def load_model(self, clicked):
s_m = str(self.cb_model.currentText())
# Fill up settings from model string
if 'Mnist' in s_m:
self.dataset = MNIST()
self.z_dim = 5
self.y_dim = 10
self.image_size = 28
self.image_channels = 1
elif 'CelebBig' in s_m:
self.dataset = CelebBig()
self.z_dim = 128
self.y_dim = 40
self.image_size = 128
self.image_channels = 3
elif 'Celeb' in s_m:
self.dataset = CelebA()
self.z_dim = 50
self.y_dim = 40
self.image_size = 32
self.image_channels = 3
elif 'Cell' in s_m:
self.dataset = Cell()
self.z_dim = 50
self.y_dim = None
self.image_size = 64
self.image_channels = 1
if 'noy' in s_m:
self.y_dim = None
if 'Mnist_Dense' in s_m:
model_class = ModelDenseMnist
elif 'Mnist_Conv' in s_m:
model_class = ModelConvMnist
elif 'Celeb_Conv' in s_m:
model_class = ModelConv32
elif 'Celeb_Subpix' in s_m:
model_class = ModelSubpix32
self.y_dim = None
elif 'CelebBig' in s_m:
model_class = ModelConv128
elif 'CelebBig' in s_m:
model_class = ModelConv128
elif 'Cell' in s_m:
model_class = ModelConv64
model = model_class(batch_size=self.batch_size,
z_dim=self.z_dim,
y_dim=self.y_dim,
is_training=False)
#
# if 'WGan' in s_m:
# self.solver = AaeWGanSolver(model=model)
# elif 'Gan' in s_m:
# self.solver = AaeGanSolver(model=model)
# else:
self.solver = AaeSolver(model=model)
self.sess.run(tf.global_variables_initializer())
# Saver
saver = tf.train.Saver()
# Restore previous
saver.restore(self.sess, s_m)
self.build_interface()
def manifold_images(model, name, gan, y=False):
# Solver
if gan == 'WGan':
print("WGan Solver")
solver = AaeWGanSolver(model=model)
gan = 'WGan_'
elif gan == 'Gan':
print("Gan Solver")
solver = AaeGanSolver(model=model)
gan = 'Gan_'
else:
solver = AaeSolver(model=model)
sess = tf.Session()
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver()
# To restore previous
print("Restoring model")
saver.restore(sess, 'models/model_%s%s.ckpt' % (gan, name))
print("Model restored")
r = 1.8
# Generate grid of z
ls = np.linspace(-r, r, samples)
lf = np.linspace(-3, 3, frames)
z = np.array(list(product(lf, ls, ls, ls, ls)))
if not os.path.exists('output/z_tmp'):
os.makedirs('output/z_tmp')
if y:
rng = 10
else:
rng = 1
for z_i in range(5):
z_t = np.roll(z, z_i, axis=1)
for code_v in range(rng):
code = [[0, 0, 0, 0, 0, 0, 0, 0, 0, 0]]
code[0][code_v] = 1
size = (samples**4 * frames)
y_lab = np.reshape(code * size, [size, 10])
img = sess.run(solver.x_from_z,
feed_dict={
solver.z_provided: z_t,
solver.y_labels: y_lab
})
files = []
w = 28 * samples**2 + (samples - 1) * 10
h = (28 * samples**2 + (samples - 1) * 10) * frames
b_i_canv = Image.new('L', (w, h), color=1)
b_i_apng = Image.new('L', (w, w), color=1)
ls = range(samples)
lf = range(frames)
prod = list(product(lf, ls, ls, ls, ls))
for i, p in enumerate(prod):
index = samples ** 4 * p[0] + samples ** 3 * p[1] + \
samples ** 2 * p[2] + samples * p[3] + p[4]
im = img[index].reshape([28, 28])
cimg = Image.fromarray(np.uint8(im * 255))
x = p[4] * 28 + p[2] * (10 + 28 * samples)
y = p[3] * 28 + p[1] * (10 + 28 * samples)
b_i_apng.paste(cimg, (x, y))
y_p = y + p[0] * (28 * samples**2 + (samples - 1) * 10)
b_i_canv.paste(cimg, (x, y_p))
if (i + 1) % samples**4 == 0:
file = "output/z_tmp/Res_%d.png" % (i // samples**4)
files.append(file)
b_i_apng.save(file)
ap = APNG()
# Create animated png
for file in files:
ap.append(file, delay=100)
for file in files[::-1]:
ap.append(file, delay=100)
if not os.path.exists('output/z_%d' % z_i):
os.makedirs('output/z_%d' % z_i)
ap.save("output/z_%d/%s_%d.apng" % (z_i, name, code_v))
b_i_canv.save("output/z_%d/%s_%d.png" % (z_i, name, code_v))
def compare_style(model, name, gan, y=False):
# Solver
if gan == 'WGan':
print("WGan Solver")
solver = AaeWGanSolver(model=model)
gan = 'WGan_'
elif gan == 'Gan':
print("Gan Solver")
solver = AaeGanSolver(model=model)
gan = 'Gan_'
else:
solver = AaeSolver(model=model)
sess = tf.Session()
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver()
# To restore previous
print("Restoring model")
saver.restore(sess, 'models/model_%s%s.ckpt' % (gan, name))
print("Model restored")
b_i = []
for i in range(91):
b_i.append(Image.new('L', (28 * 5, 28 * 5)))
ys = []
for i in range(9):
y = [0] * 10
y[i] = 1.0
ys.append(np.array(y).reshape([1, 10]))
for j in range(1, 10):
y = [0] * 10
y[i] = 1 - np.sin(np.pi / 20 * j)
y[i + 1] = np.sin(np.pi / 20 * j)
ys.append(np.array(y).reshape([1, 10]))
y = [0] * 10
y[9] = 1.0
ys.append(np.array(y).reshape([1, 10]))
y = np.concatenate(ys, axis=0)
for i in range(25):
z = np.random.uniform(-1.5, 1.5, size=[1, 5])
z = np.tile(z, [91, 1])
img = sess.run(solver.x_from_z,
feed_dict={
solver.z_provided: z,
solver.y_labels: y
})
for j in range(91):
im = img[j].reshape([28, 28])
cimg = Image.fromarray(np.uint8(im * 255))
x = 28 * (i % 5)
x2 = 28 * (i // 5)
b_i[j].paste(cimg, (x, x2))
if not os.path.exists('output/z_tmp'):
os.makedirs('output/z_tmp')
if not os.path.exists('output/style'):
os.makedirs('output/style')
ap = APNG()
files = []
for i in range(91):
file = "output/z_tmp/Res_%d.png" % i
b_i[i].save(file)
files.append(file)
# Create animated png
for file in files:
ap.append(file, delay=200)
ap.save("output/style/%s.apng" % name)
# Return solution
return Y
if __name__ == "__main__":
print(
"Run Y = tsne.tsne(X, no_dims, perplexity) to perform t-SNE on your dataset."
)
print("Running example on 2,500 MNIST digits...")
print("Draw Mnist Dense no y")
model = ModelDenseMnist(batch_size=128,
z_dim=5,
y_dim=None,
is_training=False)
solver = AaeSolver(model)
sess = tf.Session()
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver()
# To restore previous
print("Restoring model")
saver.restore(sess, 'models/model_Mnist_Dense_noy.ckpt')
print("Model restored")
z = []
y = []
pred_out = []
data = MNIST()
i = 100
def plot_samples(model, data, name):
# Solver
solver = AaeSolver(model=model)
# Session
config = tf.ConfigProto(device_count={'GPU': 0})
sess = tf.Session(config=config)
sess.run(tf.global_variables_initializer())
# Saver
saver = tf.train.Saver()
# To restore previous
print("Restoring model")
saver.restore(sess, 'models/model_%s.ckpt' % name)
print("Model restored")
z = []
y = []
pred_out = []
i = 500
for batch_x, batch_y in data.iterate_minibatches(model.batch_size,
shuffle=True):
z_enc, y_enc = sess.run([solver.z_encoded, solver.y_pred_enc],
feed_dict={
solver.x_image: batch_x,
solver.y_labels: batch_y
})
z.append(z_enc)
pred_out.extend(y_enc)
y.extend(batch_y)
if i == 0:
break
i -= 1
z = np.concatenate(z, axis=0)
f, axarr = plt.subplots(model.z_dim, model.z_dim)
for j in range(model.z_dim):
for i in range(model.z_dim):
ax = axarr[i, j]
ax.set_title('Axis [%d, %d]' % (i, j))
ax.scatter(z[:, j], z[:, i], c=pred_out)
if i != model.z_dim - 1:
plt.setp(ax.get_xticklabels(), visible=False)
if j != 0:
plt.setp(ax.get_yticklabels(), visible=False)
plt.show()
y = np.argmax(y, axis=1)
f, axarr = plt.subplots(model.z_dim, model.z_dim)
for i in range(model.z_dim):
for j in range(model.z_dim):
ax = axarr[i, j]
ax.set_title('Axis [%d, %d]' % (i, j))
ax.scatter(z[:, i], z[:, j], c=y, s=1)
if i != model.z_dim - 1:
plt.setp(ax.get_xticklabels(), visible=False)
if j != 0:
plt.setp(ax.get_yticklabels(), visible=False)
ax.set_ylim([-3, 3])
ax.set_xlim([-3, 3])
ax.set_autoscalex_on(False)
ax.set_autoscaley_on(False)
plt.show()
def draw_images(model, name, systematic=True, gan=''):
# Solver
if gan == 'WGan':
print("WGan Solver")
solver = AaeWGanSolver(model=model)
gan = 'WGan_'
elif gan == 'Gan':
print("Gan Solver")
solver = AaeGanSolver(model=model)
gan = 'Gan_'
else:
solver = AaeSolver(model=model)
sess = tf.Session()
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver()
# To restore previous
print("Restoring model")
saver.restore(sess, 'models/model_%s%s.ckpt' % (gan, name))
print("Model restored")
z = []
# If systematic then all animations pass through origin
# If random then they pass through respective axis
if systematic:
name += 'sys'
a = np.zeros([50, 50])
else:
name += 'rand'
a = np.random.uniform(-1, 1, [50, 50])
# We use sinus so that we can see edge images for longer
for v in np.linspace(-np.pi / 2, np.pi / 2, frames):
for i in range(model.z_dim):
b = np.reshape(np.copy(a[i, :]), [1, 50])
b[0][i] = 2.5 * np.sin(v)
z.append(b)
z = np.concatenate(z)
y = np.zeros([model.z_dim * frames, 1])
img = sess.run(solver.x_from_z,
feed_dict={
solver.z_provided: z,
solver.y_labels: y
})
files = []
w = 32 * 10
h = 32 * 5
b_i_apng = Image.new('RGB', (w, h), color=1)
b_i_canv = Image.new('RGB', (w, h * frames), color=1)
for f in range(frames):
for x in range(10):
for y in range(5):
index = f * 50 + x + y * 10
im = CelebA.transform2display(img[index])
cimg = Image.fromarray(np.uint8(1 + im * 254))
b_i_apng.paste(cimg, (x * 32, y * 32))
b_i_canv.paste(cimg, (x * 32, y * 32 + f * 32 * 5))
file = "output/celeb/Res_%d.png" % f
files.append(file)
# blank_image = blank_image.resize((128*10, 128*5))
b_i_apng.save(file)
ap = APNG()
# Create animated png
for file in files:
ap.append(file, delay=50)
for file in files[::-1]:
ap.append(file, delay=50)
if not os.path.exists('output/celeb'):
os.makedirs('output/celeb')
ap.save("output/celeb/%s.apng" % name)
b_i_canv.save("output/celeb/%s.png" % name)