def create_G(DIM=64):
noise = T.matrix('noise')
generator = models_uncond.build_generator_toy(noise, nd=DIM)
Tgimgs = lasagne.layers.get_output(generator)
gen_fn = theano.function([noise],
lasagne.layers.get_output(generator,
deterministic=True))
return gen_fn, generator
def create_G(loss_type=None, discriminator=None, lr=0.0002, b1=0.5, DIM=64):
noise = T.matrix('noise')
generator = models_uncond.build_generator_toy(noise,nd=DIM)
Tgimgs = lasagne.layers.get_output(generator)
Tfake_out = lasagne.layers.get_output(discriminator, Tgimgs)
if loss_type == 'trickLogD':
generator_loss = lasagne.objectives.binary_crossentropy(Tfake_out, 1).mean()
elif loss_type == 'minimax':
generator_loss = -lasagne.objectives.binary_crossentropy(Tfake_out, 0).mean()
elif loss_type == 'ls':
generator_loss = T.mean(T.sqr((Tfake_out - 1)))
generator_params = lasagne.layers.get_all_params(generator, trainable=True)
updates_g = lasagne.updates.adam(generator_loss, generator_params, learning_rate=lr, beta1=b1)
train_g = theano.function([noise],
generator_loss,
updates=updates_g)
gen_fn = theano.function([noise],
lasagne.layers.get_output(generator,
deterministic=True))
return train_g, gen_fn, generator
def main():
# Parameters
task = 'toy'
name = '25G'
DIM=512
begin_save = 0
loss_type = ['trickLogD','minimax','ls']
nloss = 3
DATASET = '25gaussians'
batchSize = 64
ncandi = 1
kD = 1 # # of discrim updates for each gen update
kG = 1 # # of discrim updates for each gen update
ntf = 256
b1 = 0.5 # momentum term of adam
nz = 2 # # of dim for Z
niter = 4 # # of iter at starting learning rate
lr = 0.0001 # initial learning rate for adam G
lrd = 0.0001 # initial learning rate for adam D
N_up = 100000
save_freq = 10000
show_freq = 10000
test_deterministic = True
beta = 1.
GP_norm = False # if use gradients penalty on discriminator
LAMBDA = 2. # hyperparameter of GP
# Load the dataset
# MODEL D
print("Building model and compiling functions...")
# Prepare Theano variables for inputs and targets
real_imgs = T.matrix('real_imgs')
fake_imgs = T.matrix('fake_imgs')
# Create neural network model
discriminator = models_uncond.build_discriminator_toy(nd=DIM, GP_norm=GP_norm)
# Create expression for passing real data through the discriminator
real_out = lasagne.layers.get_output(discriminator, real_imgs)
# Create expression for passing fake data through the discriminator
fake_out = lasagne.layers.get_output(discriminator, fake_imgs)
# Create loss expressions
discriminator_loss = (lasagne.objectives.binary_crossentropy(real_out, 1)
+ lasagne.objectives.binary_crossentropy(fake_out, 0)).mean()
# Gradients penalty norm
if GP_norm is True:
alpha = t_rng.uniform((batchSize,1), low=0.,high=1.)
differences = fake_imgs - real_imgs
interpolates = real_imgs + (alpha*differences)
gradients = theano.grad(lasagne.layers.get_output(discriminator, interpolates).sum(), wrt=interpolates)
slopes = T.sqrt(T.sum(T.sqr(gradients), axis=(1)))
gradient_penalty = T.mean((slopes-1.)**2)
D_loss = discriminator_loss + LAMBDA*gradient_penalty
b1_d = 0.
else:
D_loss = discriminator_loss
b1_d = 0.
# Create update expressions for training
discriminator_params = lasagne.layers.get_all_params(discriminator, trainable=True)
lrtd = theano.shared(lasagne.utils.floatX(lrd))
updates_d = lasagne.updates.adam(
D_loss, discriminator_params, learning_rate=lrtd, beta1=b1_d)
lrt = theano.shared(lasagne.utils.floatX(lr))
# Fd Socre
Fd = theano.gradient.grad(discriminator_loss, discriminator_params)
Fd_score = beta*T.log(sum(T.sum(T.sqr(x)) for x in Fd))
# Compile a function performing a training step on a mini-batch (by giving
# the updates dictionary) and returning the corresponding training loss:
train_d = theano.function([real_imgs, fake_imgs],
discriminator_loss,
updates=updates_d)
# Compile another function generating some data
dis_fn = theano.function([real_imgs,fake_imgs],[(fake_out).mean(),Fd_score])
disft_fn = theano.function([real_imgs,fake_imgs],
[real_out.mean(),
fake_out.mean(),
(real_out>0.5).mean(),
(fake_out>0.5).mean(),
Fd_score])
# Finally, launch the training loop.
print("Starting training...")
desc = task + '_' + name
print desc
if not os.path.isdir('logs'):
os.mkdir(os.path.join('logs'))
f_log = open('logs/%s.ndjson'%desc, 'wb')
if not os.path.isdir('models'):
os.mkdir(os.path.join('models/'))
if not os.path.isdir('models/'+desc):
os.mkdir(os.path.join('models/',desc))
gen_new_params = []
# We iterate over epochs:
for n_updates in range(N_up):
xmb = toy_dataset(DATASET=DATASET, size=batchSize*kD)
xmb = xmb[0:batchSize*kD]
# initial G cluster
if n_updates == 0:
for can_i in range(0,ncandi):
train_g, gen_fn, generator = create_G(
loss_type=loss_type[can_i%nloss],
discriminator=discriminator, lr=lr, b1=b1, DIM=DIM)
for _ in range(0,kG):
zmb = floatX(np_rng.uniform(-1., 1., size=(batchSize, nz)))
cost = train_g(zmb)
sample_zmb = floatX(np_rng.uniform(-1., 1., size=(ntf, nz)))
gen_imgs = gen_fn(sample_zmb)
gen_new_params.append(lasagne.layers.get_all_param_values(generator))
if can_i == 0:
g_imgs_old=gen_imgs
fmb = gen_imgs[0:batchSize/ncandi*kD,:]
else:
g_imgs_old = np.append(g_imgs_old,gen_imgs,axis=0)
fmb = np.append(fmb,gen_imgs[0:batchSize/ncandi*kD,:],axis=0)
#print gen_new_params
# MODEL G
noise = T.matrix('noise')
generator = models_uncond.build_generator_toy(noise,nd=DIM)
Tgimgs = lasagne.layers.get_output(generator)
Tfake_out = lasagne.layers.get_output(discriminator, Tgimgs)
g_loss_logD = lasagne.objectives.binary_crossentropy(Tfake_out, 1).mean()
g_loss_minimax = -lasagne.objectives.binary_crossentropy(Tfake_out, 0).mean()
g_loss_ls = T.mean(T.sqr((Tfake_out - 1)))
g_params = lasagne.layers.get_all_params(generator, trainable=True)
up_g_logD = lasagne.updates.adam(g_loss_logD, g_params, learning_rate=lrt, beta1=b1)
up_g_minimax = lasagne.updates.adam(g_loss_minimax, g_params, learning_rate=lrt, beta1=b1)
up_g_ls = lasagne.updates.adam(g_loss_ls, g_params, learning_rate=lrt, beta1=b1)
train_g = theano.function([noise],g_loss_logD,updates=up_g_logD)
train_g_minimax = theano.function([noise],g_loss_minimax,updates=up_g_minimax)
train_g_ls = theano.function([noise],g_loss_ls,updates=up_g_ls)
gen_fn = theano.function([noise], lasagne.layers.get_output(
generator,deterministic=True))
else:
gen_old_params = gen_new_params
for can_i in range(0,ncandi):
for type_i in range(0,nloss):
lasagne.layers.set_all_param_values(generator, gen_old_params[can_i])
if loss_type[type_i] == 'trickLogD':
for _ in range(0,kG):
zmb = floatX(np_rng.uniform(-1., 1., size=(batchSize, nz)))
cost = train_g(zmb)
elif loss_type[type_i] == 'minimax':
for _ in range(0,kG):
zmb = floatX(np_rng.uniform(-1., 1., size=(batchSize, nz)))
cost = train_g_minimax(zmb)
elif loss_type[type_i] == 'ls':
for _ in range(0,kG):
zmb = floatX(np_rng.uniform(-1., 1., size=(batchSize, nz)))
cost = train_g_ls(zmb)
sample_zmb = floatX(np_rng.uniform(-1., 1., size=(ntf, nz)))
gen_imgs = gen_fn(sample_zmb)
frr_score, fd_score = dis_fn(xmb[0:ntf],gen_imgs)
#frr = frr[0]
frr = frr_score - fd_score
if can_i*nloss + type_i < ncandi:
idx = can_i*nloss + type_i
gen_new_params[idx]=lasagne.layers.get_all_param_values(generator)
fake_rate[idx]=frr
g_imgs_old[idx*ntf:(idx+1)*ntf,:]=gen_imgs
fmb[idx*batchSize/ncandi*kD:(idx+1)*batchSize/ncandi*kD,:] = \
gen_imgs[0:batchSize/ncandi*kD,:]
else:
fr_com = fake_rate - frr
if min(fr_com) < 0:
ids_replace = np.where(fr_com==min(fr_com))
idr = ids_replace[0][0]
fake_rate[idr]=frr
gen_new_params[idr] = lasagne.layers.get_all_param_values(generator)
g_imgs_old[idr*ntf:(idr+1)*ntf,:]=gen_imgs
fmb[idr*batchSize/ncandi*kD:(idr+1)*batchSize/ncandi*kD,:] = \
gen_imgs[0:batchSize/ncandi*kD,:]
sample_xmb = toy_dataset(DATASET=DATASET, size=ncandi*ntf)
sample_xmb = sample_xmb[0:ncandi*ntf]
for i in range(0, ncandi):
xfake = g_imgs_old[i*ntf:(i+1)*ntf,:]
xreal = sample_xmb[i*ntf:(i+1)*ntf,:]
tr, fr, trp, frp, fdscore = disft_fn(xreal,xfake)
if i == 0:
fake_rate = np.array([fr])
real_rate = np.array([tr])
fake_rate_p = np.array([frp])
real_rate_p = np.array([trp])
FDL = np.array([fdscore])
else:
fake_rate = np.append(fake_rate,fr)
real_rate = np.append(real_rate,tr)
fake_rate_p = np.append(fake_rate_p,frp)
real_rate_p = np.append(real_rate_p,trp)
FDL = np.append(FDL,fdscore)
print fake_rate, fake_rate_p, FDL
print (n_updates, real_rate.mean(), real_rate_p.mean())
f_log.write(str(fake_rate)+' '+str(fake_rate_p)+'\n'+ str(n_updates) + ' ' + str(real_rate.mean())+ ' ' +str(real_rate_p.mean())+'\n')
f_log.flush()
# train D
for xreal,xfake in iter_data(xmb, shuffle(fmb), size=batchSize):
cost = train_d(xreal, xfake)
if n_updates%show_freq == 0:
s_zmb = floatX(np_rng.uniform(-1., 1., size=(512, nz)))
g_imgs = gen_fn(s_zmb)
xmb = toy_dataset(DATASET=DATASET, size=512)
generate_image(xmb,g_imgs,n_updates/save_freq,desc)
def create_G(noise=None, discriminator=None, lr=0.0002, b1=0.5, DIM=64):
alias_noise = T.matrix('noise') if noise == None else noise
generator = models_uncond.build_generator_toy(alias_noise, nd=DIM)
return GeneratorTrainer(alias_noise, generator, discriminator, lr, b1)
def main():
# Parameters
task = 'toy'
name = '8G_MOEGAN_MMDu2' #'8G_MOEGAN_PFq_NFd_t2'
DIM = 512
begin_save = 0
loss_type = ['trickLogD', 'minimax', 'ls'] #['trickLogD', 'minimax', 'ls']
nloss = 3 #2
DATASET = '8gaussians'
batchSize = 64
ncandi = 8
kD = 1 # # of discrim updates for each gen update
kG = 1 # # of discrim updates for each gen update
ntf = 256
b1 = 0.5 # momentum term of adam
nz = 2 # # of dim for Z
niter = 4 # # of iter at starting learning rate
lr = 0.0001 # initial learning rate for adam G
lrd = 0.0001 # initial learning rate for adam D
N_up = 100000
save_freq = 10000 / 10
show_freq = 10000 / 10
test_deterministic = True
beta = 1.
GP_norm = False # if use gradients penalty on discriminator
LAMBDA = 2. # hyperparameter of GP
NSGA2 = True
# Load the dataset
# MODEL D
print("Building model and compiling functions...")
# Prepare Theano variables for inputs and targets
real_imgs = T.matrix('real_imgs')
fake_imgs = T.matrix('fake_imgs')
# Create neural network model
discriminator = models_uncond.build_discriminator_toy(nd=DIM,
GP_norm=GP_norm)
# Create expression for passing real data through the discriminator
real_out = lasagne.layers.get_output(discriminator, real_imgs)
# Create expression for passing fake data through the discriminator
fake_out = lasagne.layers.get_output(discriminator, fake_imgs)
# Create loss expressions
discriminator_loss = (
lasagne.objectives.binary_crossentropy(real_out, 1) +
lasagne.objectives.binary_crossentropy(fake_out, 0)).mean()
# Gradients penalty norm
if GP_norm is True:
alpha = t_rng.uniform((batchSize, 1), low=0., high=1.)
differences = fake_imgs - real_imgs
interpolates = real_imgs + (alpha * differences)
gradients = theano.grad(lasagne.layers.get_output(
discriminator, interpolates).sum(),
wrt=interpolates)
slopes = T.sqrt(T.sum(T.sqr(gradients), axis=(1)))
gradient_penalty = T.mean((slopes - 1.)**2)
D_loss = discriminator_loss + LAMBDA * gradient_penalty
b1_d = 0.
else:
D_loss = discriminator_loss
b1_d = 0.
# Create update expressions for training
discriminator_params = lasagne.layers.get_all_params(discriminator,
trainable=True)
lrtd = theano.shared(lasagne.utils.floatX(lrd))
updates_d = lasagne.updates.adam(D_loss,
discriminator_params,
learning_rate=lrtd,
beta1=b1_d)
lrt = theano.shared(lasagne.utils.floatX(lr))
# Fd Socre
Fd = theano.gradient.grad(discriminator_loss, discriminator_params)
Fd_score = beta * T.log(sum(T.sum(T.sqr(x)) for x in Fd))
# Compile a function performing a training step on a mini-batch (by giving
# the updates dictionary) and returning the corresponding training loss:
train_d = theano.function([real_imgs, fake_imgs],
discriminator_loss,
updates=updates_d)
# Compile another function generating some data
dis_fn = theano.function([real_imgs, fake_imgs],
[(fake_out).mean(), Fd_score])
disft_fn = theano.function([real_imgs, fake_imgs], [
real_out.mean(),
fake_out.mean(), (real_out > 0.5).mean(),
(fake_out > 0.5).mean(), Fd_score
])
# Finally, launch the training loop.
print("Starting training...")
desc = task + '_' + name
print(desc)
if not os.path.isdir('logs'):
os.mkdir(os.path.join('logs'))
f_log = open('logs/%s.ndjson' % desc, 'wb')
if not os.path.isdir('models'):
os.mkdir(os.path.join('models/'))
if not os.path.isdir('models/' + desc):
os.mkdir(os.path.join('models/', desc))
gen_new_params = []
# We iterate over epochs:
for n_updates in range(N_up):
xmb = toy_dataset(DATASET=DATASET, size=batchSize * kD)
xmb = xmb[0:batchSize * kD]
# initial G cluster
if n_updates == 0:
for can_i in range(0, ncandi):
train_g, gen_fn, generator = create_G(
loss_type=loss_type[can_i % nloss],
discriminator=discriminator,
lr=lr,
b1=b1,
DIM=DIM)
for _ in range(0, kG):
zmb = floatX(np_rng.uniform(-1., 1., size=(batchSize, nz)))
cost = train_g(zmb)
sample_zmb = floatX(np_rng.uniform(-1., 1., size=(ntf, nz)))
gen_imgs = gen_fn(sample_zmb)
gen_new_params.append(
lasagne.layers.get_all_param_values(generator))
if can_i == 0:
g_imgs_old = gen_imgs
fmb = gen_imgs[0:int(batchSize / ncandi * kD), :]
else:
g_imgs_old = np.append(g_imgs_old, gen_imgs, axis=0)
newfmb = gen_imgs[0:int(batchSize / ncandi * kD), :]
fmb = np.append(fmb, newfmb, axis=0)
# print gen_new_params
# MODEL G
noise = T.matrix('noise')
generator = models_uncond.build_generator_toy(noise, nd=DIM)
Tgimgs = lasagne.layers.get_output(generator)
Tfake_out = lasagne.layers.get_output(discriminator, Tgimgs)
g_loss_logD = lasagne.objectives.binary_crossentropy(Tfake_out,
1).mean()
g_loss_minimax = - \
lasagne.objectives.binary_crossentropy(Tfake_out, 0).mean()
g_loss_ls = T.mean(T.sqr((Tfake_out - 1)))
g_params = lasagne.layers.get_all_params(generator, trainable=True)
up_g_logD = lasagne.updates.adam(g_loss_logD,
g_params,
learning_rate=lrt,
beta1=b1)
up_g_minimax = lasagne.updates.adam(g_loss_minimax,
g_params,
learning_rate=lrt,
beta1=b1)
up_g_ls = lasagne.updates.adam(g_loss_ls,
g_params,
learning_rate=lrt,
beta1=b1)
train_g = theano.function([noise], g_loss_logD, updates=up_g_logD)
train_g_minimax = theano.function([noise],
g_loss_minimax,
updates=up_g_minimax)
train_g_ls = theano.function([noise], g_loss_ls, updates=up_g_ls)
gen_fn = theano.function([noise],
lasagne.layers.get_output(
generator, deterministic=True))
else:
class Instance:
def __init__(self, fq, fd, params, img_values, image_copy):
self.fq = fq
self.fd = fd
self.params = params
self.vimg = img_values
self.cimg = image_copy
def f(self):
return self.fq - self.fd
instances = []
fq_list = np.zeros(ncandi)
fd_list = np.zeros(ncandi)
gen_old_params = gen_new_params
for can_i in range(0, ncandi):
for type_i in range(0, nloss):
lasagne.layers.set_all_param_values(
generator, gen_old_params[can_i])
if loss_type[type_i] == 'trickLogD':
for _ in range(0, kG):
zmb = floatX(
np_rng.uniform(-1., 1., size=(batchSize, nz)))
cost = train_g(zmb)
elif loss_type[type_i] == 'minimax':
for _ in range(0, kG):
zmb = floatX(
np_rng.uniform(-1., 1., size=(batchSize, nz)))
cost = train_g_minimax(zmb)
elif loss_type[type_i] == 'ls':
for _ in range(0, kG):
zmb = floatX(
np_rng.uniform(-1., 1., size=(batchSize, nz)))
cost = train_g_ls(zmb)
sample_zmb = floatX(np_rng.uniform(-1., 1.,
size=(ntf, nz)))
gen_imgs = gen_fn(sample_zmb)
frr_score, fd_score = dis_fn(xmb[0:ntf], gen_imgs)
instances.append(
Instance(
frr_score, fd_score,
lasagne.layers.get_all_param_values(generator),
gen_imgs,
gen_imgs[0:int(batchSize / ncandi * kD), :]))
if ncandi < len(instances):
if NSGA2 == True:
cromos = {
idx: [float(inst.fq), -float(inst.fd)]
for idx, inst in enumerate(instances)
}
cromos_idxs = [idx for idx, _ in enumerate(instances)]
finalpop = nsga_2_pass(ncandi, cromos, cromos_idxs)
for idx, p in enumerate(finalpop):
inst = instances[p]
gen_new_params[idx] = inst.params
fq_list[idx] = inst.fq
fd_list[idx] = inst.fd
fake_rate[idx] = inst.f()
g_imgs_old[idx * ntf:(idx + 1) * ntf, :] = inst.vimg
fmb[int(idx * batchSize / ncandi *
kD):math.ceil((idx + 1) * batchSize / ncandi *
kD), :] = inst.cimg
with open('front/%s.tsv' % desc, 'wb') as ffront:
for idx, p in enumerate(finalpop):
inst = instances[p]
ffront.write(
(str(inst.fq) + "\t" + str(inst.fd)).encode())
ffront.write("\n".encode())
else:
for idx, inst in enumerate(instances):
if idx < ncandi:
gen_new_params[idx] = inst.params
fake_rate[idx] = inst.f()
fq_list[idx] = inst.fq
fd_list[idx] = inst.fd
g_imgs_old[idx * ntf:(idx + 1) *
ntf, :] = inst.vimg
fmb[int(idx * batchSize / ncandi *
kD):math.ceil((idx + 1) * batchSize /
ncandi * kD), :] = inst.cimg
else:
fr_com = fake_rate - inst.f()
if min(fr_com) < 0:
idr = np.where(fr_com == min(fr_com))[0][0]
gen_new_params[idr] = inst.params
fake_rate[idr] = inst.f()
g_imgs_old[idr * ntf:(idr + 1) *
ntf, :] = inst.vimg
fmb[int(idr * batchSize / ncandi *
kD):math.ceil((idr + 1) * batchSize /
ncandi *
kD), :] = inst.cimg
sample_xmb = toy_dataset(DATASET=DATASET, size=ncandi * ntf)
sample_xmb = sample_xmb[0:ncandi * ntf]
for i in range(0, ncandi):
xfake = g_imgs_old[i * ntf:(i + 1) * ntf, :]
xreal = sample_xmb[i * ntf:(i + 1) * ntf, :]
tr, fr, trp, frp, fdscore = disft_fn(xreal, xfake)
if i == 0:
fake_rate = np.array([fr])
real_rate = np.array([tr])
fake_rate_p = np.array([frp])
real_rate_p = np.array([trp])
FDL = np.array([fdscore])
else:
fake_rate = np.append(fake_rate, fr)
real_rate = np.append(real_rate, tr)
fake_rate_p = np.append(fake_rate_p, frp)
real_rate_p = np.append(real_rate_p, trp)
FDL = np.append(FDL, fdscore)
print(fake_rate, fake_rate_p, FDL)
print(n_updates, real_rate.mean(), real_rate_p.mean())
f_log.write((str(fake_rate) + ' ' + str(fake_rate_p) + '\n' +
str(n_updates) + ' ' + str(real_rate.mean()) + ' ' +
str(real_rate_p.mean()) + '\n').encode())
f_log.flush()
# train D
for xreal, xfake in iter_data(xmb, shuffle(fmb), size=batchSize):
cost = train_d(xreal, xfake)
if n_updates % show_freq == 0:
s_zmb = floatX(np_rng.uniform(-1., 1., size=(512, nz)))
params_max = gen_new_params[np.argmax(fake_rate)]
lasagne.layers.set_all_param_values(generator, params_max)
g_imgs_max = gen_fn(s_zmb)
if n_updates % show_freq == 0 and n_updates != 0:
#metric
s_zmb = floatX(np_rng.uniform(-1., 1., size=(512, nz)))
xmb = toy_dataset(DATASET=DATASET, size=512)
mmd2_all = []
for i in range(0, ncandi):
lasagne.layers.set_all_param_values(generator,
gen_new_params[i])
g_imgs_min = gen_fn(s_zmb)
mmd2_all.append(compute_metric_mmd2(g_imgs_min, xmb))
mmd2_all = np.array(mmd2_all)
if NSGA2:
with open('front/%s_mmd2u.tsv' % desc, 'wb') as ffront:
for idx in range(0, ncandi):
ffront.write(
(str(fq_list[idx]) + "\t" + str(fd_list[idx]) +
"\t" + str(mmd2_all[idx])).encode())
ffront.write("\n".encode())
#save best
params = gen_new_params[np.argmin(mmd2_all)]
lasagne.layers.set_all_param_values(generator, params)
g_imgs_min = gen_fn(s_zmb)
generate_image(xmb,
g_imgs_min,
n_updates / save_freq,
desc,
postfix="_mmu2d")
np.savez('models/%s/gen_%d.npz' % (desc, n_updates / save_freq),
*lasagne.layers.get_all_param_values(discriminator))
np.savez('models/%s/dis_%d.npz' % (desc, n_updates / save_freq),
*lasagne.layers.get_all_param_values(generator))
def create_geneator(self, noise, dim):
generator = models_uncond.build_generator_toy(noise, nd=dim)
return generator