image_size = [28,28]
train_imgs,train_lbls,test_imgs,test_lbls = read_mnist(dim=image_size, n_train=60000, n_test=10000)
''' deep- belief net '''
print ("\nStarting a Deep Belief Net..")
batch_size = 20
dbn = DeepBeliefNet(sizes={"vis":image_size[0]*image_size[1], "hid":500, "pen":500, "top":2000, "lbl":10},
image_size=image_size,
n_labels=10,
batch_size=batch_size
)
''' greedy layer-wise training '''
epochs = 10
iterations = int(60000*epochs/batch_size)
dbn.train_greedylayerwise(vis_trainset=train_imgs, lbl_trainset=train_lbls, n_iterations=iterations)
#dbn.recognize(train_imgs, train_lbls)
dbn.recognize(test_imgs, test_lbls)
for digit in range(10):
digit_1hot = np.zeros(shape=(1,10))
digit_1hot[0,digit] = 1
dbn.generate(digit_1hot, name="rbms")
epochs = 1
iterations = int(60000 * epochs / batch_size)
dbn.train_greedylayerwise(vis_trainset=train_imgs,
lbl_trainset=train_lbls,
n_iterations=iterations)
#dbn.recognize(train_imgs, train_lbls)
#dbn.recognize(test_imgs, test_lbls)
#for digit in range(10):
# digit_1hot = np.zeros(shape=(1,10))
# digit_1hot[0,digit] = 1
# dbn.generate(digit_1hot, name="rbms")
''' fine-tune wake-sleep training '''
dbn.train_wakesleep_finetune(vis_trainset=train_imgs,
lbl_trainset=train_lbls,
n_iterations=iterations)
dbn.recognize(train_imgs, train_lbls)
dbn.recognize(test_imgs, test_lbls)
for digit in range(10):
digit_1hot = np.zeros(shape=(1, 10))
digit_1hot[0, digit] = 1
dbn.generate(digit_1hot, name="dbn")
dbn.train_greedylayerwise(vis_trainset=train_imgs,
lbl_trainset=train_lbls,
n_iterations=20,
verbose=True)
# print("Classify training set")
# dbn.recognize(train_imgs, train_lbls)
#
# print("Classify test set")
# dbn.recognize(test_imgs, test_lbls)
print("Image generation")
for digit in range(10):
digit_1hot = np.zeros(shape=(1, 10))
digit_1hot[0, digit] = 1
dbn.generate(digit_1hot, name="rbm")
''' fine-tune wake-sleep training '''
dbn.train_wakesleep_finetune(vis_trainset=train_imgs,
lbl_trainset=train_lbls,
n_iterations=10)
# dbn.recognize(train_imgs, train_lbls)
# dbn.recognize(test_imgs, test_lbls)
for digit in range(50):
digit_1hot = np.zeros(shape=(1, 10))
digit_1hot[0, 6] = 1
dbn.generate(digit_1hot, name="dbn" + str(digit))
#plt.legend()
#plt.savefig("pictures/4_2_recon_losses.png")
# for name, im, lb in [("train", train_imgs, train_lbls), ("test", test_imgs, test_lbls)]:
# acc = []
# for trials in range(10):
# acc.append(dbn.recognize(train_imgs, train_lbls))
# print (name, acc[-1])
# print (f"{name} & {np.mean(acc):.5f} & {np.std(acc):.5f}")
### DO IMAGE GEN HERE!
labels = np.zeros([1, 10])
labels[0, 0] = 1
vis = dbn.generate(labels)
vis = vis.reshape(28, 28)
plt.clf()
plt.subplot(4, 5, 1)
plt.imshow(vis)
plt.show(block=False)
plt.pause(0.1)
for it in range(1, 10):
labels[0, it] = 1
labels[0, it - 1] = 0
print(labels)
vis = dbn.generate(labels)
vis = vis.reshape(28, 28)
plt.subplot(4, 5, it + 1)
plt.imshow(vis)
plt.show(block=False)
batch_size=20)
''' greedy layer-wise training '''
dbn.train_greedylayerwise(vis_trainset=train_imgs,
lbl_trainset=train_lbls,
n_iterations=90000)
# dbn.recognize(train_imgs[:1000, :], train_lbls[:1000, :])
# dbn.recognize(test_imgs, test_lbls)
imgs = []
for digit in range(10):
digit_1hot = np.zeros(shape=(1, 10))
digit_1hot[0, digit] = 1
imgs.append(dbn.generate(digit_1hot, name="rbms",
init_random_pen=True))
plot_generated(np.array(imgs), np.array([i for i in range(10)]))
imgs = []
for digit in range(10):
digit_1hot = np.zeros(shape=(1, 10))
digit_1hot[0, digit] = 1
imgs.append(
dbn.generate(digit_1hot, name="rbms", init_random_pen=False))
plot_generated(np.array(imgs),
np.array([i for i in range(10)]),
dirname='rand_img')
# ''' fine-tune wake-sleep training '''
# dbn.train_wakesleep_finetune(
