np.random.seed(int(sys.argv[-1]) + 10)
Ds = [1, 12, 64]
R, BS = 110, 150
sigma_x = .1
model = networks.create_fns(BS, R, Ds, 1, lr=0.001, leakiness=0.1, var_x=sigma_x**2)
DATA = datasets.load_digits(n_class=1).images
DATA = DATA[:BS].reshape((BS, -1)) + np.random.randn(BS, Ds[-1]) * 0.1
#DATA -= DATA.mean(1, keepdims=True)
DATA /= (DATA.max() * 10)
print(DATA)
L = []
for iter in tqdm(range(36)):
plt_state()
plt.savefig('samples_{}.png'.format(iter))
plt.close()
L.append(networks.EM(model, DATA, 1, 10, pretrain=iter==0, update_var=iter>5))
print(L[-1])
plt.figure()
plt.plot(np.concatenate(L), lw=3)
plt.savefig('NLL.png')
plt.close()
Ds,
1,
lr=0.001,
leakiness=args.leakiness,
var_x=args.std**2)
for RUN in range(20):
# do the VAE case
if args.model != 'EM':
filename = 'nnnnsaving_likelihood_{}_{}_{}_{}_{}_{}.npz'
for lr_ in [0.005, 0.001, 0.0001]:
graph.reset()
lr.update(lr_)
out = networks.EM(model,
DATA,
epochs=args.epochs,
n_iter=500,
extra=emt)
np.savez(filename.format(args.dataset, args.epochs, args.model,
lr_, args.network, RUN),
L=out[0],
LL=out[1],
samples=model['sample'](4 * BS),
noise=np.random.randn(4 * BS, 2) *
np.sqrt(model['varx']()),
data=DATA)
else:
graph.reset()
filename = 'nnnsaving_likelihood_{}_{}_{}_{}_{}.npz'
out = networks.EM(model,
DATA,
plt.subplot(10, 10, 1 + i)
plt.imshow(predictions[i].reshape((8, 8)))
for i in range(50):
plt.subplot(10, 10, 51 + i)
plt.imshow(DATA[i].reshape((8, 8)))
np.random.seed(int(sys.argv[-1]) + 10)
Ds = [1, 20, 64]
R, BS = 110, 150
model = networks.create_fns(BS, R, Ds, 0, var_x=np.ones(Ds[-1]), lr=0.001)
DATA = datasets.load_digits(n_class=1).images
DATA = DATA[:BS].reshape((BS, -1))
DATA /= (0.1 + DATA.max(1, keepdims=True))
DATA -= DATA.mean(1, keepdims=True)
DATA += np.random.randn(BS, Ds[-1]) * 0.3
L = []
for iter in tqdm(range(16)):
L.append(networks.EM(model, DATA, 100))
plt_state()
plt.savefig('samples_{}.png'.format(iter))
plt.close()
plt.figure()
plt.plot(np.concatenate(L), lw=3)
plt.savefig('NLL.png')
plt.close()
#DATA = np.vstack([DATA * np.cos(DATA), DATA * np.sin(DATA)]).T
DATA += np.random.randn(BS, Ds[-1]) * 0.1
DATA -= DATA.mean(0)
DATA /= DATA.max(0)
DATA /= 1
#DATA *= 3
#DATA += 2
L = []
for iter in tqdm(range(550)):
L.append(
networks.EM(model,
DATA,
1,
min((iter + 1) * 20, 1400),
pretrain=iter == args.pretrain,
update_var=iter > 1))
# L.append(networks.EM(model, DATA, 1, 1400, pretrain=iter==args.pretrain, update_var=iter>2))
# print(L[-1])
# print(L[-1][0], L[-1][-1])
#
X = model['sample'](BS)
noise = np.random.randn(*X.shape) * np.sqrt(model['varx']())
plt.figure(figsize=(12, 6))
plt.subplot(121)
plt.scatter(DATA[:, 0], DATA[:, 1])
plt.scatter(X[:, 0], X[:, 1])
plt.scatter(X[:, 0] + noise[:, 0], X[:, 1] + noise[:, 1])
plt.subplot(122)
plt.plot(np.concatenate(L))