def predict(self, x_data, y_data=None):
x = Variable(x_data)
y = self.forward_raw(x, train=False)
if y_data is None:
return F.sigmoid(y).data.reshape(x_data.shape[0])
else:
return F.sigmoid_cross_entropy(y, Variable(y_data)).data
def forward_eye_states(self, x_batch_curr, y_batch_curr, volatile):
current_sample = Variable(x_batch_curr, volatile=volatile)
y_batch_curr = np.asarray(y_batch_curr).reshape(32, -1)
current_output = Variable(y_batch_curr, volatile=volatile)
h1_current = F.sigmoid(self.model_to_use.x_h1(current_sample))
h2_current = F.sigmoid(self.model_to_use.h1_h2(h1_current))
h3_current = F.sigmoid(self.model_to_use.h2_h3(h2_current))
h4_current = F.sigmoid(self.model_to_use.h3_h4(h3_current))
h4 = h4_current
y = self.model_to_use.h4_y(h4)
y.data = y.data.reshape(32, -1)
loss = F.sigmoid_cross_entropy(y, current_output)
current_output.data = np.squeeze(current_output.data)
accuracy = F.accuracy(y, current_output)
return accuracy, loss, y
def check_forward(self, x_data, t_data, use_cudnn='always'):
x_val = chainer.Variable(x_data)
t_val = chainer.Variable(t_data)
with chainer.using_config('use_cudnn', use_cudnn):
loss = functions.sigmoid_cross_entropy(x_val, t_val,
self.normalize)
self.assertEqual(loss.data.shape, ())
self.assertEqual(loss.data.dtype, numpy.float32)
loss_value = float(cuda.to_cpu(loss.data))
# Compute expected value
loss_expect = 0
non_ignore_count = 0
for i in six.moves.range(self.x.shape[0]):
for j in six.moves.range(self.x.shape[1]):
xd, td = self.x[i, j], self.t[i, j]
if td == -1:
continue
loss_expect -= xd * (td - (xd >= 0)) \
- math.log(1 + math.exp(-numpy.abs(xd)))
non_ignore_count += 1
if non_ignore_count == 0:
loss_expect = 0
elif self.normalize:
loss_expect /= non_ignore_count
else:
loss_expect /= self.t.shape[0]
self.assertAlmostEqual(loss_expect, loss_value, places=5)
def patch_loss_gen(self, dis):
fake = self.random(1, n=self.config.batchsize[self.config.gan_type])
dis_fake = dis(fake)
if self.config.loss_type == "wgan-gp":
loss = F.mean(dis_fake)
elif self.config.loss_type == "nsgan":
loss = F.sigmoid_cross_entropy(
dis_fake, self.xp.ones(dis_fake.shape, dtype="int32"))
else:
raise NotImplementedError()
return loss
def update_discriminator(gen, dis, dis_optimizer, x, z):
xp = cuda.get_array_module(x)
# generate
x_fake = gen(z)
#discriminate
y_real = dis(x)
y_fake = dis(x_fake)
# target
t_real = xp.ones_like(y_real.data, dtype=np.int32)
t_fake = xp.zeros_like(y_fake.data, dtype=np.int32)
loss_dis = F.sigmoid_cross_entropy(
y_real, t_real) + F.sigmoid_cross_entropy(y_fake, t_fake)
dis.cleargrads()
loss_dis.backward()
dis_optimizer.update()
return loss_dis
def __call__(self, f_atoms, f_bonds, super_node_x, atom_label,
mask_reagents, mask_reactants_reagents, batch_size):
h = self.ggnngwm(f_atoms, f_bonds, super_node_x)
h = self.mlp(h)[:, :, 0]
loss = functions.sigmoid_cross_entropy(h, atom_label)
h = functions.sigmoid(h).array
acc = atom_selection_acc(h, atom_label, mask_reagents, self.topK,
batch_size)
atom_selected = atom_selection(h, mask_reactants_reagents, self.topK)
return loss, acc, atom_selected
def evaluate(self):
train_X, train_y = self.collect_prediction_for_train_data()
model = LinearSVC()
model.fit(X=train_X, y=train_y)
iterator = self._iterators['dev']
target = self._targets['main']
# this is necessary for more-than-once-evaluation
it = copy.copy(iterator)
label_scores = []
svm_label_scores = []
summary = reporter.DictSummary()
for n, batch in enumerate(it):
observation = {}
with reporter.report_scope(observation):
padded_batch = self.converter(batch, device=self.device)
x1s = padded_batch['x1s']
x2s = padded_batch['x2s']
wordcnt = padded_batch['wordcnt']
wgt_wordcnt = padded_batch['wgt_wordcnt']
x1s_len = padded_batch['x1s_len']
x2s_len = padded_batch['x2s_len']
y = padded_batch['y']
y_score, sim_scores = target(x1s, x2s, wordcnt, wgt_wordcnt,
x1s_len, x2s_len)
# compute loss
loss = F.sigmoid_cross_entropy(x=y_score, t=y).data
reporter.report({'loss': loss}, target)
# We evaluate WikiQA by MAP and MRR
# for direct evaluation
label_score = np.c_[y, y_score.data]
label_scores.append(label_score)
# for SVM/LR
x = np.concatenate([x.data for x in sim_scores] +
[wordcnt, wgt_wordcnt, x1s_len, x2s_len],
axis=1)
y_score = model.decision_function(x)
svm_label_score = np.c_[y, y_score]
svm_label_scores.append(svm_label_score)
summary.add(observation)
stats = compute_map_mrr(label_scores)
svm_stats = compute_map_mrr(svm_label_scores)
summary_dict = summary.compute_mean()
summary_dict["validation/main/svm_map"] = svm_stats.map
summary_dict["validation/main/svm_mrr"] = svm_stats.mrr
summary_dict["validation/main/map"] = stats.map
summary_dict["validation/main/mrr"] = stats.mrr
return summary_dict
def check_backward(self, x_data, t_data, use_cudnn=True):
x = chainer.Variable(x_data)
t = chainer.Variable(t_data)
loss = functions.sigmoid_cross_entropy(x, t, use_cudnn)
loss.backward()
self.assertEqual(None, t.grad)
func = loss.creator
f = lambda: func.forward((x.data, t.data))
gx, = gradient_check.numerical_grad(f, (x.data,), (1,), eps=0.01)
gradient_check.assert_allclose(gx, x.grad)
def forward(self, inp, target=None):
out = self.calc(inp)
if not target is None:
loss = F.sigmoid_cross_entropy(out, target)
acc = F.binary_accuracy(out, target)
return loss, acc
else:
out = F.sigmoid(out)
return out.reshape(-1)
def calculate_metrics(nets, batch):
gen = nets['gen']
dis = nets['dis']
x, t = batch
xp = cuda.get_array_module(x)
batch_size = len(x)
z = xp.asarray(
np.random.random((batch_size, gen.latent_size)).astype(np.float32))
x_fake = gen(z)
y_real = dis(x)
y_fake = dis(x_fake)
real_label = xp.ones((batch_size, 1), dtype=np.int32)
fake_label = xp.zeros((batch_size, 1), dtype=np.int32)
loss_dis = F.sigmoid_cross_entropy(y_real, real_label)
loss_dis += F.sigmoid_cross_entropy(y_fake, fake_label)
loss_gen = F.sigmoid_cross_entropy(y_fake, real_label)
return {'gen': loss_gen, 'dis': loss_dis}
def mse(self, x, y, undo_norm, test_data=False): acc = 0.0 y = Variable(np.array(y, dtype=np.int32)) pred = self.prediction(x) '''accuracy''' if test_data: for i in range(len(pred)): if int(y[i].data[0]) == int(pred[i]._data[0][0]): acc = acc + 1.0 '''mse → sigomoid cross entropy''' return F.sigmoid_cross_entropy(pred, y), acc / float(len(pred))
def _elementwise_softmax_cross_entropy(x, t, two_class):
assert x.shape[:-1] == t.shape
shape = t.shape
t = F.flatten(t)
if two_class:
x = F.flatten(x)
return F.reshape(
F.sigmoid_cross_entropy(x, t, reduce='no'), shape)
else:
x = F.reshape(x, (-1, x.shape[-1]))
return F.reshape(
F.softmax_cross_entropy(x, t, reduce='no'), shape)
def check_backward(self, x_data, t_data, use_cudnn=True):
x = chainer.Variable(x_data)
t = chainer.Variable(t_data)
loss = functions.sigmoid_cross_entropy(x, t, use_cudnn)
loss.backward()
self.assertEqual(None, t.grad)
func = loss.creator
f = lambda: func.forward((x.data, t.data))
gx, = gradient_check.numerical_grad(f, (x.data,), (1,), eps=0.01)
gradient_check.assert_allclose(gx, x.grad)
def __init__(self, model, loss_fn):
super().__init__()
with self.init_scope():
self.model = model
if loss_fn == 'focal':
self.loss_fn = lambda x, t: F.sum(focal_loss(F.sigmoid(x), t))
elif loss_fn == 'sigmoid':
self.loss_fn = lambda x, t: F.sum(
F.sigmoid_cross_entropy(x, t, reduce='no'))
else:
raise ValueError('unknown loss function. {}'.format(loss_fn))
def forward(self, data, is_training=True):
"""
data: [([label], [features]), ([label], [features]), ...)]
"""
x_raw = [x[1] for x in data]
batch_size = len(x_raw)
# embed sparse featuer vector to dense vector
x_sparse = XP.iarray(x_raw)
x_sparse = F.reshape(x_sparse, [batch_size * self.nzdim])
embeddings = self.embed(x_sparse)
# FM Component
# 1st order
first_order = F.reshape(
F.sum(F.reshape(self.L1(x_sparse), (batch_size, self.nzdim)), 1),
(batch_size, 1))
# 2nd order
embeddings = F.reshape(embeddings,
(batch_size, self.nzdim * self.embed_size))
second_order = XP.fzeros((batch_size, 1))
for i in range(self.nzdim - 1):
for j in range(1, self.nzdim - i):
former = embeddings[:, i * self.embed_size:(i + 1) *
self.embed_size]
later = embeddings[:, (i + j) * self.embed_size:(i + j + 1) *
self.embed_size]
second_order += F.reshape(
F.batch_matmul(former, later, transa=True),
(batch_size, 1))
y_fm = first_order + second_order
# Deep Component
embeddings = F.reshape(embeddings,
(batch_size, self.nzdim * self.embed_size))
h = F.dropout(F.relu(self.L2(embeddings)),
ratio=0.9,
train=is_training)
h = F.dropout(F.relu(self.L3(h)), ratio=0.9, train=is_training)
h = F.dropout(F.relu(self.L4(h)), ratio=0.9, train=is_training)
y_deep = self.L5(h)
y = y_fm + y_deep
if is_training:
t_raw = [t[0] for t in data]
t = XP.iarray(t_raw)
return F.sigmoid_cross_entropy(y, t)
else:
return F.sigmoid(y)
def __call__(self, x, x1, x2, x3, x4, x5, x6, t):
h1 = self.FE(reshapei(x))
h1 = F.sigmoid(self.o1(h1))
loss1 = F.sigmoid_cross_entropy(h1, t[:, :self.n1_classes])
h2 = self.FE(reshapei(x1))
h2 = F.sigmoid(self.o2(h2))
loss2 = F.sigmoid_cross_entropy(h2, t[:, :self.n2_classes])
h3 = self.FE(reshapei(x2))
h3 = F.sigmoid(self.o3(h3))
loss3 = F.sigmoid_cross_entropy(h3, t[:, :self.n3_classes])
h4 = self.FE(reshapei(x3))
h4 = F.sigmoid(self.o4(h4))
loss4 = F.sigmoid_cross_entropy(h4, t[:, :self.n4_classes])
h5 = self.FE(reshapei(x4))
h5 = F.sigmoid(self.o5(h5))
loss5 = F.sigmoid_cross_entropy(h5, t[:, :self.n5_classes])
h6 = self.FE(reshapei(x5))
h6 = F.sigmoid(self.o(h6))
loss6 = F.sigmoid_cross_entropy(h6, t)
return h1, h2, h3, h4, h5, h6, loss1, loss2, loss3, loss4, loss5, loss6
def train(max_epoch, train_size, valid_size):
model = RNN()
# train に1000サンプル、 testに1000サンプル使用
x_train, x_test, y_train, y_test = dataset(train_size + valid_size, train_size)
optimizer = optimizers.RMSprop(lr=0.03)
optimizer.setup(model)
early_stopping = 20
min_valid_loss = 1e8
min_epoch = 0
train_loss, valid_loss = [], []
for epoch in range(1, max_epoch):
_y = model(x_test)
y = _y.data
y = np.array([1 - y, y], dtype='f').T[0]
accuracy = F.accuracy(y, y_test.data.flatten()).data
_train_loss = F.sigmoid_cross_entropy(model(x_train), y_train).data
_valid_loss = F.sigmoid_cross_entropy(_y, y_test).data
train_loss.append(_train_loss)
valid_loss.append(_valid_loss)
# valid_lossが20回連続で更新されなかった時点で学習を終了
if min_valid_loss >= _valid_loss:
min_valid_loss = _valid_loss
min_epoch = epoch
elif epoch - min_epoch >= early_stopping:
break
optimizer.update(forward, x_train, y_train, model)
print('epoch: {} acc: {} loss: {} valid_loss: {}'.format(epoch, accuracy, _train_loss, _valid_loss))
loss_plot(train_loss, valid_loss)
serializers.save_npz('model.npz', model)
def compute_exp_reg_loss(self, pred):
"""Compute expalanation loss.
Args:
pred: Shape is (Batch, 2, H, W)
"""
p_shape = pred.shape
label = self.xp.ones((p_shape[0] * p_shape[2] * p_shape[3], ),
dtype='i')
l = F.sigmoid_cross_entropy(F.reshape(pred, (-1, )),
label,
reduce='no')
return F.mean(l)
def evaluate_roc(self, trainer):
iterator = self._iterators['main']
eval_func = self.eval_func or self._targets['main']
if self.eval_hook:
self.eval_hook(self)
if hasattr(iterator, 'reset'):
iterator.reset()
it = iterator
else:
it = copy.copy(iterator)
y_total = np.array([]).reshape([0, len(self.label_name)])
t_total = np.array([]).reshape([0, len(self.label_name)])
for batch in it:
in_arrays = self.converter(batch, self.device)
with chainer.no_backprop_mode(), chainer.using_config('train', False):
y = eval_func(*in_arrays[:-1])
t = in_arrays[-1]
# y = F.sigmoid(y)
y_data = cuda.to_cpu(y.data)
t_data = cuda.to_cpu(t)
y_total = np.vstack([y_total, y_data])
t_total = np.vstack([t_total, t_data])
updater = trainer.updater
epoch = updater.iteration
out_dir = trainer.out
observation = {}
for label_index, label in enumerate(self.label_name):
y = y_total[:, label_index]
t = t_total[:, label_index]
index = numpy.where(t != -1)[0]
y = y[index]
t = t[index]
out_name = os.path.join(out_dir, str(epoch) + 'iteration_' + label + '_roc.pdf')
gather_data = self.comm.gather(np.vstack([t, y]))
if self.rank == 0:
gather_data = reduce(lambda x, y: np.hstack([x, y]), gather_data)
gather_t = np.array(gather_data[0], dtype=np.int)
gather_y = np.array(gather_data[1], dtype=np.float32)
plot_roc(y_true=gather_t, y_score=F.sigmoid(gather_y).data, out_name=out_name)
roc_auc = metrics.roc_auc_score(gather_t, F.sigmoid(gather_y).data)
with reporter.report_scope(observation):
reporter.report({'roc_auc_' + label: roc_auc}, self._targets['main'])
reporter.report({'loss': F.sigmoid_cross_entropy(gather_y, gather_t).data},
self._targets['main'])
reporter.report({'accuracy': F.binary_accuracy(gather_y, gather_t).data}, self._targets['main'])
return observation
def update_core(self):
image, labels, metadata = self.converter(
self.get_iterator('main').next())
# img = np.transpose(image[0], (1, 2, 0))
# print(img)
# cv2.imshow('img', np.uint8(img))
# cv2.imshow('lbl', np.uint8(labels[0]))
# cv2.waitKey(0)
image = Variable(image)
assert image.shape[0] == 1, "Batchsize of only 1 is allowed for now"
if self.device >= 0:
image.to_gpu(self.device)
xp = get_array_module(image.data)
to_substract = np.array((-1, 0))
noise_classes = np.unique(labels[0]).astype(np.int32)
# print(noise_classes)
target = xp.asarray([[0] * (self.no_of_classes)])
gt_labels = np.setdiff1d(
noise_classes,
to_substract) - 1 # np.unique(labels[0]).astype(np.int32)[2:] - 1
target[0][gt_labels] = 1
gcam, cl_scores, class_id, fconf, floc = self._optimizers[
'main'].target.stream_cl(image, gt_labels)
mask = self._optimizers['main'].target.get_mask(gcam)
masked_image = self._optimizers['main'].target.mask_image(image, mask)
masked_output = self._optimizers['main'].target.stream_am(masked_image)
masked_output = F.sigmoid(masked_output)
cl_loss = F.sigmoid_cross_entropy(cl_scores, target, normalize=True)
am_loss = masked_output[0][class_id][0]
finger_loss = self.finger_loss(fconf, floc, metadata, xp)
labels = Variable(labels)
if self.device >= 0:
labels.to_gpu(self.device)
segment_loss = self._optimizers['main'].target(image, labels)
total_loss = self.lambd1 * cl_loss + self.lambd2 * am_loss + self.lambd3 * segment_loss + 1 * finger_loss
report({'AM_Loss': am_loss}, self.get_optimizer('main').target)
report({'CL_Loss': cl_loss}, self.get_optimizer('main').target)
report({'SG_Loss': segment_loss}, self.get_optimizer('main').target)
report({'TotalLoss': total_loss}, self.get_optimizer('main').target)
self._optimizers['main'].target.cleargrads()
# if not np.isnan(float(total_loss.data)):
total_loss.backward()
self._optimizers['main'].update()
def evaluate_roc(self, trainer):
iterator = self._iterators['main']
eval_func = self.eval_func or self._targets['main']
if self.eval_hook:
self.eval_hook(self)
if hasattr(iterator, 'reset'):
iterator.reset()
it = iterator
else:
it = copy.copy(iterator)
y_total = []
t_total = []
length = it.dataset.__len__()
batchsize = it.batch_size
length = length // batchsize
from tqdm import tqdm
pbar = tqdm(total=length)
for batch in it:
in_arrays = self.converter(batch, self.device)
with chainer.no_backprop_mode(), chainer.using_config('train', False):
y = eval_func(*in_arrays[:-1])
t = in_arrays[-1]
y_data = cuda.to_cpu(y.data)
t_data = cuda.to_cpu(t)
y_total.extend(y_data)
t_total.extend(t_data)
pbar.update(1)
y_total = numpy.concatenate(y_total).ravel()
t_total = numpy.concatenate(t_total).ravel()
index = numpy.where(t_total != -1)[0]
y_total = y_total[index]
t_total = t_total[index]
d = {'label': t_total, 'score': y_total}
from pathlib import Path
out_path = Path('./valid_result') / str(self.epoch)
out_path.mkdir(exist_ok=True)
np.savez(out_path / ('validation_' + str(self.rank)), **d)
observation = {}
with reporter.report_scope(observation):
roc_auc = metrics.roc_auc_score(t_total, F.sigmoid(y_total).data)
with reporter.report_scope(observation):
reporter.report({'roc_auc_': roc_auc}, self._targets['main'])
reporter.report({'loss': F.sigmoid_cross_entropy(y_total, t_total).data},
self._targets['main'])
reporter.report({'accuracy': F.binary_accuracy(y_total, t_total).data}, self._targets['main'])
return observation
def __call__(self, x, t):
h = F.relu(self.conv1(x))
h = F.relu(self.conv2(h))
h = F.relu(self.conv3(h))
h = F.dropout(F.relu(self.fc4(h)), train=self.train)
h = self.fc5(h)
self.pred = F.reshape(h, (x.data.shape[0], 16, 16))
if t is not None:
self.loss = F.sigmoid_cross_entropy(self.pred, t, normalize=False)
return self.loss
else:
self.pred = F.sigmoid(self.pred)
return self.pred
def __call__(self, x, t):
h = F.relu(self.conv1(x))
h = F.relu(self.conv2(h))
h = F.relu(self.conv3(h))
h = F.dropout(F.relu(self.fc4(h)), train=self.train)
h = self.fc5(h)
self.pred = F.reshape(h, (x.data.shape[0], 16, 16))
if t is not None:
self.loss = F.sigmoid_cross_entropy(self.pred, t, normalize=False)
return self.loss
else:
self.pred = F.sigmoid(self.pred)
return self.pred
def evaluate(self):
train_X, train_y = self.collect_prediction_for_train_data()
model = LinearSVC()
# model = LinearSVC(loss="hinge",tol=0.001,max_iter=6000)
# model = SVC(C=1.0, kernel="linear")
model.fit(X=train_X, y=train_y)
iterator = self._iterators['dev']
target = self._targets['main']
# this is necessary for more-than-once-evaluation
it = copy.copy(iterator)
label_scores = []
svm_label_scores = []
summary = reporter.DictSummary()
for n, batch in enumerate(it):
observation = {}
with reporter.report_scope(observation):
x1s, x2s, wordcnt, wgt_wordcnt, x1s_len, x2s_len, y = self.converter(
batch, device=-1)
y_score, sim_scores = target(x1s, x2s, wordcnt, wgt_wordcnt,
x1s_len, x2s_len)
# batch分のy_score(feature vec を1次元に写像したもの)とsim_scoreが帰ってくる
# 正解ラベルyを使ってloss計算
loss = F.sigmoid_cross_entropy(x=y_score, t=y).data
reporter.report({'loss': loss}, target)
# We evaluate WikiQA by MAP and MRR
# for direct evaluation
label_score = np.c_[tc(y), tc(y_score.data)]
label_scores.append(label_score)
# for SVM/LR
x = np.concatenate([tc(x.data) for x in sim_scores] +
[wordcnt, wgt_wordcnt, x1s_len, x2s_len],
axis=1)
y_score = model.decision_function(x)
svm_label_score = np.c_[y, y_score]
svm_label_scores.append(svm_label_score)
summary.add(observation)
stats = compute_map_mrr(label_scores)
svm_stats = compute_map_mrr(svm_label_scores)
summary_dict = summary.compute_mean()
summary_dict["validation/svm_map"] = svm_stats.map
summary_dict["validation/svm_mrr"] = svm_stats.mrr
summary_dict["validation/map"] = stats.map
summary_dict["validation/mrr"] = stats.mrr
return summary_dict
def update_core(self):
gen_optimizer = self.get_optimizer('gen')
dis_optimizer = self.get_optimizer('dis')
xp = self.gen.xp
for i in range(self.n_dis):
batch = self.get_iterator('main').next()
batchsize = len(batch)
if i == 0:
z = self.gen.make_hidden(batchsize)
x_fake = self.gen(z)
y_fake = self.dis(x_fake)
loss_gen = F.sigmoid_cross_entropy(
y_fake, Variable(xp.ones_like(y_fake.data, dtype=xp.int8)))
self.gen.cleargrads()
loss_gen.backward()
gen_optimizer.update()
chainer.reporter.report({'gen/loss': loss_gen})
x_fake.unchain_backward()
x_real = Variable(self.converter(batch, self.device))
y_real = self.dis(x_real)
z = self.gen.make_hidden(batchsize)
x_fake = self.gen(z)
y_fake = self.dis(x_fake.data)
loss_dis = F.sigmoid_cross_entropy(
y_real, Variable(xp.ones_like(y_real.data, dtype=xp.int8)))
loss_dis += F.sigmoid_cross_entropy(
y_fake, Variable(xp.zeros_like(y_fake.data, dtype=xp.int8)))
self.dis.cleargrads()
loss_dis.backward()
dis_optimizer.update()
chainer.reporter.report({'dis/loss': loss_dis})
def noised_sigmoid_cross_entropy(y,
t,
mc_iteration,
normalize=True,
reduce='mean'):
""" Sigmoid Cross-entropy for aleatoric uncertainty estimates.
Args:
y (list of ~chainer.Variable): logits and sigma
t (~numpy.ndarray or ~cupy.ndarray): ground-truth
mc_iteration (int): number of iteration of MCMC.
normalize (bool, optional): Defaults to True.
reduce (str, optional): Defaults to 'mean'.
Returns:
[~chainer.Variable]: Loss value.
"""
assert isinstance(y, (list, tuple))
logits, log_std = y
assert logits.shape[0] == log_std.shape[0]
assert log_std.shape[1] in (logits.shape[1], 1)
assert logits.shape[2:] == log_std.shape[2:]
assert logits.shape == t.shape
xp = backend.get_array_module(t)
# std = F.sqrt(F.exp(log_var))
std = F.exp(log_std)
loss = 0.
for _ in range(mc_iteration):
noise = std * xp.random.normal(0., 1., std.shape)
loss += sigmoid_cross_entropy(logits + noise,
t,
normalize=False,
reduce='no')
if not reduce == 'mean':
return loss
if normalize:
count = loss.size * mc_iteration
else:
count = max(1, len(loss)) * mc_iteration
return F.sum(loss) / count
def patch_loss_dis(self, real, dis, n_dis=5):
for _ in range(n_dis):
fake = self.random(
n=self.config.batchsize[self.config.gan_type]).array
if not isinstance(real, Variable):
real = Variable(real)
dis_real = dis(real)
dis_fake = dis(fake)
if self.config.loss_type == "wgan-gp":
loss = F.mean(dis_real) - F.mean(dis_fake)
loss += F.mean(
F.batch_l2_norm_squared(
chainer.grad([loss], [real],
enable_double_backprop=True)[0])) * 1000
elif self.config.loss_type == "nsgan":
loss = F.sigmoid_cross_entropy(dis_real, self.xp.ones(dis_real.shape, dtype="int32")) + \
F.sigmoid_cross_entropy(dis_fake, self.xp.zeros(dis_fake.shape, dtype="int32"))
loss += F.mean(
F.batch_l2_norm_squared(
chainer.grad([loss], [real],
enable_double_backprop=True)[0])) * 1000
else:
assert "loss type: {self.config.loss_type} is not supported"
return loss
def check_forward(self, x_data, t_data, use_cudnn=True):
x_val = chainer.Variable(x_data)
t_val = chainer.Variable(t_data)
loss = functions.sigmoid_cross_entropy(x_val, t_val, use_cudnn)
loss_value = float(cuda.to_cpu(loss.data))
# Compute expected value
loss_expect = 0
for i in six.moves.range(self.x.shape[0]):
for j in six.moves.range(self.x.shape[1]):
xd, td = self.x[i, j], self.t[i, j]
loss_expect -= xd * (td - (xd >= 0)) \
- math.log(1 + math.exp(-numpy.abs(xd)))
loss_expect /= self.t.shape[0]
self.assertAlmostEqual(loss_expect, loss_value, places=5)
def __call__(self, roi_feature, labels):
# labels shape = B, T, F(9 or 8), 12
# roi_feature shape = B, T, F, D, where F is box number in one frame image
with chainer.cuda.get_device_from_array(roi_feature.data) as device:
node_out = self.forward(roi_feature, labels) # node_out B,T,F,D
node_out = F.reshape(node_out, (-1, self.out_size))
node_labels = self.xp.reshape(labels, (-1, self.out_size))
pick_index, accuracy_pick_index = self.get_loss_index(node_out, node_labels)
loss = F.sigmoid_cross_entropy(node_out[list(pick_index[0]), list(pick_index[1])],
node_labels[list(pick_index[0]), list(pick_index[1])])
accuracy = F.binary_accuracy(node_out[list(accuracy_pick_index[0]), list(accuracy_pick_index[1])],
node_labels[[list(accuracy_pick_index[0]), list(accuracy_pick_index[1])]])
return loss, accuracy
def forward_one_step(x, t, state, train=True):
# if args.gpu >= 0:
# data = cuda.to_gpu(data)
# targets = cuda.to_gpu(targets)
x = chainer.Variable(x, volatile=not train)
t = chainer.Variable(t, volatile=not train)
h_in = model.x_to_h(x) + model.h_to_h(state['h'])
c, h = F.lstm(state['c'], h_in)
y = model.h_to_y(h)
state = {'c': c, 'h': h}
sigmoid_y = 1 / (1 + np.exp(-y.data))
mean_squared_error = ((t.data - sigmoid_y)**2).sum() / t.data.size
return state, F.sigmoid_cross_entropy(y, t), mean_squared_error, h.data[0]
def __call__(self, xs, bboxes, labels): # all shape is (B, T, F, D)
node_out = self.forward(xs) # node_out B,T,F,D
node_out = F.reshape(node_out, (-1, self.out_size))
node_labels = self.xp.reshape(labels, (-1, self.out_size))
pick_index, accuracy_pick_index = self.get_loss_index(node_out, node_labels)
loss = F.sigmoid_cross_entropy(node_out,
node_labels)
accuracy = F.binary_accuracy(node_out[list(accuracy_pick_index[0]), list(accuracy_pick_index[1])],
node_labels[[list(accuracy_pick_index[0]), list(accuracy_pick_index[1])]])
report_dict = {'loss': loss, "accuracy":accuracy}
chainer.reporter.report(report_dict,
self)
return loss
def encode(self, input_batch, teacher_wp, train):
"""
Input batch of sequence and update self.c (context vector) and self.h (hidden vector)
:param teacher_wp:
:param input_batch: batch of input text embed id ex.) [[ 1, 0 ,14 ,5 ], [ ...] , ...]
:param train : True or False
"""
for batch_word in input_batch:
batch_word = chainer.Variable(xp.array(batch_word, dtype=xp.int32))
self.c_batch, self.h_batch = self.enc(batch_word, self.c_batch, self.h_batch, train=train)
self.h_enc.append(self.h_batch)
predict_mat, self.h_batch = self.wpe(self.h_enc)
if train:
return F.sigmoid_cross_entropy(predict_mat, teacher_wp)
def forward_one_step(x, t, state, train=True):
# if args.gpu >= 0:
# data = cuda.to_gpu(data)
# targets = cuda.to_gpu(targets)
x = chainer.Variable(x, volatile=not train)
t = chainer.Variable(t, volatile=not train)
h_in = model.x_to_h(x) + model.h_to_h(state['h'])
c, h = F.lstm(state['c'], h_in)
y = model.h_to_y(h)
state = {'c': c, 'h': h}
sigmoid_y = 1 / (1 + np.exp(-y.data))
mean_squared_error = ((t.data - sigmoid_y) ** 2).sum() / t.data.size
return state, F.sigmoid_cross_entropy(y, t), mean_squared_error
def check_forward(self, x_data, t_data, use_cudnn=True):
x_val = chainer.Variable(x_data)
t_val = chainer.Variable(t_data)
loss = functions.sigmoid_cross_entropy(x_val, t_val, use_cudnn)
loss_value = float(cuda.to_cpu(loss.data))
# Compute expected value
loss_expect = 0
for i in six.moves.range(self.x.shape[0]):
for j in six.moves.range(self.x.shape[1]):
xd, td = self.x[i, j], self.t[i, j]
loss_expect -= xd * (td - (xd >= 0)) \
- math.log(1 + math.exp(-numpy.abs(xd)))
loss_expect /= self.t.shape[0]
self.assertAlmostEqual(loss_expect, loss_value, places=5)
def update_core(self):
gen_optimizer = self.get_optimizer('gen')
dis_optimizer = self.get_optimizer('dis')
batch = self.get_iterator('main').next()
x_real = Variable(self.converter(batch, self.device))
batchsize = len(batch)
z = self.gen.make_hidden(batchsize)
x_fake = self.gen(z)
real_feature, y_real = self.dis(x_real)
fake_feature, y_fake = self.dis(x_fake)
loss_gen = F.sigmoid_cross_entropy(
y_fake,
Variable(self.xp.ones_like(y_fake.data, dtype=self.xp.int32)))
loss_feature = F.mean_squared_error(real_feature, fake_feature)
self.gen.cleargrads()
loss_gen.backward()
loss_feature.backward()
gen_optimizer.update()
chainer.reporter.report({'gen/loss': loss_gen})
chainer.reporter.report({'gen/loss_feature': loss_feature})
x_fake.unchain_backward()
L1 = F.sigmoid_cross_entropy(
y_real,
Variable(self.xp.ones_like(y_real.data, dtype=self.xp.int32)))
L2 = F.sigmoid_cross_entropy(
y_fake,
Variable(self.xp.zeros_like(y_fake.data, dtype=self.xp.int32)))
loss_dis = L1 + L2
self.dis.cleargrads()
loss_dis.backward()
dis_optimizer.update()
chainer.reporter.report({'dis/loss': loss_dis})
def __call__(self, xs, crf_pact_structures):
loss = 0.0
xp = chainer.cuda.cupy.get_array_module(xs.data)
accuracy = 0.0
for idx, X in enumerate(xs):
crf_pact_structure = crf_pact_structures[idx]
gt_label = self.get_gt_label_one_graph(xp,
crf_pact_structure,
is_bin=True) # N x Y
A = crf_pact_structure.A
for i in range(self.layers_num):
X = getattr(self, "attn_{}".format(i))(X, A)
loss += F.sigmoid_cross_entropy(X, gt_label)
accuracy = F.binary_accuracy(X, gt_label)
chainer.reporter.report({"loss": loss, "accuracy": accuracy})
return loss
def check_backward(self, x_data, t_data, use_cudnn=True):
x = chainer.Variable(x_data)
t = chainer.Variable(t_data)
loss = functions.sigmoid_cross_entropy(x, t, use_cudnn)
loss.backward()
self.assertEqual(None, t.grad)
# Skip too large case. That requires a long time.
if self.shape[0] == 65536:
return
func = loss.creator
f = lambda: func.forward((x.data, t.data))
gx, = gradient_check.numerical_grad(f, (x.data,), (1,), eps=0.01)
gradient_check.assert_allclose(gx, x.grad)
def forward(x_data, y_data, train=True):
x = Variable(x_data, volatile=not train)
t = Variable(y_data, volatile=not train)
feature_input = F.relu(model.bn_feature(model.img_feature2vec(x)))
l1 = F.relu(model.bn1(model.h1(feature_input)))
y = model.out(l1)
loss = F.sigmoid_cross_entropy(y, t)
predicted = cuda.to_cpu(F.sigmoid(y).data)
predicted[predicted > 0.5] = 1
predicted[predicted <= 0.5] = 0
label = cuda.to_cpu(y_data)
index = np.multiply(predicted == label, label == 1)
accuracy = float(np.sum(index)) / float(np.sum(label == 1))
return loss, accuracy
def __call__(self, x, t):
h1 = F.max_pooling_2d(F.relu(self.conv1(x)), 2)
h2 = F.max_pooling_2d(F.relu(self.conv2(h1)), 2)
h3 = F.relu(self.l1(h2))
y = self.l2(h3)
self.loss = F.sigmoid_cross_entropy(y, t)
accuracy = []
y_ = F.array.split_axis.split_axis(y, self.output_dim, 1)
t_ = F.array.split_axis.split_axis(t, self.output_dim, 1)
for y_ele, t_ele in zip(y_, t_):
accuracy.append(F.accuracy(y_ele, chainer.Variable(t_ele.data.squeeze())))
self.accuracy = accuracy
return self.loss
def forward_one_step(x, t, state, train=True):
# if args.gpu >= 0:
# data = cuda.to_gpu(data)
# targets = cuda.to_gpu(targets)
x = chainer.Variable(x, volatile=not train)
t = chainer.Variable(t, volatile=not train)
h_in = model.x_to_h(x) + model.h_to_h(state['h'])
c, h = F.lstm(state['c'], h_in)
y = model.h_to_y(h)
state = {'c': c, 'h': h}
sigmoid_y = 1 / (1 + np.exp(-y.data))
bin_y = np.round((np.sign(sigmoid_y - 0.5) + 1) / 2)
square_sum_error = ((t.data - sigmoid_y) ** 2).sum()
bin_y_error = ((t.data - bin_y) ** 2).sum()
return state, F.sigmoid_cross_entropy(y, t), square_sum_error, bin_y_error, h.data[0]
def loss(model, xs, ts, uss=None):
model.reset_state()
tags = model([Variable(
np.array([x], dtype=np.int32)
) for x in xs])
zss = []
d = Variable(np.array(0, dtype=np.float32))
for t, (y, zs) in zip(ts, tags):
d += cf.sigmoid_cross_entropy(
y, Variable(np.array([[t]], dtype=np.int32))
)
if t:
zss.append(zs)
if uss:
assert len(uss) == len(zss)
for us, zs in zip(uss, zss):
for u, z in zip(us, zs):
d += cf.softmax_cross_entropy(
z, Variable(np.array([u], dtype=np.int32))
)
return d
def forward(self, x_batch_curr, y_batch_curr, volatile=False):
current_sample = Variable(x_batch_curr, volatile=volatile)
y_batch_curr = np.asarray(y_batch_curr).reshape(1, -1)
current_output = Variable(y_batch_curr, volatile=volatile)
h1_current = F.sigmoid(self.model_to_use.x_h1(current_sample))
# h1_previous = F.sigmoid(self.model_to_use.x_h1(previous_sample))
# h1_next = F.sigmoid(self.model_to_use.x_h1(next_sample))
# h1_diff_previous = h1_current - h1_previous
# h1_diff_next = h1_next - h1_current
h2_current = F.sigmoid(self.model_to_use.h1_h2(h1_current))
# h2_diff_n = F.sigmoid(self.model_to_use.h1_h2(h1_diff_next))
# h2_diff_p = F.sigmoid(self.model_to_use.h1_h2(h1_diff_previous))
# h2_diff_next = h2_diff_n - h2_current
# h2_diff_previous = h2_current - h2_diff_p
h3_current = F.sigmoid(self.model_to_use.h2_h3(h2_current))
# h3_diff_p = F.sigmoid(self.model_to_use.h2_h3(h2_diff_previous))
# h3_diff_n = F.sigmoid(self.model_to_use.h2_h3(h2_diff_next))
# h3_diff_next = h3_diff_n - h3_current
# h3_diff_previous = h3_current - h3_diff_p
h4_current = F.sigmoid(self.model_to_use.h3_h4(h3_current))
# h4_diff_previous = F.sigmoid(self.model_to_use.h3_h4(h3_diff_previous))
# h4_diff_next = F.sigmoid(self.model_to_use.h3_h4(h3_diff_next))
# h4_diff = h4_diff_next + h4_diff_previous
# h4 = h4_current * h4_diff
h4 = h4_current
y = self.model_to_use.h4_y(h4)
loss = F.sigmoid_cross_entropy(y, current_output)
current_output.data = np.squeeze(current_output.data)
y.data = y.data.reshape(-1, 1)
accuracy = F.accuracy(y, current_output)
return accuracy, loss, y
def __call__(self, x, labels):
x = BatchTransform(self.model.mean)(x)
x = self.xp.array(x)
scores = self.model(x)
B, n_class = scores.shape[:2]
one_hot_labels = self.xp.zeros((B, n_class), dtype=np.int32)
for i, label in enumerate(labels):
one_hot_labels[i, label] = 1
# sigmoid_cross_entropy normalizes the loss
# by the size of batch and the number of classes.
# It works better to remove the normalization factor
# of the number of classes.
loss = self.loss_scale * F.sigmoid_cross_entropy(
scores, one_hot_labels)
result = calc_accuracy(scores, labels)
reporter.report({'loss': loss}, self)
reporter.report({'accuracy': result['accuracy']}, self)
reporter.report({'n_pred': result['n_pred']}, self)
reporter.report({'n_pos': result['n_pos']}, self)
return loss
def check_forward_no_reduction(self, x_data, t_data):
x_val = chainer.Variable(x_data)
t_val = chainer.Variable(t_data)
loss = functions.sigmoid_cross_entropy(
x_val, t_val, self.normalize, reduce='no')
self.assertEqual(loss.data.shape, self.x.shape)
self.assertEqual(loss.data.dtype, numpy.float32)
loss_value = cuda.to_cpu(loss.data)
# Compute expected value
if not getattr(self, 'ignore_all', False):
for i in six.moves.range(self.x.shape[0]):
for j in six.moves.range(self.x.shape[1]):
xd, td = self.x[i, j], self.t[i, j]
if td == -1:
loss_expect = 0
else:
loss_expect = -(
xd * (td - (xd >= 0)) -
math.log(1 + math.exp(-numpy.abs(xd))))
self.assertAlmostEqual(
loss_expect, loss_value[i, j], places=5)
d_opt = optimizers.Adam(alpha=0.0002, beta1=0.5)
g_opt.setup(gen)
d_opt.setup(dis)
g_opt.add_hook(chainer.optimizer.WeightDecay(0.00001))
d_opt.add_hook(chainer.optimizer.WeightDecay(0.00001))
example_z = gen.make_z(nbatch)
for epoch in range(50000):
print "epoch:", epoch
xmb = image_samples(nbatch)
x = gen(chainer.Variable(gen.make_z(nbatch)))
y1 = dis(x)
l_gen = F.sigmoid_cross_entropy(y1, chainer.Variable(np.ones((nbatch, 1), dtype=np.int32)))
l1_dis = F.sigmoid_cross_entropy(y1, chainer.Variable(np.zeros((nbatch, 1), dtype=np.int32)))
x2 = chainer.Variable(xmb)
y2 = dis(x2)
l2_dis = F.sigmoid_cross_entropy(y2, chainer.Variable(np.ones((nbatch, 1), dtype=np.int32)))
l_dis = l1_dis + l2_dis
print "loss gen:", l_gen.data
print "loss dis1:", l1_dis.data
print "loss dis2:", l2_dis.data
gen.zerograds()
dis.zerograds()
margin = 0.25
def __call__(self, xb, yb, tb):
xc = Variable(np.array(xb, dtype=np.int32))
yc = Variable(np.array(yb, dtype=np.int32))
tc = Variable(np.array(tb, dtype=np.int32))
fv = self.fwd(xc, yc)
return F.sigmoid_cross_entropy(fv, tc)
def forward(self, x_data, y_data):
x, t = Variable(x_data), Variable(y_data)
y = self.forward_raw(x, train=True)
return F.sigmoid_cross_entropy(y, t)
def forward_one_step(self, x_data, state, continuous=True, nonlinear_q='tanh', nonlinear_p='tanh', output_f = 'sigmoid', gpu=-1):
output = np.zeros( x_data.shape ).astype(np.float32)
nonlinear = {'sigmoid': F.sigmoid, 'tanh': F.tanh, 'softplus': self.softplus, 'relu': F.relu}
nonlinear_f_q = nonlinear[nonlinear_q]
nonlinear_f_p = nonlinear[nonlinear_p]
output_a_f = nonlinear[output_f]
# compute q(z|x)
for i in range(x_data.shape[0]):
x_in_t = Variable(x_data[i].reshape((1, x_data.shape[1])))
hidden_q_t = nonlinear_f_q( self.recog_in_h( x_in_t ) + self.recog_h_h( state['recog_h'] ) )
state['recog_h'] = hidden_q_t
q_mean = self.recog_mean( state['recog_h'] )
q_log_sigma = 0.5 * self.recog_log_sigma( state['recog_h'] )
eps = np.random.normal(0, 1, q_log_sigma.data.shape ).astype(np.float32)
if gpu >= 0:
eps = cuda.to_gpu(eps)
eps = Variable(eps)
z = q_mean + F.exp(q_log_sigma) * eps
# compute p( x | z)
h0 = nonlinear_f_p( self.z(z) )
out= self.output(h0)
x_0 = output_a_f( out )
state['gen_h'] = h0
if gpu >= 0:
np_x_0 = cuda.to_cpu(x_0.data)
output[0] = np_x_0
else:
output[0] = x_0.data
if continuous == True:
rec_loss = F.mean_squared_error(x_0, Variable(x_data[0].reshape((1, x_data.shape[1]))))
else:
rec_loss = F.sigmoid_cross_entropy(out, Variable(x_data[0].reshape((1, x_data.shape[1])).astype(np.int32)))
x_t = x_0
for i in range(1, x_data.shape[0]):
h_t_1 = nonlinear_f_p( self.gen_in_h( x_t ) + self.gen_h_h(state['gen_h']) )
x_t_1 = self.output(h_t_1)
state['gen_h'] = h_t_1
if continuous == True:
output_t = output_a_f( x_t_1 )
rec_loss += F.mean_squared_error(output_t, Variable(x_data[i].reshape((1, x_data.shape[1]))))
else:
out = x_t_1
rec_loss += F.sigmoid_cross_entropy(out, Variable(x_data[i].reshape((1,x_data.shape[1])).astype(np.int32)))
x_t = output_t = output_a_f( x_t_1 )
if gpu >= 0:
np_output_t = cuda.to_cpu(output_t.data)
output[i] = np_output_t
else:
output[i] = output_t.data
KLD = -0.0005 * F.sum(1 + q_log_sigma - q_mean**2 - F.exp(q_log_sigma))
return output, rec_loss, KLD, state
def loss(self, y, t):
sigmoid_y = 1 / (1 + self.mod.exp(-y.data))
mean_squared_error = ((t.data - sigmoid_y) ** 2).sum() / t.data.size
return F.sigmoid_cross_entropy(y, t), mean_squared_error
def concat_losses(p, e):
loss_x = -F.sum(F.log(sum_axis(p))) / numpy.float32(p.data.shape[0])
loss_e = F.sigmoid_cross_entropy(*e)
return loss_x + loss_e
def trainer(G,D,data,len_z=100,n_epoch=10000,pre_epoch=0,batchsize=500,save_interval=1000,
output_dir=None,G_path=None,D_path=None,show=True):
opt_g = optimizers.Adam(alpha=0.0002, beta1=0.5)
opt_d = optimizers.Adam(alpha=0.0002, beta1=0.5)
opt_g.setup(G)
opt_d.setup(D)
opt_g.add_hook(chainer.optimizer.WeightDecay(0.00001))
opt_d.add_hook(chainer.optimizer.WeightDecay(0.00001))
if D_path != None:
serializers.load_hdf5("%s"%(D_path), D)
if G_path != None:
serializers.load_hdf5("%s"%(G_path), G)
n_epoch += pre_epoch
loss_d_mem =np.zeros(n_epoch-pre_epoch)
loss_g_mem =np.zeros(n_epoch-pre_epoch)
for epoch in xrange(pre_epoch,n_epoch):
if epoch%10==0: print 'epoch',epoch
perm = np.arange(len(data))
np.random.shuffle(perm)
for i in xrange(0,len(data),batchsize):
z = Variable(np.random.uniform(-1,1,(batchsize, len_z)).astype(np.float32))
y1 = G(z)
y2 = D(y1)
# discriminator
loss_d = F.sigmoid_cross_entropy(y2,Variable(np.zeros((batchsize,1),dtype=np.int32)))
loss_g = F.sigmoid_cross_entropy(y2,Variable(np.ones((batchsize,1),dtype=np.int32)))
# get images
images = data[perm[i:i+batchsize]]
y2 = D(Variable(images))
loss_d += F.sigmoid_cross_entropy(y2,Variable(np.ones((images.shape[0],1),dtype=np.int32)))
loss_d_mem[epoch-n_epoch] += loss_d.data
loss_g_mem[epoch-n_epoch] += loss_g.data
opt_g.zero_grads()
loss_g.backward()
opt_g.update()
opt_d.zero_grads()
loss_d.backward()
opt_d.update()
#save model
if (epoch+1)%save_interval == 0:
z = Variable(np.random.uniform(-1,1,(10, len_z)).astype(np.float32))
confirm = G(z,False)
if output_dir != None:
serializers.save_hdf5("%s/gan_model_dis%d.h5"%(output_dir,epoch+1), D)
serializers.save_hdf5("%s/gan_model_gen%d.h5"%(output_dir,epoch+1), G)
serializers.save_hdf5("%s/current_gen.h5"%(output_dir), G)
if show:
if D.imshape[0] == 3:
plt.imshow(np.swapaxes(np.swapaxes(confirm.data[0], 0, 2),0,1))
else:
plt.imshow(confirm.data[0].reshape(D.imshape[1],D.imshape[2]),cmap="gray")
plt.axis('off')
plt.savefig('%s/image%d.jpg'%(output_dir,epoch+1))
print '--%d--'%(epoch+1)
print 'p_g :',D(confirm,False).data[0]
print 'p_delta:', D(Variable(images),False).data[0]
print 'done'
return loss_g_mem,loss_d_mem
def forward(self):
x = chainer.Variable(self.x)
t = chainer.Variable(self.t)
return functions.sigmoid_cross_entropy(x, t)
def __call__(self, x, t):
y = self.forward(x)
self.loss = F.sigmoid_cross_entropy(y, t)
return self.loss
def __call__(self, x, t):
y = self.predictor(x)
self.loss = F.sigmoid_cross_entropy(y, t)
return self.loss, y
