def check_invalid_reduce(self, x, t):
with chainer.using_config('use_cudnn', self.use_cudnn):
with self.assertRaises(ValueError):
functions.softmax_cross_entropy(
x, t, self.normalize, self.cache_score,
reduce='unknown_reduce_type',
enable_double_backprop=self.enable_double_backprop)
def train_one(gen, dis, optimizer_gen, optimizer_dis, x_batch, y_batch, gpu_device):
batch_size = len(x_batch)
if gpu_device == None:
xp = np
else:
xp = cuda.cupy
# train generator
y = Variable(xp.asarray(y_batch))
t = Variable(xp.asarray(y_batch + 1))
z = Variable(xp.random.uniform(-1, 1, (batch_size, LATENT_SIZE)).astype(np.float32))
x = gen((z, y))
y1 = dis(x)
loss_gen = F.softmax_cross_entropy(y1, t)
loss_dis = F.softmax_cross_entropy(y1, Variable(xp.zeros(batch_size).astype(np.int32)))
# train discriminator
y2 = dis(Variable(xp.asarray(x_batch)))
loss_dis += F.softmax_cross_entropy(y2, t)
optimizer_gen.zero_grads()
loss_gen.backward()
optimizer_gen.update()
optimizer_dis.zero_grads()
loss_dis.backward()
optimizer_dis.update()
return (float(loss_gen.data), float(loss_dis.data))
def train_one(gen, dis, optimizer_gen, optimizer_dis, x_batch, gpu_device):
batch_size = len(x_batch)
if gpu_device == None:
xp = xp
else:
xp = cuda.cupy
# train generator
z = Variable(xp.random.uniform(-1, 1, (batch_size, LATENT_SIZE)).astype(np.float32))
x = gen(z)
y1 = dis(x)
loss_gen = F.softmax_cross_entropy(y1, Variable(xp.zeros(batch_size).astype(np.int32)))
loss_dis = F.softmax_cross_entropy(y1, Variable(xp.ones(batch_size).astype(np.int32)))
# train discriminator
x2 = Variable(xp.asarray(np.reshape(x_batch, (batch_size, 1, 28, 28))))
y2 = dis(x2)
loss_dis += F.softmax_cross_entropy(y2, Variable(xp.zeros(batch_size).astype(np.int32)))
optimizer_gen.zero_grads()
loss_gen.backward()
optimizer_gen.update()
optimizer_dis.zero_grads()
loss_dis.backward()
optimizer_dis.update()
return (loss_gen.data, loss_dis.data)
def __call__(self, jline, eline):
gh = []
self.H.reset_state()
for w in jline:
wid = self.jvocab[w]
x_k = self.embedx(Variable(np.array([wid], dtype=np.int32)))
h = self.H(x_k)
gh.append(np.copy(h.data[0]))
x_k = self.embedx(Variable(np.array([self.jvocab[EOS]], dtype=np.int32)))
tx = Variable(np.array([self.evocab[eline[0]]], dtype=np.int32))
h = self.H(x_k)
ct = Variable(mk_ct(gh, h.data[0]))
h2 = F.tanh(self.Wc1(ct) + self.Wc2(h))
accum_loss = F.softmax_cross_entropy(self.W(h2), tx)
for i in range(len(eline)):
wid = self.evocab[eline[i]]
x_k = self.embedy(Variable(np.array([wid], dtype=np.int32)))
next_w = eline[i + 1] if i < len(eline) - 1 else EOS
next_wid = self.evocab[next_w]
tx = Variable(np.array([next_wid], dtype=np.int32))
h = self.H(x_k)
ct = Variable(mk_ct(gh, h.data[0]))
h2 = F.tanh(self.Wc1(ct) + self.Wc2(h))
loss = F.softmax_cross_entropy(self.W(h2), tx)
accum_loss += loss
return accum_loss
def forward(self, x_data, y_data, train=True):
x = Variable(x_data, volatile=not train)
t = Variable(y_data, volatile=not train)
h = F.relu(self.bn1_1(self.conv1_1(x)))
h = F.relu(self.bn1_2(self.conv1_2(h)))
h = F.max_pooling_2d(h, 2, stride=2)
h = F.dropout(h, ratio=0.25, train=train)
h = F.relu(self.bn2_1(self.conv2_1(h)))
h = F.relu(self.bn2_2(self.conv2_2(h)))
h = F.max_pooling_2d(h, 2, stride=2)
h = F.dropout(h, ratio=0.25, train=train)
h = F.relu(self.bn3_1(self.conv3_1(h)))
h = F.relu(self.bn3_2(self.conv3_2(h)))
h = F.relu(self.bn3_3(self.conv3_3(h)))
h = F.relu(self.bn3_4(self.conv3_4(h)))
h = F.max_pooling_2d(h, 2, stride=2)
h = F.dropout(h, ratio=0.25, train=train)
h = F.dropout(F.relu(self.fc4(h)), train=train, ratio=0.5)
h = F.dropout(F.relu(self.fc5(h)), train=train, ratio=0.5)
h = self.fc6(h)
if train:
return F.softmax_cross_entropy(h, t), F.accuracy(h, t)
else:
return F.softmax_cross_entropy(h, t), F.accuracy(h, t), h
def forward(self, x_data, y_data, train=True):
x = Variable(x_data, volatile=not train)
t = Variable(y_data, volatile=not train)
h = self.prelu1_1(self.bn1_1(self.conv1_1(x)))
h = self.prelu1_2(self.bn1_2(self.conv1_2(h)))
h = F.max_pooling_2d(h, 2, stride=2)
h = self.prelu2_1(self.bn2_1(self.conv2_1(h)))
h = self.prelu2_2(self.bn2_2(self.conv2_2(h)))
h = F.max_pooling_2d(h, 2, stride=2)
h = self.prelu3_1(self.conv3_1(h))
h = self.prelu3_2(self.conv3_2(h))
h = self.prelu3_3(self.conv3_3(h))
h = F.max_pooling_2d(h, 2, stride=1)
h = self.prelu4_1(self.conv4_1(h))
h = self.prelu4_2(self.conv4_2(h))
h = self.prelu4_3(self.conv4_3(h))
h = F.max_pooling_2d(h, 2, stride=1)
h = self.prelu5_1(self.conv5_1(h))
h = self.prelu5_2(self.conv5_2(h))
h = self.prelu5_3(self.conv5_3(h))
h = F.max_pooling_2d(h, 2, stride=1)
h = F.dropout(self.prelu6(self.fc6(h)), train=train, ratio=0.5)
h = F.dropout(self.prelu7(self.fc7(h)), train=train, ratio=0.5)
h = self.fc8(h)
if train:
return F.softmax_cross_entropy(h, t), F.accuracy(h, t)
else:
return F.softmax_cross_entropy(h, t), F.accuracy(h, t), h
def forward(self, x_data, y_data, train=True):
x = chainer.Variable(x_data, volatile=not train)
t = chainer.Variable(y_data, volatile=not train)
h = F.max_pooling_2d(
F.relu(self.norm1(self.conv1(x))), 3, stride=2, pad=1)
h = F.max_pooling_2d(
F.relu(self.norm2(self.conv2(h))), 3, stride=2, pad=1)
h = self.inc3a(h)
h = self.inc3b(h)
h = self.inc3c(h)
h = self.inc4a(h)
a = F.average_pooling_2d(h, 5, stride=3)
a = F.relu(self.norma(self.conva(a)))
a = F.relu(self.norma2(self.lina(a)))
a = self.outa(a)
a = F.softmax_cross_entropy(a, t)
h = self.inc4b(h)
h = self.inc4c(h)
h = self.inc4d(h)
b = F.average_pooling_2d(h, 5, stride=3)
b = F.relu(self.normb(self.convb(b)))
b = F.relu(self.normb2(self.linb(b)))
b = self.outb(b)
b = F.softmax_cross_entropy(b, t)
h = self.inc4e(h)
h = self.inc5a(h)
h = F.average_pooling_2d(self.inc5b(h), 7)
h = self.out(h)
return 0.3 * (a + b) + F.softmax_cross_entropy(h, t), F.accuracy(h, t)
def forward(self, x_data, y_data, train=True, models=None):
VGG_mini = models["VGG_mini"]
VGG_mini2 = models["VGG_mini2"]
VGG_mini3 = models["VGG_mini3"]
x = Variable(x_data, volatile=not train)
t = Variable(y_data, volatile=not train)
h = F.relu(self.conv1_1(x))
h = F.relu(self.conv1_2(h))
h = F.relu(self.conv1_3(h))
h = F.relu(self.conv1_4(h))
h = F.max_pooling_2d(h, 2, stride=2)
h = F.dropout(h, ratio=0.25, train=train)
h = F.relu(self.conv1_5(h))
h = F.max_pooling_2d(h, 2, stride=2)
h = F.dropout(h, ratio=0.25, train=train)
h = self.fc(h)
if train:
return F.softmax_cross_entropy(h, t), F.accuracy(h, t)
else:
# return F.softmax_cross_entropy(h, t), F.accuracy(h, t), h
return F.softmax_cross_entropy(h, t), F.accuracy(h, t)
def test_variable_assertion(self):
wrong_inst_class_weight = chainer.Variable(
numpy.array([0, 0], dtype='f'))
with self.assertRaises(ValueError):
functions.softmax_cross_entropy(
self.x, self.t, class_weight=wrong_inst_class_weight,
enable_double_backprop=self.enable_double_backprop)
def d_norm(flag, dis, img1, img2):
yl = dis(img1, img2)
if flag == 0:
return F.softmax_cross_entropy(yl, Variable(xp.zeros(batchsize, dtype=np.int32)))
elif flag == 1:
return F.softmax_cross_entropy(yl, Variable(xp.ones(batchsize, dtype=np.int32)))
else:
raise "norm flag should be either 0 / 1"
def test_NNET_train(self):
struct = 'w2vec(/project/nakamura-lab01/Work/truong-dq/chainer/vidnn/exp/word2vec_truecase_200/vectors.bin):lstm(200-2):linear(2-2)'
# struct = 'embed(2-200):lstm(200-2):linear(2-2)'
nnet = NNET_Model.parse_structure(struct)
# Testing variable
test_var = chainer.Variable(np.asarray([1], dtype=np.int32))
output_before_train = nnet(test_var).data
inp = chainer.Variable(np.asarray([1], dtype=np.int32))
target = chainer.Variable(np.asarray([0], dtype=np.int32))
output = nnet(inp)
loss = F.softmax_cross_entropy(output, target)
optimizer = optimizers.SGD(lr=0.1)
optimizer.setup(nnet)
optimizer.zero_grads()
loss.backward()
optimizer.update()
inp = chainer.Variable(np.asarray([0], dtype=np.int32))
target = chainer.Variable(np.asarray([0], dtype=np.int32))
output = nnet(inp)
loss = F.softmax_cross_entropy(output, target)
optimizer = optimizers.SGD(lr=0.1)
optimizer.setup(nnet)
optimizer.zero_grads()
loss.backward()
optimizer.update()
nnet.save('test_output')
nnet_2 = NNET_Model.load('test_output')
nnet.forget_history()
nnet_2.forget_history()
np.testing.assert_equal(nnet[0][1].W.data, nnet_2[0][1].W.data)
np.testing.assert_equal(nnet[1][1].upward.W.data, nnet_2[1][1].upward.W.data)
np.testing.assert_equal(nnet[1][1].lateral.W.data, nnet_2[1][1].lateral.W.data)
np.testing.assert_equal(nnet[1][1].upward.b.data, nnet_2[1][1].upward.b.data)
output_after_train = nnet(test_var).data
output_after_load = nnet_2(test_var).data
after_first_layer_nnet = nnet[0][1](test_var)
after_first_layer_nnet_2 = nnet_2[0][1](test_var)
np.testing.assert_equal(after_first_layer_nnet.data, after_first_layer_nnet_2.data)
after_first_layer_nnet.volatile = False
after_first_layer_nnet_2.volatile = False
after_second_layer_nnet = nnet[1][1](after_first_layer_nnet)
after_second_layer_nnet_2 = nnet_2[1][1](after_first_layer_nnet_2)
np.testing.assert_equal(after_second_layer_nnet.data, after_second_layer_nnet_2.data)
assert (output_before_train != output_after_train).any()
assert (output_before_train != output_after_load).any()
np.testing.assert_equal(output_after_train, output_after_load)
def forward(self, *inputs):
batch = len(inputs) // 6
lefts = inputs[0: batch]
rights = inputs[batch: batch * 2]
dests = inputs[batch * 2: batch * 3]
labels = inputs[batch * 3: batch * 4]
sequences = inputs[batch * 4: batch * 5]
leaf_labels = inputs[batch * 5: batch * 6]
inds = numpy.argsort([-len(l) for l in lefts])
# Sort all arrays in descending order and transpose them
lefts = F.transpose_sequence([lefts[i] for i in inds])
rights = F.transpose_sequence([rights[i] for i in inds])
dests = F.transpose_sequence([dests[i] for i in inds])
labels = F.transpose_sequence([labels[i] for i in inds])
sequences = F.transpose_sequence([sequences[i] for i in inds])
leaf_labels = F.transpose_sequence(
[leaf_labels[i] for i in inds])
batch = len(inds)
maxlen = len(sequences)
loss = 0
count = 0
correct = 0
stack = self.xp.zeros(
(batch, maxlen * 2, self.n_units), self.xp.float32)
for i, (word, label) in enumerate(zip(sequences, leaf_labels)):
batch = word.shape[0]
es = self.leaf(word)
ds = self.xp.full((batch,), i, self.xp.int32)
y = self.label(es)
loss += F.softmax_cross_entropy(y, label, normalize=False) * batch
count += batch
predict = self.xp.argmax(y.array, axis=1)
correct += (predict == label.array).sum()
stack = thin_stack.thin_stack_set(stack, ds, es)
for left, right, dest, label in zip(lefts, rights, dests, labels):
l, stack = thin_stack.thin_stack_get(stack, left)
r, stack = thin_stack.thin_stack_get(stack, right)
o = self.node(l, r)
y = self.label(o)
batch = l.shape[0]
loss += F.softmax_cross_entropy(y, label, normalize=False) * batch
count += batch
predict = self.xp.argmax(y.array, axis=1)
correct += (predict == label.array).sum()
stack = thin_stack.thin_stack_set(stack, dest, o)
loss /= count
reporter.report({'loss': loss}, self)
reporter.report({'total': count}, self)
reporter.report({'correct': correct}, self)
return loss
def train_dcgan_labeled(images, gen, dis):
o_gen = optimizers.Adam(alpha=0.0002, beta1=0.5)
o_dis = optimizers.Adam(alpha=0.0002, beta1=0.5)
o_gen.setup(gen)
o_dis.setup(dis)
o_gen.add_hook(chainer.optimizer.WeightDecay(0.00001))
o_dis.add_hook(chainer.optimizer.WeightDecay(0.00001))
zeros = Variable(xp.zeros(batchsize, dtype=np.int32))
ones = Variable(xp.ones(batchsize, dtype=np.int32))
for epoch in tqdm(range(n_epoch)):
# discriminator
# 0: from dataset
# 1: from noise
# train generator
z = xp.random.uniform(-1, 1, (batchsize, nz), dtype=np.float32)
z = Variable(z)
x = gen(z)
yl = dis(x)
L_gen = F.softmax_cross_entropy(yl, zeros)
L_dis = F.softmax_cross_entropy(yl, ones)
# train discriminator
x = generate_data(images)
yl = dis(x)
L_dis += F.softmax_cross_entropy(yl, zeros)
o_gen.zero_grads()
L_gen.backward()
o_gen.update()
o_dis.zero_grads()
L_dis.backward()
o_dis.update()
if epoch % image_save_interval == 0 and epoch > 0:
z = zvis
z[50:, :] = xp.random.uniform(-1, 1, (50, nz), dtype=np.float32)
z = Variable(z)
x = gen(z, test=True)
filename = '{}/vis_{}.png'.format(out_image_dir, epoch)
generate_and_save(filename, x.data.get())
path = join(out_model_dir, "dcgan_model_dis_{}.h5".format(epoch))
serializers.save_hdf5(path, dis)
path = join(out_model_dir, "dcgan_model_gen_%d.h5".format(epoch))
serializers.save_hdf5(path, gen)
path = join(out_model_dir, "dcgan_state_dis_%d.h5".format(epoch))
serializers.save_hdf5(path, o_dis)
path = join(out_model_dir, "dcgan_state_gen_%d.h5".format(epoch))
serializers.save_hdf5(path, o_gen)
def __call__(self, x, t):
self.clear()
test = not self.train
h = F.max_pooling_2d(
F.relu(self.norm1(self.conv1(x), test=test)), 3, stride=2, pad=1)
h = F.max_pooling_2d(
F.relu(self.norm2(self.conv2(h), test=test)), 3, stride=2, pad=1)
h = self.inc3a(h)
h = self.inc3b(h)
h = self.inc3c(h)
h = self.inc4a(h)
a = F.average_pooling_2d(h, 5, stride=3)
a = F.relu(self.norma(self.conva(a), test=test))
a = F.relu(self.norma2(self.lina(a), test=test))
a = self.outa(a)
self.loss1 = F.softmax_cross_entropy(a, t)
h = self.inc4b(h)
h = self.inc4c(h)
h = self.inc4d(h)
b = F.average_pooling_2d(h, 5, stride=3)
b = F.relu(self.normb(self.convb(b), test=test))
b = F.relu(self.normb2(self.linb(b), test=test))
b = self.outb(b)
self.loss2 = F.softmax_cross_entropy(b, t)
h = self.inc4e(h)
h = self.inc5a(h)
h = F.average_pooling_2d(self.inc5b(h), 7)
h = self.out(h)
self.loss3 = F.softmax_cross_entropy(h, t)
self.loss = 0.3 * (self.loss1 + self.loss2) + self.loss3
self.accuracy = F.accuracy(h, t)
shishi = F.softmax(h)
# kankan = shishi.data[0].tolist()
# categories = np.loadtxt("labels.txt", str, delimiter="\t")
# top_k = 10
# prediction = zip(kankan,categories)
# for feifei in categories:
# print(feifei)
# prediction.sort(cmp=lambda x,y: cmp(x[0],y[0]),reverse=True)
# cuowushuchu = ('cuowuchushu.txt','w')
# for rank,(score,name) in enumerate(prediction[:3],start=1):
# print('#%d | %s | %4.1f%%' % (rank,name,score * 100))
# print('\n')
# for rank,(score,name) in enumerate(prediction[:2],start=1):
# feijigege = score * 100
# cuowushuchu.write(str(name)+' '+str(feijigege))
# cuowushuchu.close()
return shishi
def check_value_check(self, x_data, t_data, use_cudnn):
x = chainer.Variable(x_data)
t = chainer.Variable(t_data)
if self.valid:
# Check if it throws nothing
functions.softmax_cross_entropy(x, t, use_cudnn)
else:
with self.assertRaises(ValueError):
functions.softmax_cross_entropy(x, t, use_cudnn)
def forward(self, x_img, x_doc, y_data, train=True):
x_img = cuda.cupy.asarray(x_img)
x_doc = cuda.cupy.asarray(x_doc)
y_data = cuda.cupy.asarray(y_data)
img, doc, t = Variable(x_img), Variable(x_doc), Variable(y_data)
h = F.relu(self.conv1(img))
h = F.local_response_normalization(
F.max_pooling_2d(h, 3, stride=2), n=5)
h = F.relu(self.conv2_reduce(h))
h = F.relu(self.conv2(h))
h = F.max_pooling_2d(
F.local_response_normalization(h, n=5), 3, stride=2)
h = self.inc3a(h)
h = self.inc3b(h)
h = F.max_pooling_2d(h, 3, stride=2)
h = self.inc4a(h)
l = F.average_pooling_2d(h, 5, stride=3)
l = F.relu(self.loss1_conv(l))
l = F.relu(self.loss1_fc1(l))
l = self.loss1_fc2(l)
self.loss1 = F.softmax_cross_entropy(l, t)
h = self.inc4b(h)
h = self.inc4c(h)
h = self.inc4d(h)
l = F.average_pooling_2d(h, 5, stride=3)
l = F.relu(self.loss2_conv(l))
l = F.relu(self.loss2_fc1(l))
l = self.loss2_fc2(l)
self.loss2 = F.softmax_cross_entropy(l, t)
h = self.inc4e(h)
h = F.max_pooling_2d(h, 3, stride=2)
h = self.inc5a(h)
h = self.inc5b(h)
h = F.average_pooling_2d(h, 7, stride=1)
h = self.loss3_fc1(F.dropout(h, 0.4, train=train))
h2 = F.relu(self.doc_fc1(F.dropout(doc, train=train)))
h2 = F.relu(self.doc_fc2(h2))
b = F.relu(self.bi1(h, h2))
h = self.loss3_fc2(b)
self.loss3 = F.softmax_cross_entropy(h, t)
if train:
return 0.3 * (self.loss1 + self.loss2) + self.loss3
else:
return F.accuracy(h, t)
def forward(self, x_data, y_data, train=True):
x, t = Variable(x_data), Variable(y_data)
h = F.max_pooling_2d(F.relu(self.bn1(self.conv1(x))), 3, stride=2)
h = F.max_pooling_2d(F.relu(self.bn2(self.conv2(h))), 3, stride=2)
h = F.max_pooling_2d(F.relu(self.conv3(h)), 3, stride=2)
h = self.fc4(h)
if train:
return F.softmax_cross_entropy(h, t), F.accuracy(h, t)
else:
return F.softmax_cross_entropy(h, t), F.accuracy(h, t), h
def compute_loss(self, input_ids, input_mask, token_type_ids,
start_positions, end_positions):
(start_logits, end_logits) = self.__call__(
input_ids, input_mask, token_type_ids)
start_loss = F.softmax_cross_entropy(start_logits, start_positions)
end_loss = F.softmax_cross_entropy(end_logits, end_positions)
total_loss = (start_loss + end_loss) / 2.0
chainer.report({'loss': total_loss.array}, self)
accuracy = (check_answers(start_logits, start_positions) *
check_answers(end_logits, end_positions, start_positions)).mean()
chainer.report({'accuracy': accuracy}, self)
return total_loss
def train(self, x_img, x_doc, y_data, regression, gpu=True, useImage=True, useDoc=True):
xp = cuda.cupy if gpu else np
x_img = xp.asarray(x_img)
x_doc = xp.asarray(x_doc)
y_data = xp.asarray(y_data)
img, doc, t = Variable(x_img), Variable(x_doc), Variable(y_data)
y = self.model.forward(img, doc, regression=regression, useImage=useImage, useDoc=useDoc)
# calc loss
if useImage:
if regression:
a = self.toLog(y["a"], xp)
b = self.toLog(y["b"], xp)
h = self.toLog(y["h"], xp)
t = self.toLog(t, xp)
self.loss1 = F.mean_squared_error(a, t)
self.loss2 = F.mean_squared_error(b, t)
self.loss3 = F.mean_squared_error(h, t)
else:
a = y["a"]
b = y["b"]
h = y["h"]
self.loss1 = F.softmax_cross_entropy(a, t)
self.loss2 = F.softmax_cross_entropy(b, t)
self.loss3 = F.softmax_cross_entropy(h, t)
loss = 0.3 * (self.loss1 + self.loss2) + self.loss3
else:
if regression:
h = self.toLog(y, xp)
t = self.toLog(t, xp)
self.loss1 = F.mean_squared_error(h, t)
else:
h = y
self.loss1 = F.softmax_cross_entropy(y, t)
loss = self.loss1
# random select optimizer
rnd = np.random.randint(0, len(self.myOptimizers))
self.optimizer = self.myOptimizers[rnd]
self.optimizer.setup(self.model)
self.optimizer.zero_grads()
loss.backward()
self.optimizer.update()
if regression:
h = np.array(cuda.to_cpu(h.data)).reshape((len(h)))
t = np.array(cuda.to_cpu(t.data)).reshape((len(t)))
return loss.data, h, t
else:
return loss.data, F.accuracy(h, t).data, []
def forward(self, x_data, y_data, train=True):
x = chainer.Variable(x_data, volatile=not train)
t = chainer.Variable(y_data, volatile=not train)
h = F.relu(self.conv1(x))
h = F.local_response_normalization(
F.max_pooling_2d(h, 3, stride=2), n=5)
h = F.relu(self.conv2_reduce(h))
h = F.relu(self.conv2(h))
h = F.max_pooling_2d(
F.local_response_normalization(h, n=5), 3, stride=2)
h = self.inc3a(h)
h = self.inc3b(h)
h = F.max_pooling_2d(h, 3, stride=2)
h = self.inc4a(h)
if train:
loss1 = F.average_pooling_2d(h, 5, stride=3)
loss1 = F.relu(self.loss1_conv(loss1))
loss1 = F.relu(self.loss1_fc1(loss1))
loss1 = self.loss1_fc2(loss1)
loss1 = F.softmax_cross_entropy(loss1, t)
h = self.inc4b(h)
h = self.inc4c(h)
h = self.inc4d(h)
if train:
loss2 = F.average_pooling_2d(h, 5, stride=3)
loss2 = F.relu(self.loss2_conv(loss2))
loss2 = F.relu(self.loss2_fc1(loss2))
loss2 = self.loss2_fc2(loss2)
loss2 = F.softmax_cross_entropy(loss2, t)
h = self.inc4e(h)
h = F.max_pooling_2d(h, 3, stride=2)
h = self.inc5a(h)
h = self.inc5b(h)
h = F.dropout(F.average_pooling_2d(h, 7, stride=1), 0.4, train=train)
h = self.loss3_fc(h)
loss3 = F.softmax_cross_entropy(h, t)
if train:
loss = 0.3 * (loss1 + loss2) + loss3
else:
loss = loss3
accuracy = F.accuracy(h, t)
return loss, accuracy
def train_word_embedding_batch(self, char_ids_batch): xp = self.xp word_vec = self.encode_word_batch(char_ids_batch) batchsize = char_ids_batch.shape[0] char_ids_batch = char_ids_batch.T # reconstruction loss loss_reconstruction = 0 self.word_decoder_lstm.reset_state() prev_y = None for i in xrange(char_ids_batch.shape[0]): if prev_y is None: prev_y = Variable(xp.zeros((batchsize, self.char_embed_size), dtype=xp.float32)) dec_in = F.concat((word_vec, prev_y)) y = self.word_decoder_lstm(dec_in, test=False) target = Variable(char_ids_batch[i]) if self.gpu_enabled: target.to_gpu() loss = F.softmax_cross_entropy(y, target) prev_y = self.embed_id(target) loss_reconstruction += loss self.zero_grads_generator() loss_reconstruction.backward() self.update_generator() # adversarial loss ## 0: from encoder ## 1: from noise real_z = self.sample_z(batchsize, self.word_embed_size) fake_z = word_vec y_fake = self.discriminator(fake_z, test=False) ## train generator loss_generator = F.softmax_cross_entropy(y_fake, Variable(xp.ones((batchsize,), dtype=xp.int32))) self.zero_grads_generator() loss_generator.backward() self.update_generator() # train discriminator y_real = self.discriminator(real_z, test=False) loss_discriminator = F.softmax_cross_entropy(y_fake, Variable(xp.zeros((batchsize,), dtype=xp.int32))) loss_discriminator += F.softmax_cross_entropy(y_real, Variable(xp.ones((batchsize,), dtype=xp.int32))) self.optimizer_discriminator.zero_grads() loss_discriminator.backward() self.optimizer_discriminator.update() return float(loss_reconstruction.data), float(loss_generator.data), float(loss_discriminator.data)
def __call__(self, x, t):
h = F.relu(self.conv1(x))
h = F.local_response_normalization(
F.max_pooling_2d(h, 3, stride=2), n=5)
h = F.relu(self.conv2_reduce(h))
h = F.relu(self.conv2(h))
h = F.max_pooling_2d(
F.local_response_normalization(h, n=5), 3, stride=2)
h = self.inc3a(h)
h = self.inc3b(h)
h = F.max_pooling_2d(h, 3, stride=2)
h = self.inc4a(h)
l = F.average_pooling_2d(h, 5, stride=3)
l = F.relu(self.loss1_conv(l))
l = F.relu(self.loss1_fc1(l))
l = self.loss1_fc2(l)
loss1 = F.softmax_cross_entropy(l, t)
h = self.inc4b(h)
h = self.inc4c(h)
h = self.inc4d(h)
l = F.average_pooling_2d(h, 5, stride=3)
l = F.relu(self.loss2_conv(l))
l = F.relu(self.loss2_fc1(l))
l = self.loss2_fc2(l)
loss2 = F.softmax_cross_entropy(l, t)
h = self.inc4e(h)
h = F.max_pooling_2d(h, 3, stride=2)
h = self.inc5a(h)
h = self.inc5b(h)
h = F.average_pooling_2d(h, 7, stride=1)
h = self.loss3_fc(F.dropout(h, 0.4))
loss3 = F.softmax_cross_entropy(h, t)
loss = 0.3 * (loss1 + loss2) + loss3
accuracy = F.accuracy(h, t)
chainer.report({
'loss': loss,
'loss1': loss1,
'loss2': loss2,
'loss3': loss3,
'accuracy': accuracy
}, self)
return loss
def check_value_check(self, x_data, t_data, use_cudnn):
x = chainer.Variable(x_data)
t = chainer.Variable(t_data)
with chainer.using_config('use_cudnn', use_cudnn):
if self.valid:
# Check if it throws nothing
functions.softmax_cross_entropy(
x, t, enable_double_backprop=self.enable_double_backprop)
else:
with self.assertRaises(ValueError):
functions.softmax_cross_entropy(
x, t,
enable_double_backprop=self.enable_double_backprop)
def __call__(self, z, x):
batchsize = z.data.shape[0]
# generate
x_gen = self.gen(z)
y_gen = self.dis(x_gen)
loss_gen = F.softmax_cross_entropy(y_gen, Variable(self._get_zeros(batchsize)))
loss_dis = F.softmax_cross_entropy(y_gen, Variable(self._get_ones(batchsize)))
# discriminate
y = self.dis(x)
loss_dis += F.softmax_cross_entropy(y, Variable(self._get_zeros(batchsize)))
return loss_gen, loss_dis
def __call__(self, x, t):
h = F.relu(self['conv1/7x7_s2'](x))
h = F.local_response_normalization(
F.max_pooling_2d(h, 3, stride=2), n=5, alpha=(1e-4)/5, k=1)
h = F.relu(self['conv2/3x3_reduce'](h))
h = F.relu(self['conv2/3x3'](h))
h = F.max_pooling_2d(F.local_response_normalization(
h, n=5, alpha=(1e-4)/5, k=1), 3, stride=2)
h = self.call_inception(h, 'inception_3a')
h = self.call_inception(h, 'inception_3b')
h = F.max_pooling_2d(h, 3, stride=2)
h = self.call_inception(h, 'inception_4a')
l = F.average_pooling_2d(h, 5, stride=3)
l = F.relu(self['loss1/conv'](l))
l = F.dropout(F.relu(self['loss1/fc'](l)), 0.7, train=self.train)
l = self['loss1/classifier'](l)
loss1 = F.softmax_cross_entropy(l, t)
h = self.call_inception(h, 'inception_4b')
h = self.call_inception(h, 'inception_4c')
h = self.call_inception(h, 'inception_4d')
l = F.average_pooling_2d(h, 5, stride=3)
l = F.relu(self['loss2/conv'](l))
l = F.dropout(F.relu(self['loss2/fc'](l)), 0.7, train=self.train)
l = self['loss2/classifier'](l)
loss2 = F.softmax_cross_entropy(l, t)
h = self.call_inception(h, 'inception_4e')
h = F.max_pooling_2d(h, 3, stride=2)
h = self.call_inception(h, 'inception_5a')
h = self.call_inception(h, 'inception_5b')
h = F.average_pooling_2d(h, 7, stride=1)
h = self['loss3/classifier'](F.dropout(h, 0.4, train=self.train))
loss3 = F.softmax_cross_entropy(h, t)
loss = 0.3 * (loss1 + loss2) + loss3
accuracy = F.accuracy(h, t)
chainer.report({
'loss': loss,
'loss1': loss1,
'loss2': loss2,
'loss3': loss3,
'accuracy': accuracy
}, self)
return loss
def train_gen(gen, dis, optimizer_gen, optimizer_dis, x_batch, gpu_device):
batch_size = len(x_batch)
if gpu_device == None:
xp = xp
else:
xp = cuda.cupy
z = Variable(xp.random.uniform(-1, 1, (batch_size, LATENT_SIZE)).astype(np.float32))
x = gen(z)
y1 = dis(x)
loss_gen = F.softmax_cross_entropy(y1, Variable(xp.zeros(batch_size).astype(np.int32)))
loss_dis = F.softmax_cross_entropy(y1, Variable(xp.ones(batch_size).astype(np.int32)))
optimizer_gen.zero_grads()
loss_gen.backward()
optimizer_gen.update()
return loss_gen.data
def __call__(self, x, t):
test = not self.train
finetune = self.finetune
h = self.call_conv_bn_sc(x, 'conv1/7x7_s2', test=test, finetune=finetune)
h = F.max_pooling_2d(h, 3, stride=2, pad=1)
h = self.call_conv_bn_sc(h, 'conv2/3x3_reduce', test=test, finetune=finetune)
h = self.call_conv_bn_sc(h, 'conv2/3x3', test=test)
h = F.max_pooling_2d(h, 3, stride=2, pad=1)
h = self.call_inception_bn(h, 'inception_3a', test=test, finetune=finetune)
h = self.call_inception_bn(h, 'inception_3b', test=test, finetune=finetune)
h = self.call_inception_bn(h, 'inception_3c', test=test, finetune=finetune)
a = F.average_pooling_2d(h, 5, stride=3)
a = self.call_conv_bn_sc(a, 'loss1/conv', test=test, finetune=finetune)
a = self.call_fc_bn_sc(a, 'loss1/fc', test=test, finetune=finetune)
a = self['loss1/classifier'](a)
loss1 = F.softmax_cross_entropy(a, t)
h = self.call_inception_bn(h, 'inception_4a', test=test, finetune=finetune)
h = self.call_inception_bn(h, 'inception_4b', test=test, finetune=finetune)
h = self.call_inception_bn(h, 'inception_4c', test=test, finetune=finetune)
h = self.call_inception_bn(h, 'inception_4d', test=test, finetune=finetune)
h = self.call_inception_bn(h, 'inception_4e', test=test, finetune=finetune)
b = F.average_pooling_2d(h, 5, stride=3)
b = self.call_conv_bn_sc(b, 'loss2/conv', test=test, finetune=finetune)
b = self.call_fc_bn_sc(b, 'loss2/fc', test=test, finetune=finetune)
b = self['loss2/classifier'](b)
loss2 = F.softmax_cross_entropy(b, t)
h = self.call_inception_bn(h, 'inception_5a', test=test, finetune=finetune)
h = self.call_inception_bn(h, 'inception_5b', test=test, finetune=finetune)
h = F.average_pooling_2d(h, 7, stride=1)
h = self['loss3/classifier'](h)
loss3 = F.softmax_cross_entropy(h, t)
loss = 0.3 * (loss1 + loss2) + loss3
accuracy = F.accuracy(h, t)
chainer.report({
'loss': loss,
'loss1': loss1,
'loss2': loss2,
'loss3': loss3,
'accuracy': accuracy
}, self)
return loss
def update_parameter_by_meta_learner(
self, model_params, loss,
x_l0, x_l1, y_l):
# Forward meta-learner
namedparams = model_params
for i, elm in enumerate(namedparams.items()): # parameter-loop
k, p = elm
with cuda.get_device_from_id(self.device):
shape = p.shape
xp = cuda.get_array_module(p.data)
x = p.grad
grad = xp.reshape(x, (np.prod(shape), ))
meta_learner = self.meta_learners[i]
g = meta_learner(Variable(grad)) # forward
w = p - F.reshape(g, shape)
self.model_params[k] = w
# Train meta-learner with main objective
y_pred = self.model(x_l0, self.model_params)
loss_ce = F.softmax_cross_entropy(y_pred, y_l)
self.cleargrads() # need to clear W'grad due to loss_rec.backward
for meta_learner in self.meta_learners:
meta_learner.cleargrads()
loss_ce.backward(retain_grad=True)
for opt in self.opt_meta_learners:
opt.update()
loss_ce.unchain_backward() #TODO: here is a proper place to unchain?
def forward(self, xs, ys):
xs = [x[::-1] for x in xs]
eos = self.xp.array([EOS], numpy.int32)
ys_in = [F.concat([eos, y], axis=0) for y in ys]
ys_out = [F.concat([y, eos], axis=0) for y in ys]
# Both xs and ys_in are lists of arrays.
exs = sequence_embed(self.embed_x, xs)
eys = sequence_embed(self.embed_y, ys_in)
batch = len(xs)
# None represents a zero vector in an encoder.
hx, cx, _ = self.encoder(None, None, exs)
_, _, os = self.decoder(hx, cx, eys)
# It is faster to concatenate data before calculating loss
# because only one matrix multiplication is called.
concat_os = F.concat(os, axis=0)
concat_ys_out = F.concat(ys_out, axis=0)
loss = F.sum(F.softmax_cross_entropy(
self.W(concat_os), concat_ys_out, reduce='no')) / batch
chainer.report({'loss': loss}, self)
n_words = concat_ys_out.shape[0]
perp = self.xp.exp(loss.array * batch / n_words)
chainer.report({'perp': perp}, self)
return loss
def forward(x_data, y_data, model,train=True):
# Neural net architecture
#x, t = chainer.Variable(x_data), chainer.Variable(y_data)
t = chainer.Variable(y_data)
x = {}
for n in range(500):
x[n] = chainer.Variable(x_data[n])
h = {}
initial_V = {}
initial_V_relu = {}
for nameint in range(len(l_name)-2):
initial_V[nameint] = model[l_name[nameint]](x[nameint])
#initial_V_relu[nameint] = F.relu(initial_V[nameint])
#initial_V_relu[nameint] = F.sigmoid(initial_V[nameint])
initial_V_relu[nameint] = F.tanh(initial_V[nameint])
#h[nameint] = F.dropout(F.relu(initial_V[nameint]), train=train)
#h[nameint] = F.dropout(F.sigmoid(initial_V[nameint]), train=train)
h[nameint] = F.dropout(F.tanh(initial_V[nameint]), train=train)
#h[nameint] = F.relu(model[l_name[nameint]](x[nameint]))
#h6 = F.dropout(F.relu(model.l501(Returnharray(h))), train=train)
#h6 = F.dropout(F.sigmoid(model.l501(Returnharray(h))), train=train)
h6 = F.dropout(F.tanh(model.l501(Returnharray(h))), train=train)
y = model.l502(h6)
y_pre = (y.data.argmax(axis = 1))
return F.softmax_cross_entropy(y, t), F.accuracy(y, t),y_pre,initial_V,initial_V_relu
def main():
# load MNIST images
images, labels = dataset.load_train_images()
# config
config = model.config
# settings
max_epoch = 1000
num_trains_per_epoch = 5000
num_validation_data = 10000
batchsize = 128
# seed
np.random.seed(args.seed)
if args.gpu_device != -1:
cuda.cupy.random.seed(args.seed)
# save validation accuracy per epoch
csv_results = []
# create semi-supervised split
training_images, training_labels, validation_images, validation_labels = dataset.split_data(
images, labels, num_validation_data, seed=args.seed)
# training
progress = Progress()
for epoch in xrange(1, max_epoch):
progress.start_epoch(epoch, max_epoch)
sum_loss = 0
for t in xrange(num_trains_per_epoch):
# sample from data distribution
image_batch, label_batch = dataset.sample_data(training_images,
training_labels,
batchsize,
binarize=False)
distribution = model.discriminate(image_batch, apply_softmax=False)
loss = F.softmax_cross_entropy(distribution,
model.to_variable(label_batch))
sum_loss += float(loss.data)
model.backprop(loss)
if t % 10 == 0:
progress.show(t, num_trains_per_epoch, {})
model.save(args.model_dir)
train_accuracy = compute_accuracy(training_images, training_labels)
validation_accuracy = compute_accuracy(validation_images,
validation_labels)
progress.show(
num_trains_per_epoch, num_trains_per_epoch, {
"loss": sum_loss / num_trains_per_epoch,
"accuracy (validation)": validation_accuracy,
"accuracy (train)": train_accuracy,
})
# write accuracy to csv
csv_results.append(
[epoch, validation_accuracy,
progress.get_total_time()])
data = pd.DataFrame(csv_results)
data.columns = ["epoch", "accuracy", "min"]
data.to_csv("{}/result.csv".format(args.model_dir))
def __call__(self, x, t, train=True):
y = self.fwd(x, train)
return F.softmax_cross_entropy(y, t), F.accuracy(y, t)
def forward(self,
x_img,
x_doc,
y_data,
train=True,
regression=False,
predict=False,
gpu=True):
test = not train
xp = cuda.cupy if gpu else np
x_img = xp.asarray(x_img)
y_data = xp.asarray(y_data)
img, t = Variable(x_img), Variable(y_data)
if regression and not predict:
t = self.toLog(t)
#t.data = cuda.cupy.asarray(t.data, dtype=cuda.cupy.float32).reshape((20,1))
h = F.max_pooling_2d(F.relu(self.norm1(self.conv1(img), test=test)),
3,
stride=2,
pad=1)
h = F.max_pooling_2d(F.relu(self.norm2(self.conv2(h), test=test)),
3,
stride=2,
pad=1)
h = self.inc3a(h)
h = self.inc3b(h)
h = self.inc3c(h)
h = self.inc4a(h)
if not predict:
a = F.average_pooling_2d(h, 5, stride=3)
a = F.relu(self.norma(self.conva(a), test=test))
a = F.relu(self.norma2(self.lina(a), test=test))
a = self.outa(a)
if regression:
#a = self.toLog(a)
self.loss1 = F.mean_squared_error(a, t)
else:
self.loss1 = F.softmax_cross_entropy(a, t)
h = self.inc4b(h)
h = self.inc4c(h)
h = self.inc4d(h)
if not predict:
b = F.average_pooling_2d(h, 5, stride=3)
b = F.relu(self.normb(self.convb(b), test=test))
b = F.relu(self.normb2(self.linb(b), test=test))
b = self.outb(b)
if regression:
#b = self.toLog(b)
self.loss2 = F.mean_squared_error(b, t)
else:
self.loss2 = F.softmax_cross_entropy(b, t)
h = self.inc4e(h)
h = self.inc5a(h)
h = F.average_pooling_2d(self.inc5b(h), 7)
h = self.out(h)
if predict:
#t.data = cuda.cupy.asarray(t.data, dtype=cuda.cupy.float32).reshape((20,1))
#myloss = F.mean_squared_error(h, t)
return h
if regression:
h = self.toLog(h)
self.loss3 = F.mean_squared_error(h, t)
else:
self.loss3 = F.softmax_cross_entropy(h, t)
if train or regression:
h = np.array(cuda.to_cpu(h.data)).reshape((len(h)))
t = np.array(cuda.to_cpu(t.data)).reshape((len(t)))
#print(h)
#print(t)
return 0.3 * (self.loss1 + self.loss2) + self.loss3, np.corrcoef(
h, t)
else:
return F.accuracy(h, t)
def forward(self):
x = chainer.Variable(self.x)
t = chainer.Variable(self.t)
return functions.softmax_cross_entropy(x,
t,
enable_double_backprop=False)
def __call__(self, x, t):
loss = F.softmax_cross_entropy(self.predict(x), t)
chainer.report({'loss': loss / t.shape[0]}, self)
return loss
def main():
# load MNIST images
images, labels = dataset.load_train_images()
# config
config = adgm.config
# settings
max_epoch = 1000
num_trains_per_epoch = 500
batchsize_l = 100
batchsize_u = 100
alpha = 1
# seed
np.random.seed(args.seed)
if args.gpu_device != -1:
cuda.cupy.random.seed(args.seed)
# save validation accuracy per epoch
csv_results = []
# create semi-supervised split
num_validation_data = 10000
num_labeled_data = 100
num_types_of_label = 10
training_images_l, training_labels_l, training_images_u, validation_images, validation_labels = dataset.create_semisupervised(
images,
labels,
num_validation_data,
num_labeled_data,
num_types_of_label,
seed=args.seed)
print training_labels_l
# init weightnorm layers
if config.use_weightnorm:
print "initializing weight normalization layers ..."
images_l, label_onehot_l, label_id_l = dataset.sample_labeled_data(
training_images_l, training_labels_l, batchsize_l, config.ndim_x,
config.ndim_y)
images_u = dataset.sample_unlabeled_data(training_images_u,
batchsize_u, config.ndim_x)
adgm.compute_lower_bound(images_l, label_onehot_l, images_u)
# training
progress = Progress()
for epoch in xrange(1, max_epoch):
progress.start_epoch(epoch, max_epoch)
sum_lower_bound_l = 0
sum_lower_bound_u = 0
sum_loss_classifier = 0
for t in xrange(num_trains_per_epoch):
# sample from data distribution
images_l, label_onehot_l, label_ids_l = dataset.sample_labeled_data(
training_images_l, training_labels_l, batchsize_l,
config.ndim_x, config.ndim_y)
images_u = dataset.sample_unlabeled_data(training_images_u,
batchsize_u,
config.ndim_x)
# lower bound loss
lower_bound, lb_labeled, lb_unlabeled = adgm.compute_lower_bound(
images_l, label_onehot_l, images_u)
loss_lower_bound = -lower_bound
# classification loss
a_l = adgm.encode_x_a(images_l, False)
unnormalized_y_distribution = adgm.encode_ax_y_distribution(
a_l, images_l, softmax=False)
loss_classifier = alpha * F.softmax_cross_entropy(
unnormalized_y_distribution, adgm.to_variable(label_ids_l))
# backprop
adgm.backprop(loss_classifier + loss_lower_bound)
sum_lower_bound_l += float(lb_labeled.data)
sum_lower_bound_u += float(lb_unlabeled.data)
sum_loss_classifier += float(loss_classifier.data)
progress.show(t, num_trains_per_epoch, {})
adgm.save(args.model_dir)
# validation
images_l, _, label_ids_l = dataset.sample_labeled_data(
validation_images, validation_labels, num_validation_data,
config.ndim_x, config.ndim_y)
images_l_segments = np.split(images_l, num_validation_data // 500)
label_ids_l_segments = np.split(label_ids_l,
num_validation_data // 500)
sum_accuracy = 0
for images_l, label_ids_l in zip(images_l_segments,
label_ids_l_segments):
y_distribution = adgm.encode_x_y_distribution(images_l,
softmax=True,
test=True)
accuracy = F.accuracy(y_distribution,
adgm.to_variable(label_ids_l))
sum_accuracy += float(accuracy.data)
validation_accuracy = sum_accuracy / len(images_l_segments)
progress.show(
num_trains_per_epoch, num_trains_per_epoch, {
"lb_u": sum_lower_bound_u / num_trains_per_epoch,
"lb_l": sum_lower_bound_l / num_trains_per_epoch,
"loss_spv": sum_loss_classifier / num_trains_per_epoch,
"accuracy": validation_accuracy,
})
# write accuracy to csv
csv_results.append([epoch, validation_accuracy])
data = pd.DataFrame(csv_results)
data.columns = ["epoch", "accuracy"]
data.to_csv("{}/result.csv".format(args.model_dir))
def __call__(self, x, t):
y = self.predictor(x)
loss = F.softmax_cross_entropy(y, t)
accuracy = F.accuracy(y, t)
report({'loss': loss, 'accuracy': accuracy}, self)
return loss
def __call__(self, x, t, train=True):
y = self.predictor(x, train)
self.loss = F.softmax_cross_entropy(y, t)
self.accuracy = F.accuracy(y, t)
return self.loss
def trainBatch(self, encSents, decSents, args):
"""main training"""
###encoder step
encEmbed = self.getEmbeddings(encSents, args)
hy, cy, ys = self.encNStepLSTM(hx=None, cx=None, xs=encEmbed)
encOut = F.pad_sequence(ys) #[batch, max(sentlen), Dim]
###decoder step
decEmbed = self.getEmbeddings(decSents, args) #embed のリストの状態
decEmbed = F.pad_sequence(decEmbed).transpose([1, 0, 2]) #padding して[sentLen, batch, Dim]に変更
decode_step = len(decEmbed) - 1
decoderOutList = [0] * decode_step
lstmStateList = [0] * decode_step
firstInput = chainer.Variable(xp.zeros(hy[0].shape, dtype=xp.float32)) #decLSTMの最初に入力として初期化 embedじゃない方
for i in range(decode_step):
#decEmbedじゃないdecLSTMへの入力の準備と、decLSTMのstate準備
if i == 0: #デコーダの最初のステップ
self.set_state([cy[0], hy[0]])
anoInput = firstInput
else:
self.set_state(lstmStateList[i - 1])
anoInput = decoderOutList[i - 1]
hOut = self.decLSTM(F.concat([decEmbed[i], anoInput], 1)) #decoder LSTMの出力
lstmStateList[i] = self.get_state()
decoderOutList[i] = self.attention(hOut, encOut, args) #decoder LSTMの出力をアテンションしたもの decoderの出力 decode_step * [batch, Dim]
total_loss = chainer.Variable(xp.zeros((), dtype=xp.float32))
proc = 0
correct = 0
incorrect = 0
###output層
correctLabels = F.pad_sequence(decSents, padding=-1).T.array #TODO 何か嫌だから上手く書きたい ってかこの関数全体何か汚い -1でパディングしたから1足したら0がeosトークンになるんじゃね？
for i in range(decode_step):
oVector = self.decOut(F.dropout(decoderOutList[i], args.dropout_rate))
correctLabel = correctLabels[i + 1]
proc += (xp.count_nonzero(correctLabel + 1)) ###TODO 0を数えてたらunkトークンがなくなるし、1足したら全部1以上になるンゴ
# 必ずminibatchsizeでわる
closs = F.softmax_cross_entropy(
oVector, correctLabel, normalize=False) #normalize=Falseの意味？ paddingしてるからっぽい
# これで正規化なしのloss cf. seq2seq-attn code
#total_loss_val += closs.data * cMBSize
#if train_mode > 0: # 学習データのみ backward する
total_loss += closs
# 実際の正解数を獲得したい
t_correct = 0
t_incorrect = 0
# Devのときは必ず評価，学習データのときはオプションに従って評価
# if train_mode == 0 or args.doEvalAcc > 0:
# 予測した単語のID配列 CuPy
pred_arr = oVector.data.argmax(axis=1)
# 正解と予測が同じなら0になるはず
# => 正解したところは0なので，全体から引く ###xp.count_nonzero()は間違えた数？
t_correct = (correctLabel.size -
xp.count_nonzero(correctLabel - pred_arr)) #t_correct正解した数
# 予測不要の数から正解した数を引く # +1はbroadcast
t_incorrect = xp.count_nonzero(correctLabel + 1) - t_correct #xp.count_nonzero()は予測する必要のある数 つまりt_incorrectは間違えた数
correct += t_correct
incorrect += t_incorrect
####
#total_loss.backward()
return total_loss, (correct, incorrect, decode_step, proc)
def __call__(self, imgs, masks, labels, bboxes, scales):
"""Forward FCIS and calculate losses.
Here are notations used.
* :math:`N` is the batch size.
* :math:`R` is the number of bounding boxes per image.
* :math:`H` is the image height.
* :math:`W` is the image width.
Currently, only :math:`N=1` is supported.
Args:
imgs (~chainer.Variable): A variable with a batch of images.
masks (~chainer.Variable): A batch of masks.
Its shape is :math:`(N, R, H, W)`.
labels (~chainer.Variable): A batch of labels.
Its shape is :math:`(N, R)`. The background is excluded from
the definition, which means that the range of the value
is :math:`[0, L - 1]`. :math:`L` is the number of foreground
classes.
bboxes (~chainer.Variable): A batch of bounding boxes.
Its shape is :math:`(N, R, 4)`.
scales (float or ~chainer.Variable): Amount of scaling applied to
the raw image during preprocessing.
Returns:
chainer.Variable:
Scalar loss variable.
This is the sum of losses for Region Proposal Network and
the head module.
"""
if isinstance(masks, chainer.Variable):
masks = masks.array
if isinstance(labels, chainer.Variable):
labels = labels.array
if isinstance(bboxes, chainer.Variable):
bboxes = bboxes.array
if isinstance(scales, chainer.Variable):
scales = scales.array
scales = cuda.to_cpu(scales)
batch_size, _, H, W = imgs.shape
img_size = (H, W)
assert img_size == masks.shape[2:]
if any(len(b) == 0 for b in bboxes):
return chainer.Variable(self.xp.array(0, dtype=np.float32))
rpn_features, roi_features = self.fcis.extractor(imgs)
rpn_locs, rpn_scores, rois, roi_indices, anchor = self.fcis.rpn(
rpn_features, img_size, scales)
rpn_locs = F.concat(rpn_locs, axis=0)
rpn_scores = F.concat(rpn_scores, axis=0)
gt_rpn_locs = []
gt_rpn_labels = []
for bbox in bboxes:
gt_rpn_loc, gt_rpn_label = self.anchor_target_creator(
bbox, anchor, img_size)
if cuda.get_array_module(rpn_locs.array) != np:
gt_rpn_loc = cuda.to_gpu(gt_rpn_loc)
gt_rpn_label = cuda.to_gpu(gt_rpn_label)
gt_rpn_locs.append(gt_rpn_loc)
gt_rpn_labels.append(gt_rpn_label)
del gt_rpn_loc, gt_rpn_label
gt_rpn_locs = self.xp.concatenate(gt_rpn_locs, axis=0)
gt_rpn_labels = self.xp.concatenate(gt_rpn_labels, axis=0)
batch_indices = range(batch_size)
sample_rois = []
sample_roi_indices = []
gt_roi_masks = []
gt_roi_labels = []
gt_roi_locs = []
for batch_index, mask, label, bbox in \
zip(batch_indices, masks, labels, bboxes):
roi = rois[roi_indices == batch_index]
sample_roi, gt_roi_mask, gt_roi_label, gt_roi_loc = \
self.proposal_target_creator(
roi, mask, label, bbox, self.loc_normalize_mean,
self.loc_normalize_std, self.mask_size)
del roi
sample_roi_index = self.xp.full((len(sample_roi), ),
batch_index,
dtype=np.int32)
sample_rois.append(sample_roi)
sample_roi_indices.append(sample_roi_index)
del sample_roi, sample_roi_index
gt_roi_masks.append(gt_roi_mask)
gt_roi_labels.append(gt_roi_label)
gt_roi_locs.append(gt_roi_loc)
del gt_roi_mask, gt_roi_label, gt_roi_loc
sample_rois = self.xp.concatenate(sample_rois, axis=0)
sample_roi_indices = self.xp.concatenate(sample_roi_indices, axis=0)
gt_roi_masks = self.xp.concatenate(gt_roi_masks, axis=0)
gt_roi_labels = self.xp.concatenate(gt_roi_labels, axis=0)
gt_roi_locs = self.xp.concatenate(gt_roi_locs, axis=0)
roi_ag_seg_scores, roi_ag_locs, roi_cls_scores, _, _ = self.fcis.head(
roi_features, sample_rois, sample_roi_indices, img_size,
gt_roi_labels)
# RPN losses
rpn_loc_loss = _fast_rcnn_loc_loss(rpn_locs, gt_rpn_locs,
gt_rpn_labels, self.rpn_sigma)
rpn_cls_loss = F.softmax_cross_entropy(rpn_scores, gt_rpn_labels)
if self.n_ohem_sample is None:
n_roi = roi_ag_locs.shape[0]
gt_roi_fg_labels = (gt_roi_labels > 0).astype(np.int)
roi_locs = roi_ag_locs[self.xp.arange(n_roi), gt_roi_fg_labels]
roi_loc_loss = _fast_rcnn_loc_loss(roi_locs, gt_roi_locs,
gt_roi_labels, self.roi_sigma)
roi_cls_loss = F.softmax_cross_entropy(roi_cls_scores,
gt_roi_labels)
roi_mask_loss = F.softmax_cross_entropy(
roi_ag_seg_scores, gt_roi_masks, normalize=False) \
* 10.0 / self.mask_size / self.mask_size
else:
# Losses for outputs of the head
roi_loc_loss, roi_cls_loss, roi_mask_loss = _ohem_loss(
roi_ag_locs, roi_cls_scores, roi_ag_seg_scores, gt_roi_locs,
gt_roi_labels, gt_roi_masks, self.n_ohem_sample,
self.roi_sigma, self.mask_size)
loss = rpn_loc_loss + rpn_cls_loss \
+ roi_loc_loss + roi_cls_loss + roi_mask_loss
chainer.reporter.report(
{
'rpn_loc_loss': rpn_loc_loss,
'rpn_cls_loss': rpn_cls_loss,
'roi_loc_loss': roi_loc_loss,
'roi_cls_loss': roi_cls_loss,
'roi_mask_loss': roi_mask_loss,
'loss': loss,
}, self)
return loss
def forward(self):
x = chainer.Variable(self.x)
t = chainer.Variable(self.t)
return functions.softmax_cross_entropy(x, t, self.use_cudnn)
def __call__(self, x, y):
return F.softmax_cross_entropy(self.fwd(x), y)
#以下正確なaccuracyを求めるため
test_loss = 0
test_accuracy = 0
miss_train_copy = miss_train_total.copy()
miss_test_copy = miss_test_total.copy()
correct_train_copy = correct_train_total.copy()
correct_test_copy = correct_test_total.copy()
for i in range(0, N_test, batch_size_predict):
x = Variable(cuda.to_gpu(x_test[i:i +
batch_size_predict])) #評価画像データ
t = Variable(cuda.to_gpu(t_test[i:i + batch_size_predict])) #評価ラベル
y = model.predict(x)
model.zerograds()
loss = F.softmax_cross_entropy(y, t)
#こっから誤判別を特定するプログラム
acc, data, data2, index, pred, corre, correct_index = accuracy(
y, t)
acc.to_cpu()
data.to_cpu()
data2.to_cpu()
index.to_cpu()
pred.to_cpu()
corre.to_cpu()
correct_index.to_cpu()
t.to_cpu()
x.to_cpu()
for i in range(len(data.data)):
miss_test_total[data.data[i]] += 1
def forward(self, word_list, gold_op_list, unary_limit):
is_training = gold_op_list is not None
# check args
if len(word_list) < 1:
raise ValueError('Word list is empty.')
if is_training:
n_shift = 0
n_binary = 0
for op, _ in gold_op_list:
if op == OP_SHIFT: n_shift += 1
if op == OP_BINARY: n_binary += 1
if n_shift != len(word_list) or n_binary != len(word_list) - 1:
raise ValueError(
'Invalid operation number: SHIFT=%d (required: %d), BINARY=%d (required: %d)'
% (n_shift, n_binary, len(word_list), len(word_list) - 1))
if gold_op_list[-1] != (OP_FINISH, None):
raise ValueError('Last operation is not OP_FINISH.')
# initial values
QUEUE_ZEROS = XP.fzeros((1, self.n_queue))
STACK_ZEROS = XP.fzeros((1, self.n_stack))
SRSTATE_ZEROS = XP.fzeros((1, self.n_srstate))
NEG_INF = -1e20
# queue encoding
q_list = []
qc = QUEUE_ZEROS
q = QUEUE_ZEROS
for text, wid in reversed(word_list):
qc, q = self.net_encoder(qc, XP.iarray([wid]), q)
q_list.insert(0, (text, q))
# estimate
s_list = []
zc = SRSTATE_ZEROS
z = SRSTATE_ZEROS
unary_chain = 0
if is_training:
loss = XP.fzeros(())
for i in itertools.count():
text, q = q_list[0] if q_list else ('', QUEUE_ZEROS)
t1, sc1, s1 = s_list[-1] if s_list else (None, STACK_ZEROS,
STACK_ZEROS)
t2, sc2, s2 = s_list[-2] if len(s_list) >= 2 else (None,
STACK_ZEROS,
STACK_ZEROS)
t3, sc3, s3 = s_list[-3] if len(s_list) >= 3 else (None,
STACK_ZEROS,
STACK_ZEROS)
zc, z = self.net_sr(zc, q, s1, z)
o = self.net_operation(z)
if is_training:
loss += functions.softmax_cross_entropy(
o, XP.iarray([gold_op_list[i][0]]))
o_argmax = gold_op_list[i][0]
else:
o_filter = [0.0 for _ in range(NUM_OP)]
filtered = 0
if not q_list:
o_filter[OP_SHIFT] = NEG_INF
filtered += 1
if not s_list or unary_chain >= unary_limit:
o_filter[OP_UNARY] = NEG_INF
filtered += 1
if len(s_list) < 2:
o_filter[OP_BINARY] = NEG_INF
filtered += 1
if q_list or len(s_list) > 1:
o_filter[OP_FINISH] = NEG_INF
if filtered == NUM_OP:
raise RuntimeError('No possible operation!')
o += XP.farray([o_filter])
o_argmax = int(cuda.to_cpu(o.data.argmax(1)))
if o_argmax == OP_SHIFT:
t0 = Tree(None, [text])
sc0, s0 = (STACK_ZEROS, self.net_shift(q, s1, z))
q_list.pop(0)
unary_chain = 0
label = self.net_semiterminal(s0)
elif o_argmax == OP_UNARY:
t0 = Tree(None, [t1])
sc0, s0 = self.net_unary(sc1, q, s1, s2, z)
s_list.pop()
unary_chain += 1
label = self.net_phrase(s0)
elif o_argmax == OP_BINARY:
t0 = Tree(None, [t2, t1])
sc0, s0 = self.net_binary(sc1, sc2, q, s1, s2, s3, z)
s_list.pop()
s_list.pop()
unary_chain = 0
label = self.net_phrase(s0)
else: # OP_FINISH
break
if is_training:
loss += functions.softmax_cross_entropy(
label, XP.iarray([gold_op_list[i][1]]))
label_argmax = gold_op_list[i][1]
else:
label_argmax = int(cuda.to_cpu(label.data.argmax(1)))
t0.set_label(label_argmax)
s_list.append((t0, sc0, s0))
'''
if is_training:
o_est = int(cuda.to_cpu(o.data.argmax(1)))
label_est = int(cuda.to_cpu(label.data.argmax(1)))
trace('%c %c gold=%d-%2d, est=%d-%2d, stack=%2d, queue=%2d' % (
'*' if o_est == gold_op_list[i][0] else ' ',
'*' if label_est == gold_op_list[i][1] else ' ',
gold_op_list[i][0], gold_op_list[i][1],
o_est, label_est,
len(s_list), len(q_list)))
'''
if is_training:
return loss
else:
# combine multiple trees if they exists, and return the result.
t0, _, __ = s_list.pop()
if s_list:
raise RuntimeError('There exist multiple subtrees!')
return t0
def train(args):
vocab = Vocabulary.from_conll(args.train, args.vocab)
train_dataset = [conll_to_train(x, vocab) for x in read_conll(args.train)]
dev_dataset = [conll_to_train(x, vocab) for x in read_conll(args.dev)]
parser = Parser(args.vocab, args.embed, args.hidden, args.depth)
if args.gpu >= 0:
parser.to_gpu()
opt = optimizers.AdaGrad(lr=0.01)
opt.setup(parser)
opt.add_hook(optimizer.GradientClipping(10))
opt.add_hook(optimizer.WeightDecay(0.0001))
for epoch in range(args.epoch):
random.shuffle(train_dataset)
parser.zerograds()
loss = XP.fzeros(())
for i, data in enumerate(train_dataset):
trace('epoch %3d: train sample %6d:' % (epoch + 1, i + 1))
parent_scores, root_scores = parser.forward(data)
if len(data) > 1:
parent_scores = functions.split_axis(parent_scores, len(data),
0)
else:
parent_scores = (parent_scores, )
root = -1
for j, (p_scores, (wid,
parent)) in enumerate(zip(parent_scores, data)):
if parent == -1:
trace(' %3d: root' % j)
root = j
else:
parent_est = p_scores.data.argmax()
trace('%c %3d -> %3d (%3d)' %
('*' if parent == parent_est else ' ', j, parent_est,
parent))
loss += functions.softmax_cross_entropy(
p_scores, XP.iarray([parent]))
root_est = root_scores.data.argmax()
trace('ROOT: %3d (%3d)' % (root_est, root))
loss += functions.softmax_cross_entropy(root_scores,
XP.iarray([root]))
if (i + 1) % 200 == 0:
loss.backward()
opt.update()
parser.zerograds()
loss = XP.fzeros(())
loss.backward()
opt.update()
trace('epoch %3d: trained. ' % (epoch + 1))
parent_num = 0
parent_match = 0
root_num = 0
root_match = 0
for i, data in enumerate(dev_dataset):
trace('epoch %3d: dev sample %6d:' % (epoch + 1, i + 1),
rollback=True)
parent_scores, root_scores = parser.forward(data)
if len(data) > 1:
parent_scores = functions.split_axis(parent_scores, len(data),
0)
else:
parent_scores = (parent_scores, )
root = -1
for j, (p_scores, (wid,
parent)) in enumerate(zip(parent_scores, data)):
if parent == -1:
root = j
else:
parent_est = p_scores.data.argmax()
parent_num += 1
parent_match += 1 if parent_est == parent else 0
root_est = root_scores.data.argmax()
root_num += 1
root_match += 1 if root_est == root else 0
result_str = \
'epoch %3d: dev: parent-acc = %.4f (%5d/%5d), root-acc = %.4f (%4d/%4d)' % \
( \
epoch + 1, \
parent_match / parent_num, parent_match, parent_num, \
root_match / root_num, root_match, root_num)
trace(result_str)
with open(args.model + '.log', 'a') as fp:
print(result_str, file=fp)
trace('epoch %3d: saving models ...' % (epoch + 1))
prefix = args.model + '.%03d' % (epoch + 1)
vocab.save(prefix + '.vocab')
parser.save_spec(prefix + '.parent_spec')
serializers.save_hdf5(prefix + '.parent_weights', parser)
trace('finished.')
def forward(self, x, t):
return F.softmax_cross_entropy(self.out(x), t)
def main():
# load MNIST images
images, labels = dataset.load_train_images()
# config
config = aae.config
# settings
max_epoch = 1000
num_trains_per_epoch = 5000
batchsize = 100
alpha = 1
# seed
np.random.seed(args.seed)
if args.gpu_device != -1:
cuda.cupy.random.seed(args.seed)
# classification
# 0 -> true sample
# 1 -> generated sample
class_true = aae.to_variable(np.zeros(batchsize, dtype=np.int32))
class_fake = aae.to_variable(np.ones(batchsize, dtype=np.int32))
# training
progress = Progress()
for epoch in xrange(1, max_epoch):
progress.start_epoch(epoch, max_epoch)
sum_loss_reconstruction = 0
sum_loss_discriminator = 0
sum_loss_generator = 0
for t in xrange(num_trains_per_epoch):
# sample from data distribution
images_u = dataset.sample_unlabeled_data(images, batchsize)
# reconstruction phase
qy_x_u, z_u = aae.encode_x_yz(images_u, apply_softmax=True)
reconstruction_u = aae.decode_yz_x(qy_x_u, z_u)
loss_reconstruction = F.mean_squared_error(aae.to_variable(images_u), reconstruction_u)
aae.backprop_generator(loss_reconstruction)
aae.backprop_decoder(loss_reconstruction)
# adversarial phase
y_fake_u, z_fake_u = aae.encode_x_yz(images_u, apply_softmax=True)
z_true_u = sampler.gaussian(batchsize, config.ndim_z, mean=0, var=1)
y_true_u = sampler.onehot_categorical(batchsize, config.ndim_y)
discrimination_z_true = aae.discriminate_z(z_true_u, apply_softmax=False)
discrimination_y_true = aae.discriminate_y(y_true_u, apply_softmax=False)
discrimination_z_fake = aae.discriminate_z(z_fake_u, apply_softmax=False)
discrimination_y_fake = aae.discriminate_y(y_fake_u, apply_softmax=False)
loss_discriminator_z = F.softmax_cross_entropy(discrimination_z_true, class_true) + F.softmax_cross_entropy(discrimination_z_fake, class_fake)
loss_discriminator_y = F.softmax_cross_entropy(discrimination_y_true, class_true) + F.softmax_cross_entropy(discrimination_y_fake, class_fake)
loss_discriminator = loss_discriminator_z + loss_discriminator_y
aae.backprop_discriminator(loss_discriminator)
# adversarial phase
y_fake_u, z_fake_u = aae.encode_x_yz(images_u, apply_softmax=True)
discrimination_z_fake = aae.discriminate_z(z_fake_u, apply_softmax=False)
discrimination_y_fake = aae.discriminate_y(y_fake_u, apply_softmax=False)
loss_generator_z = F.softmax_cross_entropy(discrimination_z_fake, class_true)
loss_generator_y = F.softmax_cross_entropy(discrimination_y_fake, class_true)
loss_generator = loss_generator_z + loss_generator_y
aae.backprop_generator(loss_generator)
sum_loss_reconstruction += float(loss_reconstruction.data)
sum_loss_discriminator += float(loss_discriminator.data)
sum_loss_generator += float(loss_generator.data)
if t % 10 == 0:
progress.show(t, num_trains_per_epoch, {})
aae.save(args.model_dir)
progress.show(num_trains_per_epoch, num_trains_per_epoch, {
"loss_r": sum_loss_reconstruction / num_trains_per_epoch,
"loss_d": sum_loss_discriminator / num_trains_per_epoch,
"loss_g": sum_loss_generator / num_trains_per_epoch,
})
def __call__(self, x, t):
y = self.predictor(x)
loss = F.softmax_cross_entropy(y, t)
return loss
use_dropout = 0.25
label_size = 3
knowledge_size = 2
input_hidden = 3
kelic_hidden = 5
enrich_hidden = 5
mlp_hidden = 3
input_layers = 1
enrich_layer = 1
model = kim(word_embed, emb_dim, label_size, knowledge_size, input_hidden,
kelic_hidden, enrich_hidden, mlp_hidden, input_layers,
enrich_layer, use_dropout)
optimizer = optimizers.AdaGrad()
optimizer.use_cleargrads()
optimizer.setup(model)
optimizer.add_hook(WeightDecay(0.0001))
if args.gpu >= 0:
model.to_gpu()
word_embed.to_gpu()
for i in range(1, args.epoch + 1):
system_start, system_end = model(x_list, y_list, x_mask, y_mask,
knowledge)
loss = F.softmax_cross_entropy(system_start, gold)
print(F.argmax(system_start, axis=1), "loss:", loss.data)
model.cleargrads()
loss.backward()
optimizer.update()
def __call__(self, x, t):
return F.softmax_cross_entropy(self.out(x), t)
def train():
# model
model = Mynet(train=True)
if GPU >= 0:
chainer.cuda.get_device(GPU).use()
model.to_gpu()
opt = chainer.optimizers.MomentumSGD(0.01, momentum=0.9)
opt.setup(model)
#opt.add_hook(chainer.optimizer.WeightDecay(0.0005))
xs, ts, paths = data_load('../Dataset/train/images/', hf=True, vf=True)
# training
mb = 4
mbi = 0
train_ind = np.arange(len(xs))
np.random.seed(0)
np.random.shuffle(train_ind)
for i in range(500):
if mbi + mb > len(xs):
mb_ind = train_ind[mbi:]
np.random.shuffle(train_ind)
mb_ind = np.hstack((mb_ind, train_ind[:(mb - (len(xs) - mbi))]))
mbi = mb - (len(xs) - mbi)
else:
mb_ind = train_ind[mbi:mbi + mb]
mbi += mb
x = xs[mb_ind]
t = ts[mb_ind]
if GPU >= 0:
x = chainer.cuda.to_gpu(x)
t = chainer.cuda.to_gpu(t)
#else:
# x = chainer.Variable(x)
# t = chainer.Variable(t)
y = model(x)
#accu = F.accuracy(y, t[..., 0])
y = F.transpose(y, axes=(0, 2, 3, 1))
y = F.reshape(y, [-1, num_classes + 1])
t = F.reshape(t, [-1])
loss = F.softmax_cross_entropy(y, t)
accu = F.accuracy(y, t)
model.cleargrads()
loss.backward()
opt.update()
loss = loss.data
accu = accu.data
if GPU >= 0:
loss = chainer.cuda.to_cpu(loss)
accu = chainer.cuda.to_cpu(accu)
print("iter >>", i + 1, ',loss >>', loss.item(), ',accuracy >>', accu)
chainer.serializers.save_npz('cnn.npz', model)
dis.zerograds()
styleparam_g = Variable(chainer.cuda.to_gpu(
styleg[batch:batch + batchsize]))
styleparam = Variable(chainer.cuda.to_gpu(
style[batch:batch + batchsize]))
style_vector = gen(styleparam_g)
dis_out1 = dis(style_vector) # にせもの
dis_out2 = dis(styleparam) # ほんもの
# if debug:
# print('params:')
# print(style_vector.data[2][0:10])
# print(styleparam.data[2][0:10])
# print(style_vector.data.shape,styleparam.data.shape)
# da = chainer.cuda.to_gpu(data[1])
loss_gen = F.softmax_cross_entropy(dis_out1, Variable(xp.zeros(batchsize, dtype=np.int32))) # 生成したパラメータがどうなのかadversarial loss, 0本物に近づけたい
loss_dis = F.softmax_cross_entropy(dis_out1, Variable(xp.ones(batchsize, dtype=np.int32))) # 1(偽物)に近づけたい
loss_dis += F.softmax_cross_entropy(dis_out2, Variable(xp.zeros(batchsize, dtype=np.int32))) # 本物をdiscriminatorに入力. 0(本物)に近づけたい
# data[1] = chainer.cuda.to_cpu(data[1])
Lsum_gen += loss_gen
Lsum_dis += loss_dis
if batch < style.shape[0] * 0.9:
loss_gen.backward()
Optimizer_gen.update()
loss_dis.backward()
Optimizer_dis.update()
# else:
# print("val loss: gen:%s dis:%s" % (loss_gen.data, loss_dis.data))
batch += batchsize
# if(batch % 1000 == 0):
# print('batch loss: ' + str(loss.data))
def __call__(self, hs, ys):
'''Decoder forward
:param Variable hs:
:param Variable ys:
:return:
'''
self.loss = None
# prepare input and output word sequences with sos/eos IDs
eos = self.xp.array([self.eos], 'i')
sos = self.xp.array([self.sos], 'i')
ys_in = [F.concat([sos, y], axis=0) for y in ys]
ys_out = [F.concat([y, eos], axis=0) for y in ys]
# padding for ys with -1
# pys: utt x olen
pad_ys_in = F.pad_sequence(ys_in, padding=self.eos)
pad_ys_out = F.pad_sequence(ys_out, padding=-1)
# get dim, length info
batch = pad_ys_out.shape[0]
olength = pad_ys_out.shape[1]
logging.info(self.__class__.__name__ + ' input lengths: ' +
str(self.xp.array([h.shape[0] for h in hs])))
logging.info(self.__class__.__name__ + ' output lengths: ' +
str(self.xp.array([y.shape[0] for y in ys_out])))
# initialization
c_list = [None] # list of cell state of each layer
z_list = [None] # list of hidden state of each layer
for l in six.moves.range(1, self.dlayers):
c_list.append(None)
z_list.append(None)
att_w = None
z_all = []
self.att.reset() # reset pre-computation of h
# pre-computation of embedding
eys = self.embed(pad_ys_in) # utt x olen x zdim
eys = F.separate(eys, axis=1)
# loop for an output sequence
for i in six.moves.range(olength):
att_c, att_w = self.att(hs, z_list[0], att_w)
if i > 0 and random.random() < self.sampling_probability:
logging.info(' scheduled sampling ')
z_out = self.output(z_all[-1])
z_out = F.argmax(F.log_softmax(z_out), axis=1)
z_out = self.embed(z_out)
ey = F.hstack((z_out, att_c)) # utt x (zdim + hdim)
else:
ey = F.hstack((eys[i], att_c)) # utt x (zdim + hdim)
c_list[0], z_list[0] = self.lstm0(c_list[0], z_list[0], ey)
for l in six.moves.range(1, self.dlayers):
c_list[l], z_list[l] = self['lstm%d' % l](c_list[l], z_list[l],
z_list[l - 1])
z_all.append(z_list[-1])
z_all = F.reshape(F.stack(z_all, axis=1),
(batch * olength, self.dunits))
# compute loss
y_all = self.output(z_all)
self.loss = F.softmax_cross_entropy(y_all, F.flatten(pad_ys_out))
# -1: eos, which is removed in the loss computation
self.loss *= (np.mean([len(x) for x in ys_in]) - 1)
acc = F.accuracy(y_all, F.flatten(pad_ys_out), ignore_label=-1)
logging.info('att loss:' + str(self.loss.data))
# show predicted character sequence for debug
if self.verbose > 0 and self.char_list is not None:
y_hat = F.reshape(y_all, (batch, olength, -1))
y_true = pad_ys_out
for (i, y_hat_), y_true_ in zip(enumerate(y_hat.data),
y_true.data):
if i == MAX_DECODER_OUTPUT:
break
idx_hat = self.xp.argmax(y_hat_[y_true_ != -1], axis=1)
idx_true = y_true_[y_true_ != -1]
seq_hat = [self.char_list[int(idx)] for idx in idx_hat]
seq_true = [self.char_list[int(idx)] for idx in idx_true]
seq_hat = "".join(seq_hat).replace('<space>', ' ')
seq_true = "".join(seq_true).replace('<space>', ' ')
logging.info("groundtruth[%d]: " % i + seq_true)
logging.info("prediction [%d]: " % i + seq_hat)
if self.labeldist is not None:
if self.vlabeldist is None:
self.vlabeldist = chainer.Variable(
self.xp.asarray(self.labeldist))
loss_reg = -F.sum(
F.scale(F.log_softmax(y_all), self.vlabeldist,
axis=1)) / len(ys_in)
self.loss = (
1. - self.lsm_weight) * self.loss + self.lsm_weight * loss_reg
return self.loss, acc
def _check_forward(self, mb_locs, mb_confs, gt_mb_locs, gt_mb_labels, k):
if self.variable:
mb_locs = chainer.Variable(mb_locs)
mb_confs = chainer.Variable(mb_confs)
gt_mb_locs = chainer.Variable(gt_mb_locs)
gt_mb_labels = chainer.Variable(gt_mb_labels)
loc_loss, conf_loss = multibox_loss(mb_locs, mb_confs, gt_mb_locs,
gt_mb_labels, k)
self.assertIsInstance(loc_loss, chainer.Variable)
self.assertEqual(loc_loss.shape, ())
self.assertEqual(loc_loss.dtype, mb_locs.dtype)
self.assertIsInstance(conf_loss, chainer.Variable)
self.assertEqual(conf_loss.shape, ())
self.assertEqual(conf_loss.dtype, mb_confs.dtype)
if self.variable:
mb_locs = mb_locs.array
mb_confs = mb_confs.array
gt_mb_locs = gt_mb_locs.array
gt_mb_labels = gt_mb_labels.array
mb_locs = cuda.to_cpu(mb_locs)
mb_confs = cuda.to_cpu(mb_confs)
gt_mb_locs = cuda.to_cpu(gt_mb_locs)
gt_mb_labels = cuda.to_cpu(gt_mb_labels)
loc_loss = cuda.to_cpu(loc_loss.array)
conf_loss = cuda.to_cpu(conf_loss.array)
n_positive_total = 0
expect_loc_loss = 0
expect_conf_loss = 0
for i in six.moves.xrange(gt_mb_labels.shape[0]):
n_positive = 0
negatives = []
for j in six.moves.xrange(gt_mb_labels.shape[1]):
loc = F.huber_loss(mb_locs[np.newaxis, i, j],
gt_mb_locs[np.newaxis, i, j], 1).array
conf = F.softmax_cross_entropy(mb_confs[np.newaxis, i, j],
gt_mb_labels[np.newaxis, i,
j]).array
if gt_mb_labels[i, j] > 0:
n_positive += 1
expect_loc_loss += loc
expect_conf_loss += conf
else:
negatives.append(conf)
n_positive_total += n_positive
if n_positive > 0:
expect_conf_loss += sum(sorted(negatives)[-n_positive * k:])
if n_positive_total == 0:
expect_loc_loss = 0
expect_conf_loss = 0
else:
expect_loc_loss /= n_positive_total
expect_conf_loss /= n_positive_total
np.testing.assert_almost_equal(loc_loss, expect_loc_loss, decimal=2)
np.testing.assert_almost_equal(conf_loss, expect_conf_loss, decimal=2)
# 5. Write a training loop
import numpy as np
from chainer.dataset import concat_examples
from chainer.cuda import to_cpu, to_gpu
max_epoch = 10
gpu_id = 0
model.to_gpu()
while train_iter.epoch < max_epoch:
train_batch = train_iter.next()
image_train, target_train = concat_examples(train_batch, gpu_id)
prediction_train = model(image_train)
loss = F.softmax_cross_entropy(prediction_train, target_train)
model.cleargrads()
loss.backward()
optimizer.update()
if train_iter.is_new_epoch:
print('epoch:{:02d} train_loss:{:.04f} '.format(train_iter.epoch, float(to_cpu(loss.data))), end='')
test_losses = []
test_accuracies = []
while True:
test_batch = test_iter.next()
image_test, target_test = concat_examples(test_batch, gpu_id)
prediction_test = model(image_test)
loss_test = F.softmax_cross_entropy(prediction_test, target_test)
test_losses.append(to_cpu(loss_test.data))
accuracy = F.accuracy(prediction_test, target_test)
def _elementwise_softmax_cross_entropy(x, t):
assert x.shape[:-1] == t.shape
shape = t.shape
x = F.reshape(x, (-1, x.shape[-1]))
t = F.flatten(t)
return F.reshape(F.softmax_cross_entropy(x, t, reduce='no'), shape)
def update_core(self):
# TODO: support n_Classfier <- いる？
# 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')
clfr_optimizer = self.get_optimizer('classfier')
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)
gen_data = gen_data.reshape(batchsize, 1, 28, 28)
# Critic(Discrimintor)
critic_real = self.critic(real_data)
critic_fake = self.critic(gen_data)
# Classifier
# classifier_real = self.classifier(real_data)
# classifier_fake = self.classifier(gen_data)
# Loss
## Critic Loss
# print(critic_fake.shape, critic_real.shape, gen_data.shape, real_data.shape)
# critic_loss = F.mean(critic_fake - critic_real)
e = xp.random.uniform(0., 1.,
(batchsize, 1, 1, 1)).astype(np.float32)
x_hat = e * real_data + (1 - e) * gen_data # recreate Variable
loss_gan = F.average(critic_fake - critic_real)
# x_hat.backward(retain_grad=True, enable_double_backprop=True)
grad, = chainer.grad([self.critic(x_hat)], [x_hat],
enable_double_backprop=True)
grad = F.sqrt(F.batch_l2_norm_squared(grad))
loss_grad = self.l * F.mean_absolute_error(grad,
xp.ones_like(grad.data))
critic_loss = loss_gan + loss_grad
self.critic.cleargrads()
critic_loss.backward()
critic_optimizer.update()
chainer.report({'critic_loss': critic_loss})
chainer.report({'loss_grad': loss_grad})
chainer.report({'loss_gan': loss_gan})
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.
z = Variable(
xp.asarray(
self.generator.make_input_z_with_label(batchsize,
real_label.data)))
# Generator
gen_data = self.generator(z)
# Critic(Discrimintor)
critic_fake = self.critic(gen_data)
# Classifier
classifier_real = self.classifier(real_data)
classifier_fake = self.classifier(gen_data)
## Categorical Loss
c_f_loss = F.softmax_cross_entropy(classifier_fake, real_label)
c_r_loss = F.softmax_cross_entropy(classifier_real, real_label)
c_loss = (c_r_loss + c_f_loss) / 2
self.classifier.cleargrads()
c_loss.backward()
clfr_optimizer.update()
chainer.report({'c_r_loss': c_r_loss})
chainer.report({'c_loss': c_loss})
# Generator Loss
gen_loss = F.average(-critic_fake)
self.generator.cleargrads()
gen_loss.backward()
gen_optimizer.update()
chainer.report({'gen_loss': gen_loss})
self.classifier.cleargrads()
c_f_loss.backward()
gen_optimizer.update()
chainer.report({'c_f_loss': c_f_loss})
def train_dcgan_labeled(evol, dis, proj, epoch0=0):
global epoch
o_evol = optimizers.Adam(alpha=0.0002, beta1=0.5)
o_evol.setup(evol)
o_dis = optimizers.Adam(alpha=0.0002, beta1=0.5)
o_dis.setup(dis)
o_proj = optimizers.Adam(alpha=0.0002, beta1=0.5)
o_proj.setup(proj)
if not args.fresh_start:
serializers.load_hdf5("%s/dcgan_model_evol.h5" % (out_model_dir), evol)
serializers.load_hdf5("%s/dcgan_state_evol.h5" % (out_model_dir),
o_evol)
serializers.load_hdf5("%s/dcgan_model_dis.h5" % (out_model_dir), dis)
serializers.load_hdf5("%s/dcgan_state_dis.h5" % (out_model_dir), o_dis)
serializers.load_hdf5("%s/dcgan_model_proj.h5" % (out_model_dir), proj)
serializers.load_hdf5("%s/dcgan_state_proj.h5" % (out_model_dir),
o_proj)
o_evol.add_hook(chainer.optimizer.WeightDecay(0.00001))
o_dis.add_hook(chainer.optimizer.WeightDecay(0.00001))
o_proj.add_hook(chainer.optimizer.WeightDecay(0.00001))
vis_process = None
for epoch in xrange(epoch0, n_epoch):
for train_ctr in xrange(0, n_train, batchsize):
print "epoch:", epoch, "train:", train_ctr,
# discriminator
# 0: from dataset
# 1: from noise
good_movie = True
prediction_movie = n_movie * [None]
try:
current_movie = load_movie()
except:
continue
for i in range(n_timeseries - 1):
if current_movie[i] is None:
good_movie = False
else:
prediction_movie[i] = current_movie[i]
if not good_movie: continue
for i in range(n_timeseries - 1, n_movie):
prediction_movie[i] = evolve_image(
evol, proj, prediction_movie[i - n_timeseries + 1:i])
if train_ctr % save_interval == 0:
for answer_mode in ['predict', 'observe']:
for offset in [n_timeseries, 16, 32, 64, 119]:
if offset >= n_movie: continue
img_prediction = prediction_movie[offset]
if answer_mode == 'observe':
img_prediction = current_movie[offset]
if img_prediction is None: continue
imgfn = '%s/futuresun_%d_%04d_%s+%03d.png' % (
out_image_dir, epoch, train_ctr, answer_mode,
offset)
plt.rcParams['figure.figsize'] = (12.0, 12.0)
plt.close('all')
plt.imshow(img_prediction, vmin=0, vmax=1.4)
plt.suptitle(imgfn)
plt.savefig(imgfn)
subprocess.call("cp %s ~/public_html/futuresun/" %
(imgfn),
shell=True)
# we don't have enough disk for history
history_dir = 'history/' #%d-%d'%(epoch, train_ctr)
subprocess.call("mkdir -p %s " % (history_dir), shell=True)
subprocess.call("cp %s/*.h5 %s " %
(out_model_dir, history_dir),
shell=True)
if epoch > 0 or train_ctr > 0:
print 'saving model...'
serializers.save_hdf5(
"%s/dcgan_model_evol.h5" % (out_model_dir), evol)
serializers.save_hdf5(
"%s/dcgan_state_evol.h5" % (out_model_dir), o_evol)
serializers.save_hdf5(
"%s/dcgan_model_dis.h5" % (out_model_dir), dis)
serializers.save_hdf5(
"%s/dcgan_state_dis.h5" % (out_model_dir), o_dis)
serializers.save_hdf5(
"%s/dcgan_model_proj.h5" % (out_model_dir), proj)
serializers.save_hdf5(
"%s/dcgan_state_proj.h5" % (out_model_dir), o_proj)
print '...saved.'
movie_in = None
movie_out = None
movie_out_predict = None
evol_scores = {}
proj_scores = {}
matsuoka_shuzo = {}
shuzo_evoke_timestep = []
difficulties = ['normal', 'hard']
vis_kit = {}
for difficulty in difficulties:
evol_scores[difficulty] = [0.0]
proj_scores[difficulty] = [0.0]
matsuoka_shuzo[difficulty] = True
vis_kit[difficulty] = None
matsuoka_shuzo[
'normal'] = False # dameda, dameda.... Akirameyou....
if vis_process is not None:
vis_process.join()
vis_process = None
# start main training routine.
print
next_shuzo_scale = 10.0 * (1 + epoch)
next_shuzo_offset = 1 + abs(
int(round(np.random.normal(scale=next_shuzo_scale))))
for train_offset in range(0, n_movie - n_timeseries):
for difficulty in difficulties:
movie_clip = current_movie
if not matsuoka_shuzo[difficulty]:
# Doushitesokode yamerunda...
continue
else:
# Akiramen'nayo!
pass
if difficulty == 'normal':
movie_clip_in = movie_clip
else:
movie_clip_in = prediction_movie
maybe_dat = create_batch(train_offset, movie_clip_in,
movie_clip)
if not maybe_dat:
#print "Warning: skip offset", train_offset, "because of unavailable data."
continue
data_in, data_out, data_other = maybe_dat
movie_in = Variable(cuda.to_gpu(data_in))
movie_out = Variable(cuda.to_gpu(data_out))
movie_other = Variable(cuda.to_gpu(data_other))
movie_out_predict_before = evol(movie_in)
movie_out_predict = proj(
movie_out_predict_before) # no proj
vis_kit[difficulty] = (movie_in.data.get(),
movie_out.data.get(),
movie_out_predict_before.data.get(),
movie_out_predict.data.get())
if args.norm == 'dcgan':
yl = dis(movie_in, movie_out_predict)
L_evol = F.softmax_cross_entropy(
yl, Variable(xp.zeros(batchsize, dtype=np.int32)))
L_dis = F.softmax_cross_entropy(
yl, Variable(xp.ones(batchsize, dtype=np.int32)))
# train discriminator
yl_train = dis(movie_in, movie_out)
L_dis += F.softmax_cross_entropy(
yl_train,
Variable(xp.zeros(batchsize, dtype=np.int32)))
elif args.norm == 'CA':
L_evol = d_norm(0, dis, movie_out,
movie_out_predict_before)
L_proj = d_norm(0, dis, movie_out, movie_out_predict)
L_dis = d_norm(1, dis, movie_out,
movie_out_predict_before)
# L_dis += d_norm(1, dis, movie_out, movie_out_predict)
L_dis += d_norm(0, dis, movie_out, movie_other)
# L_dis += d_norm(0, dis, movie_other, movie_out)
else:
L2norm = (movie_out - movie_out_predict)**2
yl = F.sum(L2norm) / L2norm.data.size
L_evol = yl
evol_scores[difficulty] += [
L_evol.data.get()
] # np.average(F.softmax(yl).data.get()[:,0])
proj_scores[difficulty] += [
L_proj.data.get()
] # np.average(F.softmax(yl).data.get()[:,0])
# stop learning on normal mode.
if difficulty == 'hard':
o_evol.zero_grads()
L_evol.backward()
o_evol.update()
o_dis.zero_grads()
L_dis.backward()
o_dis.update()
o_proj.zero_grads()
L_proj.backward()
o_proj.update()
movie_in.unchain_backward()
movie_out_predict.unchain_backward()
movie_out_predict_before.unchain_backward()
movie_other.unchain_backward()
L_evol.unchain_backward()
if args.norm == 'dcgan' or args.norm == 'CA':
L_dis.unchain_backward()
sys.stdout.write(
'%d %6s %s: %f -> %f, %f -> %f shuzo:%s\r' %
(train_offset, difficulty, args.norm,
np.average(evol_scores['normal']),
np.average(proj_scores['normal']),
np.average(evol_scores['hard']),
np.average(proj_scores['hard']),
str(shuzo_evoke_timestep[-10:])))
sys.stdout.flush()
# update the prediction as results of learning.
prediction_movie[
train_offset + n_timeseries - 1] = evolve_image(
evol, proj,
prediction_movie[train_offset:train_offset +
n_timeseries - 1])
# prevent too much learning from noisy prediction.
# if len(evol_scores['hard'])>=10 and np.average(evol_scores['hard'][-5:-1]) > 5 * np.average(evol_scores['normal']):
if train_offset == next_shuzo_offset:
next_shuzo_offset = train_offset + 1 + abs(
int(round(
np.random.normal(scale=next_shuzo_scale))))
# Zettaini, akiramennna yo!
# matsuoka_shuzo['hard'] = False
shuzo_evoke_timestep += [train_offset]
evol_scores['hard'] = [0.0]
proj_scores['hard'] = [0.0]
for t in range(train_offset,
train_offset + n_timeseries):
if current_movie[t] is not None:
prediction_movie[t] = current_movie[t]
print
def visualize_vis_kit(vis_kit):
print "visualizing...",
sys.stdout.flush()
for difficulty in difficulties:
if vis_kit[difficulty] is None:
continue
movie_data, movie_out_data, movie_pred_data, movie_proj_data = vis_kit[
difficulty]
imgfn = '%s/batch-%s_%d_%04d.png' % (
out_image_dir, difficulty, epoch, train_ctr)
n_col = n_timeseries + 3
plt.rcParams['figure.figsize'] = (1.0 * n_col,
1.0 * batchsize)
plt.close('all')
for ib in range(batchsize):
for j in range(n_timeseries - 1):
plt.subplot(batchsize, n_col, 1 + ib * n_col + j)
if j < 2:
vmin = -1
vmax = 1
else:
vmin = 0
vmax = 1.4
plt.imshow(movie_data[ib, j, :, :],
vmin=vmin,
vmax=vmax)
plt.axis('off')
plt.subplot(batchsize, n_col,
1 + ib * n_col + n_timeseries - 1)
plt.imshow(movie_pred_data[ib, 0, :, :],
vmin=0,
vmax=1.4)
plt.axis('off')
plt.subplot(batchsize, n_col,
1 + ib * n_col + n_timeseries)
plt.imshow(movie_proj_data[ib, 0, :, :],
vmin=0,
vmax=1.4)
plt.axis('off')
plt.subplot(batchsize, n_col,
1 + ib * n_col + n_timeseries + 2)
plt.imshow(movie_out_data[ib, 0, :, :],
vmin=0,
vmax=1.4)
plt.axis('off')
plt.suptitle(imgfn)
plt.savefig(imgfn)
subprocess.call(
"cp %s ~/public_html/suntomorrow-batch-%s-%s.png" %
(imgfn, difficulty, args.gpu),
shell=True)
print "visualized.",
sys.stdout.flush()
vis_process = Process(target=visualize_vis_kit, args=(vis_kit, ))
vis_process.start()
model = MLP(10, 10, 2)
optimizer = optimizers.SGD()
optimizer.setup(model)
train_data_variable = Variable(train_data.astype(np.float32))
train_label_variable = Variable(train_label.astype(np.int32))
loss_log = []
for epoch in range(200):
model.cleargrads()
prod_label = model(train_data_variable)
loss = F.softmax_cross_entropy(prod_label, train_label_variable)
loss.backward()
optimizer.update()
loss_log.append(loss.data)
#print(loss_log)
print(test_data[0:10])
print("-----")
test_data_variable = Variable(test_data.astype(np.float32))
y = model(test_data_variable)
y = F.softmax(y)
print(y.data[0:50])
pred_label = np.argmax(y.data, 1)
print(pred_label[0:50])
def __call__(self, x, t):
h = self.predict(x)
loss = F.softmax_cross_entropy(h, t)
chainer.report({'loss': loss, 'accuracy': F.accuracy(h, t)}, self)
return loss