def __call__(self, h, dist):
"""
Args:
h (numpy.ndarray): axis 0 represents minibatch index,
axis 1 represents atom_index and axis2 represents
feature dimension.
dist (numpy.ndarray): axis 0 represents minibatch index,
axis 1 and 2 represent distance between atoms.
"""
mb, atom, ch = h.shape
if ch != self.hidden_dim:
raise ValueError('h.shape[2] {} and hidden_dim {} must be same!'
.format(ch, self.hidden_dim))
embedlist = self.xp.arange(
self.num_rbf).astype('f') * self.radius_resolution
dist = functions.reshape(dist, (mb, atom, atom, 1))
dist = functions.broadcast_to(dist, (mb, atom, atom, self.num_rbf))
dist = functions.exp(- self.gamma * (dist - embedlist) ** 2)
dist = functions.reshape(dist, (-1, self.num_rbf))
dist = self.dense1(dist)
dist = functions.softplus(dist)
dist = self.dense2(dist)
dist = functions.softplus(dist)
dist = functions.reshape(dist, (mb, atom, atom, self.hidden_dim))
h = functions.reshape(h, (mb, atom, 1, self.hidden_dim))
h = functions.broadcast_to(h, (mb, atom, atom, self.hidden_dim))
h = functions.sum(h * dist, axis=1)
return h
def dis_loss(discriminator, y, t):
y_dis = discriminator(y)
t_dis = discriminator(t)
loss = F.mean(F.softplus(-t_dis)) + F.mean(F.softplus(y_dis))
return loss
def train(self, expert_states, expert_next_states, expert_action_probs, fake_states, fake_next_states,
fake_action_probs, gamma):
def logits(states, next_states, log_action_probs):
# p(expert|state, action) = sigmoid(logits)
rewards = self.reward_net(states)
# print(F.mean(rewards))
state_values = self.value_net(states)
next_state_values = self.value_net(next_states)
return rewards + gamma * next_state_values - state_values - log_action_probs[:, None].array
# This parameter stabilise training
# softplus(logits) == log(sigmoid(logits))
# print('expert: ', end='')
loss = F.mean(F.softplus(-logits(expert_states, expert_next_states, expert_action_probs)))
# print('fake: ', end='')
loss += F.mean(F.softplus(logits(fake_states, fake_next_states, fake_action_probs)))
# add gradient penalty for reward
# xp = chainer.cuda.get_array_module(expert_states)
# e = xp.random.uniform(0., 1., len(expert_states))[:, None].astype(xp.float32)
# x_hat = chainer.Variable((e * expert_states + (1 - e) * fake_states), requires_grad=True)
# grad, = chainer.grad([self.reward_net(x_hat)], [x_hat], enable_double_backprop=True)
# loss_grad = 0.1 * F.mean(F.sqrt(F.batch_l2_norm_squared(grad)))
# loss += loss_grad
self.reward_net.cleargrads()
self.value_net.cleargrads()
loss.backward()
self.reward_optimizer.update()
self.value_optimizer.update()
return loss
def __call__(self, h, dist):
"""
Args:
h (numpy.ndarray): axis 0 represents minibatch index,
axis 1 represents atom_index and axis2 represents
feature dimension.
dist (numpy.ndarray): axis 0 represents minibatch index,
axis 1 and 2 represent distance between atoms.
"""
mb, atom, ch = h.shape
if ch != self.hidden_dim:
raise ValueError(
'h.shape[2] {} and hidden_dim {} must be same!'.format(
ch, self.hidden_dim))
embedlist = self.xp.arange(
self.num_rbf).astype('f') * self.radius_resolution
dist = functions.reshape(dist, (mb, atom, atom, 1))
dist = functions.broadcast_to(dist, (mb, atom, atom, self.num_rbf))
dist = functions.exp(-self.gamma * (dist - embedlist)**2)
dist = functions.reshape(dist, (-1, self.num_rbf))
dist = self.dense1(dist)
dist = functions.softplus(dist)
dist = self.dense2(dist)
dist = functions.softplus(dist)
dist = functions.reshape(dist, (mb, atom, atom, self.hidden_dim))
h = functions.reshape(h, (mb, atom, 1, self.hidden_dim))
h = functions.broadcast_to(h, (mb, atom, atom, self.hidden_dim))
h = functions.sum(h * dist, axis=1)
return h
def loss_dis(self, dis, y_fake, y_real):
batchsize = len(y_fake)
L1 = F.sum(F.softplus(-y_real)) / batchsize
L2 = F.sum(F.softplus(y_fake)) / batchsize
loss = L1 + L2
chainer.report({'loss': loss}, dis)
return loss
def loss_gen_multi(self, gen, y_flowgan_fake, y_texgan_fake, C=0.1):
batchsize = y_flowgan_fake.data.shape[0]
loss_flow = F.sum(F.softplus(-y_flowgan_fake)) / batchsize
loss_tex = F.sum(F.softplus(-y_texgan_fake)) / batchsize
loss = loss_flow + C * loss_tex
chainer.report({'loss': loss}, gen)
return loss
def loss_dis(self, dis, y_fake, y_real):
batchsize = len(y_fake)
L1= F.sum(F.softplus(-y_real)) / batchsize
L2 = F.sum(F.softplus(y_fake)) / batchsize
loss = L1 + L2
train_loss_dis.append(loss)
return loss
def loss_dis(y_in, y_out):
batchsize, _, w, h = y_in.data.shape
L1 = F.sum(F.softplus(-y_in)) / batchsize / w / h
L2 = F.sum(F.softplus(y_out)) / batchsize / w / h
loss = L1 + L2
# print("loss_dis:", loss)
return loss
def loss_discriminator(self, discriminator, real, fake):
batchsize, _, h, w = fake.data.shape
loss1 = F.sum(F.softplus(-real)) / batchsize / h / w
loss2 = F.sum(F.softplus(fake)) / batchsize / h / w
loss = loss1 + loss2
chainer.report({'loss': loss}, discriminator)
return loss
def loss_dis(self, dis, y_in, y_out, a=1, b=1):
L1 = a*F.sum(F.softplus(-y_in)) / len(y_in)
L2 = b*F.sum(F.softplus(y_out)) /len(y_out)
loss = (L1 + L2) / (a+b)
chainer.report({'loss': loss}, dis)
return loss
def compute_discriminator_loss(self, image_real, image_fake, image_labeled,
label):
# predict
prediction_real = self.discriminator(image_real)
prediction_fake = self.discriminator(image_fake)
prediction_labeled = self.discriminator(image_labeled)
# discriminator loss
prediction_real_lse = cf.logsumexp(prediction_real, axis=1)
prediction_fake_lse = cf.logsumexp(prediction_fake, axis=1)
loss_discriminator = (
0.5 * cf.sum(cf.softplus(prediction_real_lse)) /
prediction_real_lse.size +
0.5 * cf.sum(-prediction_real_lse) / prediction_real_lse.size +
0.5 * cf.sum(cf.softplus(prediction_fake_lse)) /
prediction_fake_lse.size)
# classifier loss
loss_classifier = cf.softmax_cross_entropy(prediction_labeled, label)
loss = loss_discriminator + loss_classifier
chainer.reporter.report(
{
'loss_discriminator': loss_discriminator,
'loss_classifier': loss_classifier
}, self)
return loss
def loss_dis(self, dis, y_fake, y_real):
batchsize = len(y_fake)
L1 = F.sum(F.softplus(-y_real)) / batchsize
L2 = F.sum(F.softplus(y_fake)) / batchsize
loss = L1 + L2
chainer.report({'loss': loss}, dis)
return loss
def loss_dis(self, dis, y_in, y_out):
batchsize,_,w,h = y_in.data.shape
L1 = F.sum(F.softplus(-y_in)) / batchsize / w / h
L2 = F.sum(F.softplus(y_out)) / batchsize / w / h
loss = L1 + L2
chainer.report({'loss': loss}, dis)
return loss
def loss_dis(self, dis, completed_y, real_y):
batchsize = len(completed_y)
L1 = F.sum(F.softplus(-real_y)) / batchsize
L2 = F.sum(F.softplus(completed_y)) / batchsize
loss = L1 + L2 # GAN loss
chainer.report({'loss': loss}, dis)
return loss
def dis_loss(opt, real_d, fake_d, observer=None):
#adversarial loss
adv_loss = 0
real_loss = 0
fake_loss = 0
if opt.adv_loss_mode == 'bce':
real_loss = F.mean(F.softplus(-real_d))
fake_loss = F.mean(F.softplus(fake_d))
if opt.adv_loss_mode == 'mse':
xp = cuda.get_array_module(real_d.array)
real_loss = F.mean_squared_error(real_d, xp.ones_like(real_d.array))
fake_loss = F.mean_squared_error(fake_d, xp.zeros_like(fake_d.array))
if opt.adv_loss_mode == 'hinge':
real_loss = F.mean(F.relu(1.0 - real_d))
fake_loss = F.mean(F.relu(1.0 + fake_d))
adv_loss = (real_loss + fake_loss) * 0.5
loss = adv_loss
if observer is not None:
report(
{
'loss': loss,
'adv_loss': adv_loss,
'real_loss': real_loss,
'fake_loss': fake_loss
},
observer=observer)
return loss
def learning_GAN_log(generator,
discliminator,
optgen,
optdis,
data,
result,
T=200):
for time in range(T):
optgen.target.cleargrads()
ytemp = generator(data[1])
with chainer.using_config('train', False):
ytrain_false = discliminator(ytemp)
loss_train_gen = 0.5 * F.mean(F.softplus(-ytrain_false))
loss_train_gen.backward()
optgen.update()
#偽物画像
optdis.target.cleargrads()
ytrain_false = discliminator(ytemp)
ytrain_true = discliminator(data[0])
loss1 = 0.5 * F.mean(F.softplus(ytrain_false))
loss2 = 0.5 * F.mean(F.softplus(-ytrain_true))
loss_train_dis = loss1 + loss2
loss_train_dis.backward()
optdis.update()
result[0].append(cuda.to_cpu(loss_train_gen.data))
result[1].append(cuda.to_cpu(loss1.data))
result[2].append(cuda.to_cpu(loss2.data))
def __call__(self, x):
# <question : is batchsize>1 possible for RNN ? if No, I will implement calculations without batch dimension.>
self.chi = F.concat((x, self.r))
(self.nu, self.xi) = \
F.split_axis(self.l_dl(self.chi), [self.Y], 1)
(self.kr, self.betar, self.kw, self.betaw,
self.e, self.v, self.f, self.ga, self.gw, self.pi
) = F.split_axis(self.xi, np.cumsum(
[self.W*self.R, self.R, self.W, 1, self.W, self.W, self.R, 1, 1]), 1)
self.kr = F.reshape(self.kr, (self.R, self.W)) # R * W
self.betar = 1 + F.softplus(self.betar) # 1 * R
# self.kw: 1 * W
self.betaw = 1 + F.softplus(self.betaw) # 1 * 1
self.e = F.sigmoid(self.e) # 1 * W
# self.v : 1 * W
self.f = F.sigmoid(self.f) # 1 * R
self.ga = F.sigmoid(self.ga) # 1 * 1
self.gw = F.sigmoid(self.gw) # 1 * 1
self.pi = F.softmax(F.reshape(self.pi, (self.R, 3))) # R * 3 (softmax for 3)
# self.wr : N * R
self.psi_mat = 1 - F.matmul(Variable(np.ones((self.N, 1)).astype(np.float32)), self.f) * self.wr # N * R
self.psi = Variable(np.ones((self.N, 1)).astype(np.float32)) # N * 1
for i in range(self.R):
self.psi = self.psi * F.reshape(self.psi_mat[:,i],(self.N,1)) # N * 1
# self.ww, self.u : N * 1
self.u = (self.u + self.ww - (self.u * self.ww)) * self.psi
self.a = u2a(self.u) # N * 1
self.cw = C(self.M, self.kw, self.betaw) # N * 1
self.ww = F.matmul(F.matmul(self.a, self.ga) + F.matmul(self.cw, 1.0 - self.ga), self.gw) # N * 1
self.M = self.M * (np.ones((self.N, self.W)).astype(np.float32) - F.matmul(self.ww, self.e)) + F.matmul(self.ww, self.v) # N * W
self.p = (1.0 - F.matmul(Variable(np.ones((self.N,1)).astype(np.float32)), F.reshape(F.sum(self.ww),(1,1)))) \
* self.p + self.ww # N * 1
self.wwrep = F.matmul(self.ww, Variable(np.ones((1, self.N)).astype(np.float32))) # N * N
self.L = (1.0 - self.wwrep - F.transpose(self.wwrep)) * self.L + F.matmul(self.ww, F.transpose(self.p)) # N * N
self.L = self.L * (np.ones((self.N, self.N)) - np.eye(self.N)) # force L[i,i] == 0
self.fo = F.matmul(self.L, self.wr) # N * R
self.ba = F.matmul(F.transpose(self.L), self.wr) # N * R
self.cr_list = [0] * self.R
for i in range(self.R):
self.cr_list[i] = C(self.M, F.reshape(self.kr[i,:],(1, self.W)),
F.reshape(self.betar[0,i],(1, 1))) # N * 1
self.cr = F.concat(self.cr_list) # N * R
self.bacrfo = F.concat((F.reshape(F.transpose(self.ba),(self.R,self.N,1)),
F.reshape(F.transpose(self.cr),(self.R,self.N,1)),
F.reshape(F.transpose(self.fo) ,(self.R,self.N,1)),),2) # R * N * 3
self.pi = F.reshape(self.pi, (self.R,3,1)) # R * 3 * 1
self.wr = F.transpose(F.reshape(F.batch_matmul(self.bacrfo, self.pi), (self.R, self.N))) # N * R
self.r = F.reshape(F.matmul(F.transpose(self.M), self.wr),(1, self.R * self.W)) # W * R (-> 1 * RW)
self.y = self.l_Wr(self.r) + self.nu # 1 * Y
return self.y
def update_core(self):
gen_optimizer = self.get_optimizer('opt_gen')
dis_optimizer = self.get_optimizer('opt_dis')
xp = self.gen.xp
batch = self.get_iterator('main').next()
batchsize = len(batch)
x = []
for i in range(batchsize):
x.append(np.asarray(batch[i][0]).astype("f"))
x_real = Variable(xp.asarray(x))
y_real = self.dis(x_real)
z = Variable(xp.asarray(self.gen.make_hidden(batchsize)))
x_fake = self.gen(z)
y_fake = self.dis(x_fake)
loss_dis = F.sum(F.softplus(-y_real)) / batchsize
loss_dis += F.sum(F.softplus(y_fake)) / batchsize
loss_gen = F.sum(F.softplus(-y_fake)) / batchsize
self.gen.cleargrads()
loss_gen.backward()
gen_optimizer.update()
x_fake.unchain_backward()
self.dis.cleargrads()
loss_dis.backward()
dis_optimizer.update()
chainer.reporter.report({'loss_gen': loss_gen})
chainer.reporter.report({'loss_dis': loss_dis})
def loss_dis_removal(self, dis_removal, y_in_removal, y_out_removal):
batchsize, _, w, h = y_in_removal.data.shape
L1 = F.sum(F.softplus(-y_in_removal)) / batchsize / w / h
L2 = F.sum(F.softplus(y_out_removal)) / batchsize / w / h
loss = L1 + L2
#chainer.report({'loss': loss}, dis)
return loss
def __call__(self, x, r):
"""Forward propagaion
Args:
x (numpy.ndarray): axis 0 represents minibatch index,
axis 1 represents atom_index and axis 2 represents feature
dimension.
r (numpy.ndarray): axis 0 represents minibatch index,
axis 1 and 2 represent distance between atoms.
"""
# s0 minibatch, s1 atom (from), s2 atom (to)
s0, s1, s2 = r.shape
# a0 minibatch, a1 atom, a2 ch. a2 must be equal to self.hidden_dim
a0, a1, a2 = x.shape
if a2 != self.hidden_dim:
raise ValueError(
"x.shape[2] {} and hidden_dim {} must be same!".format(
a2, self.hidden_dim))
embedlist = self.xp.arange(
self.num_rbf, dtype=self.xp.float32) * self.radius_resolution
r = functions.reshape(r, (s0, s1, s2, 1))
r = functions.broadcast_to(r, (s0, s1, s2, self.num_rbf))
r = functions.exp(-self.gamma * (r - embedlist)**2)
r = functions.reshape(r, (s0 * s1 * s2, self.num_rbf))
r = self.dense1(r)
r = functions.softplus(r)
r = self.dense2(r)
r = functions.softplus(r)
r = functions.reshape(r, (s0, s1, s2, self.hidden_dim))
x = functions.reshape(x, (a0, a1, 1, self.hidden_dim))
x = functions.broadcast_to(x, (a0, a1, s2, self.hidden_dim))
x = functions.sum(x * r, axis=1)
return x
def loss_dis(self, dis, y_fake, y_real):
batchsize = len(y_fake)
loss_real = F.sum(F.softplus(-y_real)) / batchsize
loss_fake = F.sum(F.softplus(y_fake)) / batchsize
loss = loss_fake + loss_real
chainer.report({'loss': loss}, dis)
return loss
def update_core(self):
iterator = self.get_iterator('main') # type: Iterator
g_opt = self.get_optimizer('gen') # type: chainer.Optimizer
d_opt = self.get_optimizer('dis') # type: chainer.Optimizer
batch = iterator.next()
batch_size = len(batch)
x1, x2 = self.converter(batch, self.device)
generated = self.gen(x1)
dis_real = self.dis(x1, x2)
dis_fake = self.dis(x1, generated)
g_loss = F.sum(F.softplus(
-dis_fake)) / dis_fake.size + self.lambda_ * F.mean_absolute_error(
generated, x2)
reporter.report({'loss': g_loss}, self.gen)
self.gen.cleargrads()
g_loss.backward()
g_opt.update()
del g_loss
d_loss = F.sum(F.softplus(-dis_real)) / dis_real.size + F.sum(
F.softplus(dis_fake)) / dis_fake.size
reporter.report({'loss': d_loss}, self.dis)
self.dis.cleargrads()
d_loss.backward()
d_opt.update()
del d_loss, dis_fake, dis_real, x1, x2, batch, generated
def loss_dis(self, dis, y_in, y_out):
batchsize, _, w, h = y_in.data.shape
L1 = F.sum(F.softplus(-y_in)) / batchsize / w / h
L2 = F.sum(F.softplus(y_out)) / batchsize / w / h
loss = L1 + L2
chainer.report({'loss': loss}, dis)
return loss
def loss_discriminator(self, discriminator, x_fake, x_real):
batchsize = len(x_fake)
loss1 = F.sum(F.softplus(-x_real)) / batchsize
loss2 = F.sum(F.softplus(x_fake)) / batchsize
loss = loss1 + loss2
chainer.report({'loss': loss}, discriminator)
return loss
def dis_loss(discriminator, y_fake, x, y_label, x_label):
fake = discriminator(y_fake, F.concat([y_label, x_label]))
real = discriminator(x, F.concat([x_label, y_label]))
#fake = discriminator(y_fake, y_label)
#"real = discriminator(x, x_label)
return F.mean(F.softplus(-real)) + F.mean(F.softplus(fake))
def loss_dis2(self, dis2, y_in, y_out):
batchsize,_,w,h = y_in.data.shape
L1 = F.sum(F.softplus(-y_in)) / batchsize / w / h
L2 = F.sum(F.softplus(y_out)) / batchsize / w / h
loss = L1 + L2
#chainer.report({'loss': loss}, dis2)
#print("dis2", {'loss': loss})
return loss
def dis_loss(self, discriminator, y, t):
y_dis = discriminator(y)
t_dis = discriminator(t)
loss = self.zero_centered_gradient_penalty_fake(y_dis, y)
loss += self.zero_centered_gradient_penalty_real(discriminator, t)
return F.mean(F.softplus(-t_dis)) + F.mean(F.softplus(y_dis)) + loss
def loss_dis(self, dis, y_fake, y_real, y_class, labels):
batchsize = len(y_fake)
L1 = F.sum(F.softplus(-y_real)) / batchsize
L2 = F.sum(F.softplus(y_fake)) / batchsize
L4 = F.softmax_cross_entropy(y_class, labels) / batchsize
loss = L1 + L2 + L4 # + L3 + L4
chainer.report({'loss': loss}, dis)
return loss
def process_batch(self,
x,
y,
one_hot,
last_joint,
return_sample=False,
return_mean=True):
xp = cuda.cupy
x = F.dropout(x, ratio=0.5)
x = F.concat((one_hot, x), axis=-1)
x = self.ln1_(x)
h0 = self.l1_(x)
h = F.dropout(h0, ratio=0.5)
h1 = self.ln2_(F.concat((x, h), axis=1))
h1 = self.l2_(h1)
h = F.dropout(h1, ratio=0.5)
h2 = self.ln3_(F.concat((x, h), axis=1))
h2 = self.l3_(h2)
final_h = F.concat((h0, h1, h2), axis=-1)
mu = self.mu_(final_h)
mu = F.reshape(mu, (-1, self.future_out_dim, self.NUM_MIXTURE))
sigma_orig = self.sigma_(final_h)
mixing_orig = self.mixing_(final_h)
sigma = F.softplus(sigma_orig)
mixing = F.softmax(mixing_orig)
y = F.expand_dims(y, axis=2)
y_broad = F.broadcast_to(y, mu.shape)
normalizer = 2 * np.pi * sigma
exponent = -0.5 * (1. / F.square(sigma)) * F.sum(
(y_broad - mu)**2, axis=1) + F.log(
mixing) - (self.future_out_dim * 0.5) * F.log(normalizer)
cost = -F.logsumexp(exponent, axis=1)
cost = F.mean(cost)
#sampling
if return_sample:
mixing = mixing_orig * (1 + self.sampling_bias)
sigma = F.softplus(sigma_orig - self.sampling_bias)
mixing = F.softmax(mixing)
argmax_mixing = F.argmax(mixing, axis=1)
mixing_one_hot = xp.zeros(mixing.shape, dtype=xp.float32)
mixing_one_hot[xp.arange(mixing.shape[0]), argmax_mixing.data] = 1
component_expanded = F.broadcast_to(
F.expand_dims(mixing_one_hot, axis=1), mu.shape)
component_mean = F.sum(mu * component_expanded, axis=2)
if return_mean:
return cost, component_mean
component_std = F.sum(sigma * component, axis=2, keepdims=True)
component_std = F.broadcast_to(component_std, component_mean.shape)
sample = xp.random.normal(component_mean.data, component_std.data)
return cost, sample
return cost, None
def update_core(self):
# TODO: support n_Classfier
# TIPS: in case of experiments, set n_critic as 5 is best result.
gen_optimizer = self.get_optimizer('gen')
critic_optimizer = self.get_optimizer('critic')
xp = self.generator.xp
for i in range(self.n_critic):
# grab data
batch = self.get_iterator('main').next()
batchsize = len(batch)
batch = self.converter(batch, self.device)
real_data, real_label = batch
real_label = Variable(real_label)
real_data = Variable(real_data) / 255.
# TODO: cWGANってuniformで良いんだっけ...?
z = Variable(
xp.asarray(
self.generator.make_input_z_with_label(
batchsize, real_label.data)))
# Generator
gen_data = self.generator(z)
# Critic(Discrimintor)
critic_real = self.critic(real_data, real_label)
critic_fake = self.critic(gen_data, real_label)
# Loss
## Critic Loss
# print(critic_fake.shape, critic_real.shape, gen_data.shape, real_data.shape)
# critic_loss = F.mean(critic_fake - critic_real)
critic_loss = F.sum(F.softplus(-critic_real)) / batchsize
critic_loss += F.sum(F.softplus(critic_fake)) / batchsize
self.critic.cleargrads()
critic_loss.backward()
critic_optimizer.update()
chainer.report({'critic_loss': critic_loss})
#
# if chainer.is_debug():
# graph = cg.build_computational_graph(critic_loss)
# with open(os.path.join(), 'w') as file:
# file.write(graph.dump())
if i == 0:
# Generator Loss
gen_loss = F.sum(F.softplus(-critic_fake)) / batchsize
self.generator.cleargrads()
gen_loss.backward()
gen_optimizer.update()
chainer.report({'gen_loss': gen_loss})
def loss_dis(self, dis, dis_real, dis_fake):
batchsize, _, w, h = dis_real.data.shape
L1 = (2 + np.random.rand()) * F.sum(
F.softplus(-dis_real)) / batchsize / w / h
L2 = (2 + np.random.rand()) * F.sum(
F.softplus(dis_fake)) / batchsize / w / h
loss = L1 + L2
chainer.report({'loss': loss}, dis)
return loss
def __call__(self, x):
h = F.reshape(x, (x.shape[0], -1))
h = self.l1(h)
h = self.l2(F.softplus(h))
h = self.l3(F.softplus(h))
self.feature = F.softplus(h)
h = self.l4(self.feature)
return h
def loss_dis(self, dis, y_fake, y_real):
batchsize = len(y_fake)
#G(x, E(x))->1, G(D(z), z)->0
#real label 1
L1 = F.sum(F.softplus(-y_real)) / batchsize
#fake label 0
L2 = F.sum(F.softplus(y_fake)) / batchsize
loss = L1 + L2
chainer.report({'loss': loss}, dis)
return loss
def dis_loss(opt, real_d, fake_d, real_g, fake_g, observer=None, tag=str()):
#gradient penalty
real_gp = 0
fake_gp = 0
if opt.zero_gp_mode == 'real' or opt.zero_gp_mode == 'real_fake':
real_gp = opt.gp_coef * compute_grad(real_d, real_g)
if opt.zero_gp_mode == 'fake' or opt.zero_gp_mode == 'real_fake':
fake_gp = opt.gp_coef * compute_grad(fake_d, fake_g)
#adversarial loss
adv_loss = 0
real_loss = 0
fake_loss = 0
if opt.adv_loss_mode == 'wgan':
adv_loss = -F.mean(real_d - fake_d)
gp = real_gp + fake_gp
else:
if opt.adv_loss_mode == 'bce':
real_loss = F.mean(F.softplus(-real_d))
fake_loss = F.mean(F.softplus(fake_d))
if opt.adv_loss_mode == 'mse':
xp = cuda.get_array_module(real_d.array)
real_loss = F.mean_squared_error(real_d,
xp.ones_like(real_d.array))
fake_loss = F.mean_squared_error(fake_d,
xp.zeros_like(fake_d.array))
if opt.adv_loss_mode == 'hinge':
real_loss = F.mean(F.relu(1.0 - real_d))
fake_loss = F.mean(F.relu(1.0 + fake_d))
adv_loss = (real_loss + fake_loss) * 0.5
gp = (real_gp + fake_gp) * 0.5
loss = adv_loss + gp
if observer is not None:
if tag:
tag += '_'
report(
{
tag + 'loss': l2f(loss),
tag + 'adv_loss': l2f(adv_loss),
tag + 'real_loss': l2f(real_loss),
tag + 'fake_loss': l2f(fake_loss),
tag + 'gp': l2f(gp),
tag + 'adv_loss_with_gp': l2f(adv_loss + gp)
},
observer=observer)
return loss
def dis_loss(self, discriminator, y_fake, x, y_label, x_label, residual):
x = chainer.Variable(x.data)
if residual:
fake = discriminator(y_fake + x, y_label)
else:
fake = discriminator(y_fake, y_label)
real = discriminator(x, x_label)
return F.mean(F.softplus(-real)), F.mean(F.softplus(fake))
def __call__(self, x, dist):
v = self.linear[0](x)
v = self.cfconv(v, dist)
v = self.linear[1](v)
v = functions.softplus(v)
v = self.linear[2](v)
return x + v
def bernoulli_nll(x, y):
"""Calculate negative log-likelihood of Bernoulli distribution.
This function calculates negative log-likelihood on a Bernoulli
distribution.
.. math::
-B(x; p) = -\\sum_i {x_i \\log(p_i) + (1 - x_i)\\log(1 - p_i)},
where :math:`p = \\sigma(y)`, and :math:`\\sigma(\\cdot)` is a sigmoid
funciton.
.. note::
As this funtion uses a sigmoid function, you can pass a result of
fully-connected layer (that means :class:`Linear`) to this function
directly.
Args:
x (~chainer.Variable): Input variable.
y (~chainer.Variable): A variable representing the parameter of
Bernoulli distribution.
Returns:
~chainer.Variable: A variable representing negative log-likelihood.
"""
assert isinstance(x, variable.Variable)
assert isinstance(y, variable.Variable)
return F.sum(F.softplus(-y)) + F.sum(y) - F.sum(y * x)
def __call__(self, x):
# x.shape == (batchsize, 3, 128, 64)
batchsize = x.shape[0]
h = F.elu(self.bn1(self.conv1_1(x)))
h = F.elu(self.bn2(self.conv1_2(h)))
h = F.max_pooling_2d(h, 3, 2, cover_all=False)
h = self.conv2_1(h)
h = self.conv2_3(h)
h = self.conv3_1(h)
h = self.conv3_3(h)
h = self.conv4_1(h)
h = self.conv4_3(h)
h = h.reshape(batchsize, -1)
h = F.dropout(h, ratio=0.6)
h = F.elu(self.fc1_bn(self.fc1(h)))
# Features in rows, normalize axis 1.
weights = self.mean_vectors
features = self.ball(h)
features = F.normalize(features, eps=1e-8)
scale = F.softplus(self.scale)
normalized_weight = F.normalize(weights, axis=0, eps=1e-8)
logits = F.tile(scale[None, ], (batchsize, 1)) * \
F.matmul(features, normalized_weight)
return logits
def loss_dec(self, dec, x_out, t_out, y_out, lam1=100, lam2=1):
batchsize, _, w, h = y_out.data.shape
loss_rec = lam1*(F.mean_absolute_error(x_out, t_out))
loss_adv = lam2*F.sum(F.softplus(-y_out)) / batchsize / w / h
loss = loss_rec + loss_adv
chainer.report({'loss': loss}, dec)
return loss
def check_forward(self, x_data):
x = chainer.Variable(x_data)
y = functions.softplus(x, beta=self.beta)
x_value = cuda.to_cpu(x_data)
y_exp = numpy.log(1 + numpy.exp(self.beta * x_value)) / self.beta
self.assertEqual(y.data.dtype, self.dtype)
testing.assert_allclose(
y_exp, y.data, **self.check_forward_options)
def _loss_predictor(self, predictor, output, target, d_fake):
b, _, t = d_fake.data.shape
loss_mse = (F.mean_absolute_error(output, target))
chainer.report({'mse': loss_mse}, predictor)
loss_adv = F.sum(F.softplus(-d_fake)) / (b * t)
chainer.report({'adversarial': loss_adv}, predictor)
loss = self.loss_config.mse * loss_mse + self.loss_config.adversarial * loss_adv
chainer.report({'loss': loss}, predictor)
return loss
def check_backward(self, x_data, y_grad):
x = chainer.Variable(x_data)
y = functions.softplus(x, beta=self.beta)
self.assertEqual(y.data.dtype, numpy.float32)
y.grad = y_grad
y.backward()
func = y.creator
f = lambda: func.forward((x.data,))
gx, = gradient_check.numerical_grad(f, (x.data,), (y.grad,))
gradient_check.assert_allclose(gx, x.grad)
def _loss_discriminator(self, discriminator, d_real, d_fake):
b, _, t = d_real.data.shape
loss_real = F.sum(F.softplus(-d_real)) / (b * t)
chainer.report({'real': loss_real}, discriminator)
loss_fake = F.sum(F.softplus(d_fake)) / (b * t)
chainer.report({'fake': loss_fake}, discriminator)
loss = loss_real + loss_fake
chainer.report({'loss': loss}, discriminator)
tp = (d_real.data > 0.5).sum()
fp = (d_fake.data > 0.5).sum()
fn = (d_real.data <= 0.5).sum()
tn = (d_fake.data <= 0.5).sum()
accuracy = (tp + tn) / (tp + fp + fn + tn)
precision = tp / (tp + fp)
recall = tp / (tp + fn)
chainer.report({'accuracy': accuracy}, self.discriminator)
chainer.report({'precision': precision}, self.discriminator)
chainer.report({'recall': recall}, self.discriminator)
return loss
def bernoulli_nll_keepbatch(self, x, y): nll = F.softplus(y) - x * y return F.sum(nll, axis=1)
def f(x):
y = functions.softplus(x, beta=self.beta)
return y * y
def forward(self, inputs, device):
x, = inputs
return functions.softplus(x, beta=self.beta),
def _encode(self, xs):
exs = self.embed_mat(xs)
h = F.tanh(self.l1(exs))
logits = F.softplus(self.l2(h))
logits = F.log(logits + 1e-10).reshape(-1, self.M, self.K)
return logits, exs
def __call__(self, x): return F.softplus(x, self.use_cudnn)
def loss_gen(self, gen, y_fake):
batchsize = y_fake.data.shape[0]
loss = F.sum(F.softplus(-y_fake)) / batchsize
chainer.report({'loss': loss}, gen)
return loss
def loss_gen(self, gen, y_fake):
batchsize = len(y_fake)
loss = F.sum(F.softplus(-y_fake)) / batchsize
train_loss_gen.append(loss)
return loss
def check_forward(self, x_data):
x = chainer.Variable(x_data)
y = functions.softplus(x, beta=self.beta)
x_value = cuda.to_cpu(x_data)
y_exp = numpy.log(1 + numpy.exp(self.beta * x_value)) / self.beta
gradient_check.assert_allclose(y_exp, y.data)
def __call__(self, x): return functions.softplus(x, self.beta)
def f(x):
return functions.softplus(x, beta=self.beta)
def loss_gen(self, gen, y_fake):
batchsize = len(y_fake)
loss = F.sum(F.softplus(-y_fake)) / batchsize
chainer.report({'loss': loss}, gen)
return loss
def loss(self, source, target, weight):
word = F.dropout(self.embed(target), ratio=self.dropout_ratio)
inner = F.sum(source * word, axis=1)
sp = F.sum(F.softplus(-inner) * weight)
return sp
def __call__(self, h):
h = self.linear1(h)
h = functions.softplus(h)
h = self.linear2(h)
h = functions.sum(h, axis=1)
return h
