def predict(cfs):
dataset = MNIST(filenames=cfs)
net = LSGAN(filenames=cfs)
net.define_net()
print(net.pretty_settings())
net.load('AutoEncoder')
x = net.gen_latent()
p = net.predict('Decoder', [x])
imgs = dataset.visualize(p[0])
subplot_images((imgs, ), is_gray=True, size=3.0, tight_c=0.5)
def predict():
dataset = MNIST(**data_settings)
if net_name == 'ae':
net = AutoEncoder1D(**net_settings)
elif net_name == 'vae':
net = VAE1D(**net_settings)
net.define_net()
print(net.pretty_settings())
net.load('AutoEncoder')
s = next(dataset)
p = net.predict('Decoder', [net.gen_latent()])
imgs = dataset.visualize(p[0])
subplot_images((imgs, ), is_gray=True, size=3.0, tight_c=0.5)
def train(filenames=None, settings=None, **kwargs):
dataset = MNIST(filenames=filenames)
print("=" * 30)
print("DATASET SETTINGS:")
print(dataset.pretty_settings())
print("=" * 30)
net = AAE1D(filenames=filenames)
net.define_net()
print("=" * 30)
print("NETWORK SETTINGS:")
print(net.pretty_settings())
print("=" * 30)
nb_batches = settings['nb_batches']
net.load()
ptp = ProgressTimer(nb_batches)
# for i in range(nb_batches // 3 * 2):
# loss_ae = train_ae(net, dataset)
# msg = 'T:AuE, loss=%05f' % (loss_ae)
# ptp.event(net.step, msg)
# net.lr_decay()
# for i in range(nb_batches // 3):
# loss_ae = train_ae(net, dataset)
# msg = 'T:AuE, loss=%05f' % (loss_ae)
# ptp.event(net.step, msg)
for i in range(nb_batches // 3):
loss_cri = train_cri(net, dataset)
msg = '|T:Cri, loss=%05f|' % (loss_cri)
loss_gen = train_gen(net, dataset)
msg += '|T:Gen, loss=%05f|' % (loss_gen)
loss_ae = train_ae(net, dataset)
msg = '|T:AuE, loss=%05f|' % (loss_ae)
ptp.event(net.step, msg)
ptp = ProgressTimer(nb_batches)
# for i in range(nb_batches // 2):
# loss_ae = train_ae(net, dataset)
# msg = 'step #%5d, AuE, loss=%05f' % (net.step, loss_ae)
# ptp.event(net.step, msg)
# net.lr_decay()
# for i in range(nb_batches // 2):
# loss_ae = train_ae(net, dataset)
# msg = 'step #%5d, AuE, loss=%05f' % (net.step, loss_ae)
# ptp.event(net.step, msg)
# for i in range(nb_batches):
# loss_cri = train_cri(net, dataset)
# msg = 'Cri, loss=%05f ' % (loss_cri)
# loss_gen = train_gen(net, dataset)
# msg += 'Gen, loss=%05f ' % (loss_gen)
# ptp.event(net.step, msg)
net.save('net', is_print=True)
def show_latent(cfs, nb_sample=10000):
print("AE1D Test. Show latent called.")
net = AAE1D(filenames=cfs)
net.define_net()
net.load('enc')
nb_batch_show = nb_sample // net._batch_size
dataset = MNIST(filenames=cfs)
latents = []
for i in range(10):
latents.append([])
for i in tqdm(range(nb_batch_show)):
s = next(dataset)
p = net.predict('enc', [s[0]])
for j in range(dataset._batch_size):
latents[s[1][j]].append(p[0][j])
x = []
y = []
for i in range(10):
pos = np.array(latents[i])
x.append(pos[:, 0])
y.append(pos[:, 1])
para = []
for i in range(10):
para.append(x[i])
para.append(y[i])
para.append('.')
plt.plot(*para)
plt.legend(list(map(str, range(10))))
def show_latent(nb_sample=10000):
print("AE1D Test. Show latent called.")
if net_name == 'ae':
net = AutoEncoder1D(**net_settings)
elif net_name == 'vae':
net = VAE1D(**net_settings)
net.define_net()
net.load('AutoEncoder')
nb_batch_show = nb_sample // batch_size
dataset = MNIST(**data_settings)
latents = []
for i in range(10):
latents.append([])
for i in tqdm(range(nb_batch_show)):
s = next(dataset)
p = net.predict('Encoder', [s[0]])
for j in range(batch_size):
latents[s[1][j]].append(p[0][j])
x = []
y = []
for i in range(10):
pos = np.array(latents[i])
x.append(pos[:, 0])
y.append(pos[:, 1])
para = []
for i in range(10):
para.append(x[i])
para.append(y[i])
para.append('.')
plt.plot(*para)
plt.legend(list(map(str, range(10))))
def train(is_print_loss=False):
dataset = MNIST(**data_settings)
if net_name == 'ae':
net = AutoEncoder1D(**net_settings)
elif net_name == 'vae':
net = VAE1D(**net_settings)
net.define_net()
# net.load()
print(net.pretty_settings())
for i in tqdm(range(nb_batches), ascii=True, ncols=50):
s = next(dataset)
loss_v = net.train_on_batch('AutoEncoder', [s[0]], [s[0]])
if is_print_loss:
print(' loss = ', loss_v)
if i % 1000 == 0:
net.save('AutoEncoder')
# for i in tqdm(range(nb_batches), ascii=True, ncols=50):
# net.reset_lr([1e-3])
# s = next(dataset)
# loss_v = net.train_on_batch('AutoEncoder', [s[0]], [s[0]])
# if is_print_loss:
# print(' loss = ', loss_v)
# if i % 1000 == 0:
# net.save('AutoEncoder')
# for i in tqdm(range(nb_batches), ascii=True, ncols=50):
# net.reset_lr([1e-3])
# s = next(dataset)
# loss_v = net.train_on_batch('AutoEncoder', [s[0]], [s[0]])
# if is_print_loss:
# print(' loss = ', loss_v)
# if i % 1000 == 0:
# net.save('AutoEncoder')
net.save('AutoEncoder', is_print=True)
def train():
dataset = MNIST(**data_settings)
net = WGAN1D(**net_settings)
net.define_net()
for i in tqdm(range(nb_batches)):
z = net.gen_latent()
s = next(dataset)
for j in range(5):
net.train_on_batch('Cri', [s[0], z])
z = net.gen_latent()
net.train_on_batch('Gen', [z])
def show_mainfold(cfs):
dataset = MNIST(filenames=cfs)
nb_axis = int(np.sqrt(dataset._batch_size))
x = np.linspace(-1.5, 1.5, nb_axis)
y = np.linspace(-1.5, 1.5, nb_axis)
pos = np.meshgrid(x, y)
xs = pos[0]
ys = pos[1]
xs = xs.reshape([-1])
ys = ys.reshape([-1])
net = LSGAN(filenames=cfs)
net.define_net()
net.load('Gen')
latents = np.array([xs, ys]).T
p = net.predict('Gen', [latents])
imgs = dataset.visualize(p[0])
subplot_images((imgs, ),
nb_max_row=nb_axis,
is_gray=True,
size=1.0,
tight_c=0.5)
def show_mainfold():
dataset = MNIST(**data_settings)
nb_axis = int(np.sqrt(batch_size))
x = np.linspace(-1.5, 1.5, nb_axis)
y = np.linspace(-1.5, 1.5, nb_axis)
pos = np.meshgrid(x, y)
xs = pos[0]
ys = pos[1]
xs = xs.reshape([-1])
ys = ys.reshape([-1])
if net_name == 'ae':
net = AutoEncoder1D(**net_settings)
elif net_name == 'vae':
net = VAE1D(**net_settings)
net.define_net()
net.load('AutoEncoder')
latents = np.array([xs, ys]).T
p = net.predict('Decoder', [latents])
imgs = dataset.visualize(p[0])
subplot_images((imgs, ), nb_max_row=nb_axis,
is_gray=True, size=1.0, tight_c=0.5)
def train(nb_batches, cfs):
dataset = MNIST(filenames=cfs)
print("=" * 30)
print("DATASET SETTINGS:")
print(dataset.pretty_settings())
print("=" * 30)
net = LSGAN(filenames=cfs)
net.define_net()
print("=" * 30)
print("NETWORK SETTINGS:")
print(net.pretty_settings())
print("=" * 30)
ptp = ProgressTimer(net.pre_train)
for i in range(net.pre_train):
s = next(dataset)
z = net.gen_latent()
loss_c = net.train_on_batch('Cri', [s[0], z], [])
msg = 'loss_c= %f' % loss_c
ptp.event(i, msg)
pt = ProgressTimer(nb_batches)
loss_c = np.nan
loss_g = np.nan
for i in range(nb_batches):
s = next(dataset)
z = net.gen_latent()
if i % net.gen_freq > 0:
loss_c = net.train_on_batch('Cri', [s[0], z], [])
msg = 'c_step, loss_c= %f; loss_g= %f' % (loss_c, loss_g)
pt.event(i, msg)
else:
loss_g = net.train_on_batch('WGan', [s[0], z], [])
msg = 'g_step, loss_c= %f; loss_g= %f' % (loss_c, loss_g)
pt.event(i, msg)
if i % 1000 == 0:
net.save('AutoEncoder')
net.save('AutoEncoder', is_print=True)
def show_data_mainfold(cfs):
""" show latent main fold for data """
dataset = MNIST(filenames=cfs)
# nb_axis = int(np.sqrt(dataset._batch_size))
nb_axis = 32
x = np.linspace(-5.0, 20.0, nb_axis)
y = np.linspace(-5.0, 20.0, nb_axis)
pos = np.meshgrid(x, y)
xs = pos[0]
ys = pos[1]
xs = xs.reshape([-1])
ys = ys.reshape([-1])
net = AAE1D(filenames=cfs)
net.define_net()
net.load('Gen')
latents = np.array([xs, ys]).T
pall = None
nb_latents = latents.shape[0]
nb_batches = int(np.ceil(nb_latents / net.batch_size))
nb_pad = nb_batches * net.batch_size - nb_latents
latents_pad = np.pad(latents, ((0, nb_pad), (0, 0)), mode='constant')
for i in tqdm(range(nb_batches)):
data_batch = latents_pad[i * net.batch_size:(i + 1) *
net.batch_size, :]
p = net.predict('Gen', [data_batch])
if pall is None:
pall = p[0]
else:
pall = np.concatenate((pall, p[0]))
p = pall[:nb_latents, ...]
imgs = dataset.visualize(p)
subplot_images((imgs, ),
nb_max_row=nb_axis,
is_gray=True,
size=1.0,
tight_c=0.5)
def create_dataset_net(is_load=False):
dataset = MNIST(**data_settings)
net = WGAN1D(**net_settings)
if is_load:
net.load()
return dataset, net