def get_loss_func(self,x, C=1.0):
batchsize = len(self.encode(x)[0])
z=list()
mu, ln_var = self.encode(x)
for l in six.moves.range(self.sampling_number):
z.append(F.gaussian(mu, ln_var))
for iii in range(self.sampling_number):
if iii==0:
rec_loss=0
z = F.gaussian(mu, ln_var)
rec_loss += F.sum(F.bernoulli_nll(x, self.decode(z, sigmoid=False), reduce='no'),axis=1)/(batchsize)
loss=rec_loss+F.sum(C * gaussian_kl_divergence(mu, ln_var,reduce='no'),axis=1)/ batchsize
loss=F.reshape(loss,[batchsize,1])
else:
rec_loss=0
z = F.gaussian(mu, ln_var)
rec_loss += F.sum(F.bernoulli_nll(x, self.decode(z, sigmoid=False), reduce='no'),axis=1)/(batchsize)
tmp_loss=rec_loss+F.sum(C * gaussian_kl_divergence(mu, ln_var,reduce='no'),axis=1)/ batchsize
tmp_loss=F.reshape(tmp_loss,[batchsize,1])
loss=F.concat((loss,tmp_loss),axis=1)
importance_weight = F.softmax(loss)
self.total_loss=F.sum(importance_weight*loss)
return self.total_loss
def check_bernoulli_nll(self, x, y):
if self.wrap_x:
x = chainer.Variable(x)
if self.wrap_y:
y = chainer.Variable(y)
actual = cuda.to_cpu(F.bernoulli_nll(x, y, self.reduce).data)
testing.assert_allclose(self.expect, actual)
def lf(x):
mu, ln_var = self.encode(x)
mean_mu, mean_sigma = calculate_means(mu, ln_var)
batchsize = len(mu.data)
std_mu, std_ln_var = generate_std_params(mu)
# reconstruction loss
rec_loss = 0
kl_loss = 0
for l in six.moves.range(k):
z = F.gaussian(mu, ln_var)
rec_loss += F.bernoulli_nll(x, self.decode(z, sigmoid=False)) / (k * batchsize)
kl_loss += -F.gaussian_nll(z, mu, ln_var) / (k * batchsize)
kl_loss += F.gaussian_nll(z, std_mu, std_ln_var) / (k * batchsize)
self.rec_loss = rec_loss
self.kl_loss = kl_loss
self.loss = self.rec_loss + C * self.kl_loss
chainer.report(
{
'rec_loss': rec_loss,
'kl': self.kl_loss,
'loss': self.loss,
'mu': mean_mu,
'sigma': mean_sigma,
},
observer=self
)
return self.loss
def lf(x):
# x = x.reshape(-1, 3*64*64)
mu, ln_var = self.encode(x)
batchsize = len(mu.data)
# reconstruction loss
rec_loss = 0
for _ in range(self.k):
z = F.gaussian(mu, ln_var)
rec_loss += F.bernoulli_nll(x, self.decode(z, sigmoid=False)) \
/ (self.k * batchsize)
self.rec_loss = rec_loss
# latent loss
lat_loss = self.beta * gaussian_kl_divergence(mu,
ln_var) / batchsize
self.lat_loss = lat_loss
self.loss = rec_loss + lat_loss
chainer.report(
{
"rec_loss": rec_loss,
"lat_loss": lat_loss,
"loss": self.loss
},
observer=self)
return self.loss
def __call__(self, xs, zs, es):
batch_size = xs.shape[0]
encoded_zs = self.encoder(xs, es)
ys = self.decoder(encoded_zs)
d_loss = F.bernoulli_nll(xs, ys) / batch_size
t_loss = F.sum(self.discriminator(xs, encoded_zs)) / batch_size
return t_loss + d_loss, encoded_zs
def forward(self, inp, target=None, reward=None):
self.gru.reset_state()
if target is None and reward is None:
sample_size = inp
return self.sample(sample_size)
else:
batch_size, seq_len = inp.shape
inp = inp.T
target = target.T
loss = 0
if not target is None and reward is None:
for i in range(seq_len):
out = self.calc(inp[i])
loss += F.bernoulli_nll(out, target[i])
return loss
elif not target is None and not reward is None:
for i in range(seq_len):
out = self.calc(inp[i])
for j in range(batch_size):
loss += -out[j][target[i][j]] * reward[j]
return loss / batch_size
def bernoulli_nll(x, y):
x = np.array(x, dtype=np.float32)
y = np.array(y, dtype=np.float32)
loss = F.bernoulli_nll(x, y, reduce='no')
plt.plot(loss.data)
plt.show()
print loss
def check_bernoulli_nll(self, x, y):
if self.wrap_x:
x = chainer.Variable(x)
if self.wrap_y:
y = chainer.Variable(y)
actual = cuda.to_cpu(F.bernoulli_nll(x, y, self.reduce).data)
testing.assert_allclose(self.expect, actual)
def update_core(self):
batch = self._iterators['main'].next()
x = Variable(self.converter(batch, self.device))
xp = cuda.get_array_module(x.data)
enc = self.enc
opt_enc = self._optimizers['enc']
dec = self.dec
opt_dec = self._optimizers['dec']
mu, ln_var = enc(x)
batchsize = len(mu.data)
rec_loss = 0
k = 10
for l in range(k):
z = F.gaussian(mu, ln_var)
rec_loss += F.bernoulli_nll(x, dec(
z, sigmoid=False)) / (k * batchsize)
loss = rec_loss + 1.0 * F.gaussian_kl_divergence(mu,
ln_var) / batchsize
enc.cleargrads()
dec.cleargrads()
loss.backward()
opt_enc.update()
opt_dec.update()
chainer.report({'rec_loss': rec_loss})
chainer.report({'loss': loss})
def __call__(self, x, t):
# データ数
num_data = x.shape[0]
# Forwardとlossの計算
mu, var = self.vae.encoder(x)
z = F.gaussian(mu, var)
reconst_loss = 0
for i in range(self.k):
# MSEで誤差計算を行う
if self.loss_function == 'mse':
reconst = self.vae.decoder(z, use_sigmoid=True)
reconst_loss += F.mean_squared_error(x, reconst) / self.k
# その他の場合はベルヌーイ分布により計算
else:
# bernoulli_nllがsigmoidを内包しているので学習時はsigmoid=False
reconst = self.vae.decoder(z, use_sigmoid=False)
reconst_loss += F.bernoulli_nll(x,
reconst) / (self.k * num_data)
kld = gaussian_kl_divergence(mu, var, reduce='mean')
loss = reconst_loss + self.beta * kld
# report
reporter.report({'loss': loss}, self)
reporter.report({'reconst_loss': reconst_loss}, self)
reporter.report({'kld': kld}, self)
return loss
def free_energy(self, x):
#return -(free energy)
enc_mu, enc_log_sigma_2 = self.encode(x)
kl = F.gaussian_kl_divergence(enc_mu, enc_log_sigma_2)
z = F.gaussian(enc_mu, enc_log_sigma_2)
dec_mu = self.decode(z)
nll = F.bernoulli_nll(x, dec_mu)
return nll + kl
def free_energy(self,x):
#return -(free energy)
enc_mu, enc_log_sigma_2 = self.encode(x)
kl = F.gaussian_kl_divergence(enc_mu,enc_log_sigma_2)
z = F.gaussian(enc_mu,enc_log_sigma_2)
dec_mu = self.decode(z)
nll = F.bernoulli_nll(x,dec_mu)
return nll+kl
def lf(x):
z = self.encode(x)
batchsize = len(x.data)
# reconstruction loss
#self.rec_loss = F.mean_squared_error(x, self.decode(z))
self.rec_loss = F.bernoulli_nll(x, self.decode(z, sigmoid=False)) / (batchsize)
# total_loss vanilla AEの場合はreconstruction lossとtotal lossは一緒
self.loss = self.rec_loss
return self.loss
def calc_reconst_loss(self, encoder_input, decoder_output):
if self.is_gauss_dist:
dec_mu, dec_var = decoder_output
m_vae = 0.5* (encoder_input - dec_mu)**2 * F.exp(-dec_var)
a_vae = 0.5* (log2pi+dec_var)
# reconst = F.sum(m_vae + a_vae, axis=1)
reconst = F.sum(m_vae + a_vae)
else:
reconst = F.bernoulli_nll(encoder_input, decoder_output)
return reconst
def train(model, epoch0=0):
optimizer = optimizers.Adam()
optimizer.setup(model)
for epoch in xrange(epoch0, n_epoch):
logger.print_epoch(epoch)
# training
perm = np.random.permutation(N_train)
for i in xrange(0, N_train, batchsize):
x = Variable(xp.asarray(x_train[perm[i:i + batchsize]]))
mu, ln_var = model.encode(x)
rec_loss = 0
for l in xrange(1):
z = F.gaussian(mu, ln_var)
rec_loss += F.bernoulli_nll(x, model.decode(z, sigmoid=False)) / batchsize
reg_loss = gaussian_kl_divergence(mu, ln_var) / batchsize
loss = rec_loss + reg_loss
optimizer.zero_grads()
loss.backward()
loss.unchain_backward()
optimizer.update()
logger.save_loss(reg_loss.data, rec_loss.data, train=True)
# evaluation
for i in xrange(0, N_test, batchsize):
x = Variable(xp.asarray(x_train[i:i + batchsize]), volatile='on')
mu, ln_var = model.encode(x)
rec_loss = 0
for l in xrange(1):
z = F.gaussian(mu, ln_var)
rec_loss += F.bernoulli_nll(x, model.decode(z, sigmoid=False)) / batchsize
reg_loss = gaussian_kl_divergence(mu, ln_var) / batchsize
loss = rec_loss + reg_loss
logger.save_loss(reg_loss.data, rec_loss.data, train=False)
logger.epoch_end()
logger.terminate()
serializer.save(model, optimizer, epoch+1)
# everything works well
return 0
def update_core(self):
vae_optimizer = self.get_optimizer('opt_vae')
xp = self.vae.xp
batch = self.get_iterator('main').next()
batchsize = len(batch)
x = chainer.dataset.concat_examples(batch, device=self.device)
latent_dist = self.vae.encode(x)
# reconstruction loss
rec_loss = 0
for _ in range(self.vae.k):
reconstructions = self.vae(x, sigmoid=False, mode="sample")
rec_loss += F.bernoulli_nll(x, reconstructions) \
/ (self.vae.k * batchsize)
### latent loss
# latent loss for continuous
cont_capacity_loss = 0
if self.vae.is_continuous:
mu, ln_var = latent_dist['cont']
kl_cont_loss = gaussian_kl_divergence(mu, ln_var) / batchsize
# Anealing loss
cont_min, cont_max, cont_num_iters, cont_gamma = \
self.vae.cont_capacity
cont_cap_now = (cont_max - cont_min) * self.iteration / float(cont_num_iters) + cont_min
cont_cap_now = min(cont_cap_now, cont_max)
cont_capacity_loss = cont_gamma * F.absolute(cont_cap_now - kl_cont_loss)
# latent loss for discrete
disc_capacity_loss = 0
if self.vae.is_discrete:
kl_disc_loss = kl_multiple_discrete_loss(latent_dist['disc'])
# Anealing loss
disc_min, disc_max, disc_num_iters, disc_gamma = \
self.vae.disc_capacity
disc_cap_now = (disc_max - disc_min) * self.iteration / float(disc_num_iters) + disc_min
disc_cap_now = min(disc_cap_now, disc_max)
# Require float conversion here to not end up with numpy float
disc_theoretical_max = 0
for disc_dim in self.vae.latent_spec["disc"]:
disc_theoretical_max += xp.log(disc_dim)
disc_cap_now = min(disc_cap_now, disc_theoretical_max.astype("float32"))
disc_capacity_loss = disc_gamma * F.absolute(disc_cap_now - kl_disc_loss)
joint_vae_loss = rec_loss + cont_capacity_loss + disc_capacity_loss
self.vae.cleargrads()
joint_vae_loss.backward()
vae_optimizer.update()
chainer.reporter.report({"rec_loss": rec_loss, "cont_loss": cont_capacity_loss,
"disc_loss": disc_capacity_loss, "vae_loss": joint_vae_loss, })
return
def cost(self, x_var, C=1.0, k=1): mu, ln_var = self.encode(x_var) batchsize = len(mu.data) rec_loss = 0 for l in six.moves.range(k): z = F.gaussian(mu, ln_var) rec_loss += F.bernoulli_nll(x_var, self.decode(z, sigmoid=False)) \ / (k * batchsize) self.rec_loss = rec_loss self.loss = self.rec_loss + C * gaussian_kl_divergence(mu, ln_var) / batchsize return self.loss
def __call__(self, x):
x = Variable(x)
start = time.time()
zm, zv = self.encoder((x,))
z = F.gaussian(zm, zv)
y = self.decoder((z,))[0]
kl_loss = F.gaussian_kl_divergence(zm, zv)
nll_loss = F.bernoulli_nll(x, y)
loss = kl_loss + nll_loss
return loss
def lf(x, x_next):
z = self.encode(x)
batchsize = len(z.data)
# reconstruction loss
loss = 0
for l in six.moves.range(k):
loss += F.bernoulli_nll(x_next, self.decode(z, batchsize, sigmoid=False)) \
/ (k * batchsize)
self.loss = loss
chainer.report({'loss': self.loss}, observer=self)
return self.loss
def get_loss(self, x, y, train=True):
mu, ln_var = self.encode(x, y)
batchsize = len(mu.data)
# reconstruction loss
rec_loss = 0
z = F.gaussian(mu, ln_var)
rec_loss += F.bernoulli_nll(y, self.decode(z, x)) / (batchsize)
self.rec_loss = rec_loss
self.loss = self.rec_loss + F.gaussian_kl_divergence(
mu, ln_var) / batchsize
return self.loss
def reconstruction_loss(self, x, k):
batchsize = len(x)
mu, ln_var = self.encode(x)
samples = []
loss = 0.
for _ in range(k):
z = F.gaussian(mu, ln_var)
samples.append(z)
loss += F.bernoulli_nll(x, self.decode(z)) / (k * batchsize)
dist = Distribution(mu, ln_var, samples)
return loss, dist
def __call__(self, text, x, t, textlens, xlens):
batchsize = text.shape[0]
vk = self.text_enc(text)
v = vk[:, :self.d, :]
k = vk[:, self.d:, :]
q = self.audio_enc(x)
a = F.matmul(F.transpose(k, (0, 2, 1)), q)
a = F.softmax(a / self.xp.sqrt(self.d))
r = F.matmul(v, a)
rd = F.concat((r, q))
y = self.audio_dec(rd)
loss_bin = 0
for i in range(batchsize):
loss_bin += F.mean(
F.bernoulli_nll(t[i, :, :xlens[i]], y[i, :, :xlens[i]], 'no'))
loss_bin /= batchsize
y = F.sigmoid(y)
loss_l1 = 0
for i in range(batchsize):
loss_l1 += F.mean_absolute_error(t[i, :, :xlens[i]],
y[i, :, :xlens[i]])
loss_l1 /= batchsize
loss_att = 0
for i in range(batchsize):
N = textlens[i]
T = xlens[i]
def w_fun(n, t):
return 1 - np.exp(-((n / (N - 1) - t / (T - 1))**2) /
(2 * self.g**2))
w = np.fromfunction(w_fun, (a.shape[1], T), dtype='f')
w = self.xp.array(w)
loss_att += F.mean(w * a[i, :, :T])
loss_att /= batchsize
loss = loss_bin + loss_l1 + loss_att
chainer.reporter.report({
'loss_bin': loss_bin,
'loss_l1': loss_l1,
'loss_att': loss_att,
})
return loss, y, a
def update_core(self):
vae_optimizer = self.get_optimizer('opt_vae')
dis_optimizer = self.get_optimizer('opt_dis')
xp = self.vae.xp
batch = self.get_iterator('main').next()
batchsize = len(batch)
x = chainer.dataset.concat_examples(batch, device=self.device)
mu, ln_var = self.vae.encode(x)
z_sampled = F.gaussian(mu, ln_var)
z_shuffled = shuffle_codes(z_sampled)
logits_z, probs_z = self.dis(z_sampled)
_, probs_z_shuffle = self.dis(z_shuffled)
reconstructions = self.vae.decode(z_sampled, sigmoid=False)
# reconstruction loss
rec_loss = 0
for _ in range(self.vae.k):
rec_loss += F.bernoulli_nll(x, reconstructions) \
/ (self.vae.k * batchsize)
# latent loss
lat_loss = self.vae.beta * gaussian_kl_divergence(mu,
ln_var) / batchsize
tc_loss = F.mean(logits_z[:, 0] - logits_z[:, 1])
factor_vae_loss = rec_loss + lat_loss + self.vae.gamma * tc_loss
dis_loss = -(0.5 * F.mean(F.log(probs_z[:, 0])) \
+ 0.5 * F.mean(F.log(probs_z_shuffle[:, 1])))
self.vae.cleargrads()
self.dis.cleargrads()
factor_vae_loss.backward()
vae_optimizer.update()
# avoid backword duplicate
z_sampled.unchain_backward()
self.dis.cleargrads()
self.vae.cleargrads()
dis_loss.backward()
dis_optimizer.update()
chainer.reporter.report({
"rec_loss": rec_loss,
"lat_loss": lat_loss,
"tc_loss": tc_loss,
"vae_loss": factor_vae_loss,
"dis_loss": dis_loss
})
return
def lf(x):
mu, ln_var = self.encode(x)
batchsize = len(mu.data)
# reconstruction loss
rec_loss = 0
for l in six.moves.range(k):
z = F.gaussian(mu, ln_var)
rec_loss += F.bernoulli_nll(x, self.decode(z, sigmoid=False)) \
/ (k * batchsize)
self.rec_loss = rec_loss
self.loss = self.rec_loss + \
C * gaussian_kl_divergence(mu, ln_var) / batchsize
return self.loss
def lf(x):
mu, ln_var = self.encode(x)
batchsize = len(mu.data)
# reconstruction loss
rec_loss = 0
for l in six.moves.range(k):
z = F.gaussian(mu, ln_var)
rec_loss += F.bernoulli_nll(x, self.decode(z, sigmoid=False)) \
/ (k * batchsize)
self.rec_loss = rec_loss
self.loss = self.rec_loss + \
C * gaussian_kl_divergence(mu, ln_var) / batchsize
return self.loss
def lf(self, x, mu, ln_var, split=False):
batchsize = len(mu.data)
# reconstruction loss
rec_loss = 0
for l in range(k):
z = F.gaussian(mu, ln_var)
rec_loss += F.bernoulli_nll(x, self.decoder_model(z, sigmoid=False)) / (k * batchsize)
rec_loss = rec_loss
kl_loss = C * F.gaussian_kl_divergence(mu, ln_var) / batchsize
loss = rec_loss + kl_loss
if split:
return rec_loss, kl_loss
else:
return loss
def lf(x):
mu, ln_var = self.encode(x)
batchsize = len(mu)
# 復元誤差の計算
rec_loss = 0
for l in six.moves.range(k):
z = F.gaussian(mu, ln_var)
rec_loss += F.bernoulli_nll(x, self.decode(z, sigmoid=False)) / (k * batchsize)
self.rec_loss = rec_loss
self.loss = self.rec_loss + \
beta * gaussian_kl_divergence(mu, ln_var) / batchsize
chainer.report(
{'rec_loss': rec_loss, 'loss': self.loss}, observer=self)
return self.loss
def __call__(self, x, sigmoid=True):
"""AutoEncoder"""
mu, ln_var = self.encode(x)
batchsize = len(mu.data)
# reconstruction loss
rec_loss = 0
for l in six.moves.range(self.k):
z = F.gaussian(mu, ln_var)
rec_loss += F.bernoulli_nll(x, self.decode(z, sigmoid=False)) \
/ (self.k * batchsize)
loss = rec_loss + \
self.C * gaussian_kl_divergence(mu, ln_var) / batchsize
chainer.report({'loss': loss}, self)
return loss
def __call__(self, x, C=1.0, k=1):
mu, ln_var = self.encode(x)
mb_size = mu.data.shape[0]
# reconstruction loss
rec_loss = 0
for l in range(k):
z = F.gaussian(mu, ln_var)
rec_loss += F.bernoulli_nll(x, self.decode(z, sigmoid=False))
rec_loss /= (k * mb_size)
kld_loss = gaussian_kl_divergence(mu, ln_var) / mb_size
loss = rec_loss + C * kld_loss
return loss, float(rec_loss.data), float(kld_loss.data)
def lf(x):
mu, ln_var = self.encode(x)
batchsize = len(mu)
# reconstruction loss
rec_loss = 0
for l in six.moves.range(k):
z = F.gaussian(mu, ln_var)
rec_loss += F.bernoulli_nll(x, self.decode(z, sigmoid=False)) \
/ (k * batchsize)
self.rec_loss = rec_loss
self.loss = self.rec_loss + \
beta * gaussian_kl_divergence(mu, ln_var) / batchsize
chainer.report(
{'rec_loss': rec_loss, 'loss': self.loss}, observer=self)
return self.loss
def lf(x):
in_img = x[0]
in_labels = x[1:-1]
mask = x[-1]
# escape dividing by 0 when there are no labelled data points in the batch
non_masked = sum(mask) + 1
mask_flipped = 1 - mask
rec_loss = 0
label_loss = 0
label_acc = 0
mu, ln_var = self.encode(in_img)
batchsize = len(mu.data)
for l in six.moves.range(k):
z = F.gaussian(mu, ln_var)
out_img = self.decode(z, sigmoid=False)
rec_loss += F.bernoulli_nll(in_img, out_img) / (k * batchsize)
out_labels = self.predict_label(mu, ln_var, softmax=False)
for i in range(self.n_latent):
n = self.groups_len[i] - 1
# certain labels should not contribute to the calculation of the label loss values
fixed_labels = (cupy.tile(cupy.array([1] + [-100] * n),
(batchsize, 1)) *
mask_flipped[:, cupy.newaxis])
out_labels[i] = out_labels[
i] * mask[:, cupy.newaxis] + fixed_labels
label_acc += F.accuracy(out_labels[i], in_labels[i])
label_loss += F.softmax_cross_entropy(
out_labels[i], in_labels[i]) / (k * non_masked)
self.rec_loss = self.alpha * rec_loss
self.label_loss = self.gamma * label_loss
self.label_acc = label_acc
kl = gaussian_kl_divergence(mu, ln_var) / (batchsize)
self.kl = self.beta * kl
self.loss = self.rec_loss + self.label_loss + self.kl
return self.loss, self.rec_loss, self.label_loss, self.label_acc, self.kl
def lf(x):
# getting the mu and ln_var of the prior of z with the encoder
mu, ln_var = self.encode(x)
batchsize = len(mu.data)
# creating the latent variable z by sampling from the encoder output
z = F.gaussian(mu, ln_var)
# computing the reconstruction loss
self.rec_loss = F.bernoulli_nll(x, self.decode(
z, sigmoid=False)) / batchsize
#self.rec_loss = F.sigmoid_cross_entropy(x, self.decode(z, sigmoid=False))
# computing the KL divergence
self.KL_loss = gaussian_kl_divergence(mu, ln_var) / batchsize
# computing the total loss
self.loss = self.rec_loss + self.KL_loss
# returning the losses separately
return [self.rec_loss, self.loss]
def __call__(self, x, t):
y = self.ssrn(x)
loss_l1 = F.mean_absolute_error(t, y)
loss_bin = F.mean(F.bernoulli_nll(t, y, 'no'))
loss = loss_l1 + loss_bin
chainer.reporter.report(
{
'loss_l1': loss_l1,
'loss_bin': loss_bin,
}
)
return loss, y
def lf(x):
mu, ln_var = self.encode(x)
batch_size = len(mu.data)
# reconstruction loss
rec_loss = 0
for l in range(k):
z = f.gaussian(mu, ln_var)
z.name = "z"
rec_loss += f.bernoulli_nll(x, self.decode(z, sigmoid=True)) \
/ (k * batch_size)
self.rec_loss = rec_loss
self.rec_loss.name = "reconstruction error"
self.latent_loss = C * gaussian_kl_divergence(mu, ln_var) / batch_size
self.latent_loss.name = "latent loss"
self.loss = self.rec_loss + self.latent_loss
self.loss.name = "loss"
return self.loss
def loss_func(self, x, t):
y = self.predict(x)
# loss = F.sigmoid_cross_entropy(y, t)
loss = F.bernoulli_nll(t.astype("f"), y) / len(y)
# labelが付いている(t_が1)場合: -log(y_) y_は0となるかも?
# 付いていない(t_が0)場合: -log(1-y_) ここでt_,y_ はx, yの要素
# 以上の総和をバッチサイズで割る
chainer.reporter.report({'loss': loss}, self)
accuracy = self.accuracy(y.data, t)
chainer.reporter.report({'accuracy': accuracy[0]},
self) # dataひとつひとつのlabelが完全一致している確率
chainer.reporter.report({'frequent_error': accuracy[1]},
self) # batchの中で最も多く間違って判断したlabel
chainer.reporter.report({'acc_66': accuracy[2]}, self) # 66番ラベルの正解率
return loss
def lf(x):
self.other_loss = 0.
ze = self.encode(x)
decoded_x = self.decode(ze, sigmoid=False)
batchsize = x.shape[0]
# self.rec_loss = F.mean_squared_error(x, decoded_x)
self.rec_loss = F.bernoulli_nll(x, decoded_x) / batchsize
self.loss = self.rec_loss + self.other_loss
chainer.report(
{
'rec_loss': self.rec_loss,
'other_loss': self.other_loss,
'loss': self.loss
},
observer=self)
del self.rec_loss
del self.other_loss
return self.loss
sum_loss_reconstruction = 0
start_time = time.time()
perm = np.random.permutation(N)
for i in range(0, N, batchsize):
x = chainer.Variable(xp.asarray(x_train[perm[i:i + batchsize]]))
y = chainer.Variable(xp.asarray(y_train[perm[i:i + batchsize]].astype(np.float32)))
y_label = np.argmax(y_train[perm[i:i + batchsize]], axis=1)
# Reconstruction phase
z_fake_batch = model.encode(x)
_x = model.decode(z_fake_batch, sigmoid=False)
loss_reconstruction = F.bernoulli_nll(x, _x) / batchsize
# Adversarial phase
z_real_batch = utils.sample_z_from_n_2d_gaussian_mixture(batchsize, n_z, y_label, 10, use_gpu)
z_real_batch_with_label = F.concat((z_real_batch, y))
p_real_batch = dis(z_real_batch_with_label)
z_fake_batch_with_label = F.concat((z_fake_batch, y))
p_fake_batch = dis(z_fake_batch_with_label)
loss_dis_real = F.softmax_cross_entropy(p_real_batch, chainer.Variable(xp.zeros(batchsize, dtype=np.int32)))
loss_dis_fake = F.softmax_cross_entropy(p_fake_batch, chainer.Variable(xp.ones(batchsize, dtype=np.int32)))
loss_dis = loss_dis_fake + loss_dis_real
loss_gen = F.softmax_cross_entropy(p_fake_batch, chainer.Variable(xp.zeros(batchsize, dtype=np.int32)))
loss_aae = loss_reconstruction + C * loss_gen
def check_bernoulli_nll(self, x_data, y_data):
x = chainer.Variable(x_data)
y = chainer.Variable(y_data)
actual = cuda.to_cpu(F.bernoulli_nll(x, y, self.reduce).data)
testing.assert_allclose(self.expect, actual)
def check_invalid_option(self, xp):
x = chainer.Variable(xp.asarray(self.x))
y = chainer.Variable(xp.asarray(self.y))
with self.assertRaises(ValueError):
F.bernoulli_nll(x, y, 'invalid_option')
