_normal = load_dataset('adi_test')
# drfine train and valid iamge for train
trains = x_train[10:]
valid = x_train[:10]
print(trains.shape, valid.shape, test_anomaly.shape, test_normal.shape)
# try to plot
plt.imshow(x_train[10].reshape(256, 256))
plt.gray()
plt.show()
# train
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
def step_decay(epoch):
initial_lrate = 0.0001
decay_rate = 0.5
decay_steps = 8.0
lrate = initial_lrate * math.pow(decay_rate,
math.floor((1 + epoch) / decay_steps))
return lrate
callback = []
callback.append(
HistoryCheckpoint(filepath='tb/LearningCurve_{history}.png',
verbose=1,
period=300))
callback.append(LearningRateScheduler(step_decay))
model = VAE()
model, loss = model.vae_net()
#model.load_weights("vae_model.h5")
model.add_loss(loss)
model.compile(optimizer=Adam(lr=0.0001))
model.summary()
try:
model.fit(trains,
batch_size=20,
epochs=300,
callbacks=callback,
validation_data=(valid, None))
finally:
model.save('weight/vae_model.h5')
def test_gaussianVAE_trainsaveload(self):
mb_size = 49 # because dummyDataset98.npz is a 98 data size
epoch_nb = 5
# if exists remove 'saveloadTest' folder
if os.path.exists('./saveloadTest'):
shutil.rmtree('./saveloadTest')
# create a VAE
X_dim = 1024
Z_dim = 10
IOh_dims_Enc = [X_dim, 600, Z_dim]
IOh_dims_Dec = [Z_dim, 600, X_dim]
NL_types_Enc = ['relu6']
NL_types_Dec = ['relu6']
vae = VAE(X_dim,
Z_dim,
IOh_dims_Enc,
IOh_dims_Dec,
NL_types_Enc,
NL_types_Dec,
mb_size,
bernoulli=False,
gaussian=True)
# prepare dataset
datasetName = 'dummyDataset98.npz'
datasetDir = './dummyDataset/'
saveDir = './dummySaveTest/'
testDataset = NPZ_Dataset(datasetName, datasetDir, 'Spectrums')
train_loader = torch.utils.data.DataLoader(testDataset,
batch_size=mb_size,
shuffle=True)
# train it for 10 epochs
vae.trainVAE(train_loader, epoch_nb)
# save it
savefile = vae.save(datasetName, saveDir)
# reload the savefile of VAE
vae = loadVAE(savefile, saveDir)
# continue training
# vae.trainVAE(train_loader, 10)
# vae.save(datasetName, datasetDir)
self.assertTrue(vae.created and vae.loaded and vae.saved)
if args.type == 'bernoulli':
vae = VAE(X_dim, Z_dim, IOh_dims_Enc,
IOh_dims_Dec, NL_types_Enc, NL_types_Dec, mb_size,
bernoulli=True, gaussian=False, beta=beta, Nwu=Nwu, noiseIn=noise, noiseGain=noiseGain)
elif args.type == 'gaussian':
vae = VAE(X_dim, Z_dim, IOh_dims_Enc,
IOh_dims_Dec, NL_types_Enc, NL_types_Dec, mb_size,
bernoulli=False, gaussian=True, beta=beta, Nwu=Nwu, noiseIn=noise, noiseGain=noiseGain)
else:
print("ERROR script: Chose VAE type -> either bernoulli or gaussian")
# train it for N epochs
vae.trainVAE(train_loader, epoch_nb)
# save it
vae.save(datasetName, saveDir)
elif mode == "load":
# get savefile path
directory = args.vae_path
savefile = directory.split("/")[-1]
directory = directory.replace(savefile, "")
# load vae
vaeLoaded = loadVAE(savefile, directory)
# Now do some stuff (e.g. PCA, t-sne, generate ...)
# Here for example it generates 200 images. Each image is 200x1024.
# It corresponds to 200 stacked spectra (of size 1024).
# Hence, x axis corresponds to frequency while y axis is a latent dimension varying from -10 to 10.
# When changing image it varies another latent dimension also from -10 to 10.
