def compute_loss(imgs, pafs_ys, heatmaps_ys, pafs_t, heatmaps_t, ignore_mask):
heatmap_loss_log = []
paf_loss_log = []
total_loss = 0
paf_masks = ignore_mask[:, None].repeat(pafs_t.shape[1], axis=1)
heatmap_masks = ignore_mask[:, None].repeat(heatmaps_t.shape[1], axis=1)
# compute loss on each stage
for pafs_y, heatmaps_y in zip(pafs_ys, heatmaps_ys):
stage_pafs_t = pafs_t.copy()
stage_heatmaps_t = heatmaps_t.copy()
stage_paf_masks = paf_masks.copy()
stage_heatmap_masks = heatmap_masks.copy()
if pafs_y.shape != stage_pafs_t.shape:
stage_pafs_t = F.resize_images(stage_pafs_t, pafs_y.shape[2:]).data
stage_heatmaps_t = F.resize_images(stage_heatmaps_t, pafs_y.shape[2:]).data
stage_paf_masks = F.resize_images(stage_paf_masks.astype('f'), pafs_y.shape[2:]).data > 0
stage_heatmap_masks = F.resize_images(stage_heatmap_masks.astype('f'), pafs_y.shape[2:]).data > 0
stage_pafs_t[stage_paf_masks == True] = pafs_y.data[stage_paf_masks == True]
stage_heatmaps_t[stage_heatmap_masks == True] = heatmaps_y.data[stage_heatmap_masks == True]
pafs_loss = F.mean_squared_error(pafs_y, stage_pafs_t)
heatmaps_loss = F.mean_squared_error(heatmaps_y, stage_heatmaps_t)
total_loss += pafs_loss + heatmaps_loss
paf_loss_log.append(float(cuda.to_cpu(pafs_loss.data)))
heatmap_loss_log.append(float(cuda.to_cpu(heatmaps_loss.data)))
return total_loss, paf_loss_log, heatmap_loss_log
def __call__(self, x_recon, x, enc_hiddens, dec_hiddens):
"""
Parameters
-----------------
x_recon: Variable to be reconstructed as label
x: Variable to be reconstructed as label
enc_hiddens: list of Variable
dec_hiddens: list of Varialbe
"""
# Lateral Recon Loss
recon_loss = 0
if self.rc and enc_hiddens is not None:
for h0, h1 in zip(enc_hiddens[::-1], dec_hiddens):
d = np.prod(h0.data.shape[1:])
recon_loss += F.mean_squared_error(h0, h1) / d
# Reconstruction Loss
if x_recon is not None:
d = np.prod(x.data.shape[1:])
recon_loss += F.mean_squared_error(x_recon, x) / d
self.loss = recon_loss
return self.loss
def forward(x, p, a, A=None,P=None):
conv1_1, conv2_1, conv3_1, conv4_1,conv5_1, = func(inputs={'data': x}, outputs=['conv1_1', 'conv2_1', 'conv3_1', 'conv4_1', 'conv5_1'])
conv1_1F,conv2_1F, conv3_1F, conv4_1F,conv5_1F, = [ reshape2(x) for x in [conv1_1,conv2_1, conv3_1, conv4_1,conv5_1]]
conv1_1G,conv2_1G, conv3_1G, conv4_1G,conv5_1G, = [ Fu.matmul(x, x, transa=False, transb=True) for x in [conv1_1F,conv2_1F, conv3_1F, conv4_1F,conv5_1F]]
# Because P an A is not change over iteration, it's better to calcurate onece.
if A is None and B is None:
#compute matrix P
conv1_1,conv2_1, conv3_1, conv4_1,conv5_1, = func(inputs={'data': p}, outputs=['conv1_1', 'conv2_1', 'conv3_1', 'conv4_1', 'conv5_1'])
conv1_1P,conv2_1P, conv3_1P, conv4_1P,conv5_1P, = [ reshape2(x) for x in [conv1_1,conv2_1, conv3_1, conv4_1,conv5_1]]
#compute matrix A
conv1_1,conv2_1, conv3_1, conv4_1,conv5_1, = func(inputs={'data': a}, outputs=['conv1_1', 'conv2_1', 'conv3_1', 'conv4_1', 'conv5_1'])
conv1_1A0,conv2_1A0, conv3_1A0, conv4_1A0,conv5_1A0, = [ reshape2(x) for x in [conv1_1,conv2_1, conv3_1, conv4_1,conv5_1]]
conv1_1A,conv2_1A, conv3_1A, conv4_1A,conv5_1A, = [ Fu.matmul(x, x, transa=False, transb=True) for x in [conv1_1A0,conv2_1A0, conv3_1A0, conv4_1A0,conv5_1A0]]
else:
conv1_1P,conv2_1P, conv3_1P, conv4_1P,conv5_1P,=P
conv1_1A,conv2_1A, conv3_1A, conv4_1A,conv5_1A,=A
L_content = Fu.mean_squared_error(conv4_1F,conv4_1P)/2
#caution! the deviding number is hard coding!
#this part is correspnding to equation (4) in the original paper
#to check the current N and M, run the following
#[x.data.shape for x in [conv1_1F,conv2_1F, conv3_1F, conv4_1F,conv5_1F]]
L_style = (Fu.mean_squared_error(conv1_1G,conv1_1A)/(4*64*64*50176*50176)
+ Fu.mean_squared_error(conv2_1G,conv2_1A)/(4*128**128*12544*12544)
+ Fu.mean_squared_error(conv3_1G,conv3_1A)/(4*256*256*3136*3136)
+ Fu.mean_squared_error(conv4_1G,conv4_1A)/(4*512*512*784*784)\
)/4 # this is equal weighting of E_l
loss = a_p_ratio*L_content + L_style
return loss
def test_invalid_dtype2(self):
x0 = chainer.Variable(
numpy.random.uniform(-1, 1, (4, 3)).astype(numpy.float32))
x1 = chainer.Variable(
numpy.random.uniform(-1, 1, (4, 3)).astype(numpy.float16))
with self.assertRaises(type_check.InvalidType):
functions.mean_squared_error(x0, x1)
def calcLoss(self,y,t):
if self.mode == MODE_TRAIN:
self.loss = F.mean_squared_error(y,t)
self.loss.backward() # 誤差逆伝播
self.optimizer.update() # 最適化
else :
self.loss = F.mean_squared_error(y,t)
def __call__(self, x):
# encode
self.z = self.encode(x)
xp = cuda.get_array_module(self.z.data)
volatile = "off" if self.train else "on"
z_t = Variable(xp.ones_like(self.z.data), volatile=volatile)
loss_z = F.mean_squared_error(self.z, z_t)
# decode
self.y = self.decode(self.z)
loss_y = F.mean_squared_error(x, self.y)
self.loss = loss_z + loss_y
return self.loss
def __call__(self, x_recon, x):
bs = x.shape[0]
d = np.prod(x.shape[1:])
if x.shape[1:] == 3:
h_recon = F.average_pooling_2d(x_recon, (2, 2))
h = F.average_pooling_2d(x, (2, 2))
self.loss = F.mean_squared_error(x_recon, x) / d
else:
self.loss = F.mean_squared_error(x_recon, x) / d
return self.loss
def generate_image(img_orig, img_style, width, nw, nh, max_iter, lr, img_gen=None):
batch_size = img_orig.shape[0]
mid_orig = nn.forward(Variable(img_orig, volatile=True))
style_mats = [get_matrix(y) for y in nn.forward(Variable(img_style, volatile=True))]
if img_gen is None:
if args.gpu >= 0:
img_gen_ = xp.random.uniform(-20,20,(3,width,width),dtype=np.float32)
img_gen = xp.random.uniform(-20,20,(batch_size,3,width,width),dtype=np.float32)
img_gen[:,:,:,:] = img_gen_
else:
img_gen_ = np.random.uniform(-20,20,(3,width,width)).astype(np.float32)
img_gen = np.random.uniform(-20,20,(batch_size,3,width,width)).astype(np.float32)
img_gen[:,:,:,:] = img_gen_
x = Variable(img_gen)
xg = xp.zeros_like(x.data)
optimizer = optimizers.Adam(alpha=lr)
optimizer.setup((img_gen,xg))
for i in range(max_iter):
x = Variable(img_gen)
y = nn.forward(x)
optimizer.zero_grads()
L = Variable(xp.zeros((), dtype=np.float32))
for l in range(len(y)):
gogh_matrix = get_matrix(y[l])
L1 = np.float32(args.lam) * np.float32(nn.alpha[l])*F.mean_squared_error(y[l], Variable(mid_orig[l].data))
L2 = np.float32(nn.beta[l])*F.mean_squared_error(gogh_matrix, Variable(style_mats[l].data))/np.float32(len(y))
L += L1+L2
if i%100==0:
print i,l,L1.data,L2.data
L.backward()
xg += x.grad
optimizer.update()
tmp_shape = img_gen.shape
if args.gpu >= 0:
img_gen += Clip().forward(img_gen).reshape(tmp_shape) - img_gen
else:
def clip(x):
return -120 if x<-120 else (136 if x>136 else x)
img_gen += np.vectorize(clip)(img_gen).reshape(tmp_shape) - img_gen
if i%50==0:
for j in range(img_gen.shape[0]):
save_image(img_gen[j], W, nw[j], nh[j], args.out_dir+"_%d/im_%05d.png"%(j,i))
for j in range(img_gen.shape[0]):
save_image(img_gen[j], W, nw[j], nh[j], args.out_dir+"_last/im_%d.png"%(j))
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 DNN(self, x_train, y_train, x_test, y_test, seed):
np.random.seed(seed)
dnn = Deep()
dnn.compute_accuracy = False
if args.gpu >= 0:
dnn.to_gpu()
optimizer = optimizers.Adam()
optimizer.setup(dnn)
end_counter = 0
min_loss = 100
final_epoch = 0
final_pred = xp.empty([x_test.shape[0], 1], dtype=xp.float32)
x_train, y_train = resample(x_train, y_train, n_samples=x_train.shape[0])
for epoch in range(n_epoch):
indexes = np.random.permutation(x_train.shape[0])
for i in range(0, x_train.shape[0], batchsize):
x_train_dnn = Variable(x_train[indexes[i : i + batchsize]])
y_train_dnn = Variable(y_train[indexes[i : i + batchsize]])
dnn.zerograds()
loss = F.mean_squared_error(dnn(x_train_dnn), y_train_dnn)
loss.backward()
optimizer.update()
end_counter += 1
#evaluation
if epoch % evaluation == 0:
y_pred = dnn(Variable(x_test, volatile='on'))
loss = F.mean_squared_error(y_pred, Variable(y_test, volatile='on'))
if min_loss > loss.data:
min_loss = loss.data
print "epoch{}".format(epoch)
print "Current minimum loss is {}".format(min_loss)
serializers.save_npz('network/DNN{}.model'.format(seed), dnn)
final_epoch = epoch
final_pred = y_pred
end_counter = 0
if end_counter > end_counter_max:
f = open("network/final_epoch.txt", "a")
f.write("DNN{}:{}".format(seed, final_epoch) + "\n")
f.close()
break
return final_pred.data, min_loss
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)
y, y1 = self.forward_super(x, train=train)
L = F.mean_squared_error(y, t)
L1 = F.mean_squared_error(y1, t)
if train:
Loss = L + (L1 * 0.5)
else:
Loss = L
return Loss
def loss_unlabeled(x):
lam = [1000, 10, 0.1, 0.1, 0.1, 0.1, 0.1]
y, zs = enc(x, eta=0.3, test=False)
zs2 = dec(F.softmax(y), zs, test=False)
y3, zs3 = enc(x, eta=0.0, test=False)
mus = [enc.bn1.mu, enc.bn2.mu, enc.bn3.mu, enc.bn4.mu, enc.bn5.mu, enc.bn6.mu]
vrs = [enc.bn1.vr, enc.bn2.vr, enc.bn3.vr, enc.bn4.vr, enc.bn5.vr, enc.bn6.vr]
L = 0
for i in range(len(zs2)):
if i==0:
L += lam[i] * F.mean_squared_error(zs2[i], zs3[i])
else:
L += lam[i] * F.mean_squared_error((zs2[i]-mus[i-1])/(vrs[i-1]+1e-10), zs3[i])
return L
def forward(self, state, action, Reward, state_dash, episode_end):
num_of_batch = state.shape[0]
s = Variable(state)
s_dash = Variable(state_dash)
Q = self.Q_func(s) # Get Q-value
# Generate Target Signals
tmp2 = self.Q_func(s_dash)
tmp2 = list(map(np.argmax, tmp2.data.get())) # argmaxQ(s',a)
tmp = self.Q_func_target(s_dash) # Q'(s',*)
tmp = list(tmp.data.get())
# select Q'(s',*) due to argmaxQ(s',a)
res1 = []
for i in range(num_of_batch):
res1.append(tmp[i][tmp2[i]])
#max_Q_dash = np.asanyarray(tmp, dtype=np.float32)
max_Q_dash = np.asanyarray(res1, dtype=np.float32)
target = np.asanyarray(Q.data.get(), dtype=np.float32)
for i in xrange(num_of_batch):
if not episode_end[i][0]:
tmp_ = np.sign(Reward[i]) + self.gamma * max_Q_dash[i]
else:
tmp_ = np.sign(Reward[i])
action_index = self.action_to_index(action[i])
target[i, action_index] = tmp_
# TD-error clipping
td = Variable(cuda.to_gpu(target)) - Q # TD error
td_tmp = td.data + 1000.0 * (abs(td.data) <= 1) # Avoid zero division
td_clip = td * (abs(td.data) <= 1) + td/abs(td_tmp) * (abs(td.data) > 1)
zero_val = Variable(cuda.to_gpu(np.zeros((self.replay_size, self.num_of_actions), dtype=np.float32)))
loss = F.mean_squared_error(td_clip, zero_val)
return loss, Q
def cost(self, x_data):
x = Variable(x_data)
t = Variable(x_data.reshape(x_data.shape[0], x_data.shape[1]*x_data.shape[2]*x_data.shape[3]))
h = F.dropout(x)
h = self.encode(h)
y = self.decode(h)
return F.mean_squared_error(y, t)
def forward(self, state, action, Reward, state_dash, episode_end):
num_of_batch = state.shape[0]
s = Variable(state)
s_dash = Variable(state_dash)
Q = self.model.Q_func(s,train=True) # Get Q-value
# Generate Target Signals
tmp = self.model_target.Q_func(s_dash,train=self.targetFlag) # Q(s',*)
tmp = list(map(np.max, tmp.data.get())) # max_a Q(s',a)
max_Q_dash = np.asanyarray(tmp, dtype=np.float32)
target = np.asanyarray(Q.data.get(), dtype=np.float32)
for i in xrange(num_of_batch):
if not episode_end[i][0]:
tmp_ = np.sign(Reward[i]) + self.gamma * max_Q_dash[i]
else:
tmp_ = np.sign(Reward[i])
#print action
action_index = self.action_to_index(action[i])
target[i, action_index] = tmp_
# TD-error clipping
td = Variable(cuda.to_gpu(target,self.gpu_id)) - Q # TD error
td_tmp = td.data + 1000.0 * (abs(td.data) <= 1) # Avoid zero division
td_clip = td * (abs(td.data) <= 1) + td/abs(td_tmp) * (abs(td.data) > 1)
zero_val = Variable(cuda.to_gpu(np.zeros((self.replay_size, self.num_of_actions), dtype=np.float32),self.gpu_id))
loss = F.mean_squared_error(td_clip, zero_val)
return loss, Q
def get_loss(self, state, action, reward, state_prime, episode_end):
s = Variable(cuda.to_gpu(state))
s_dash = Variable(cuda.to_gpu(state_prime))
q = self.model.q_function(s) # Get Q-value
# Generate Target Signals
tmp = self.model_target.q_function(s_dash) # Q(s',*)
tmp = list(map(np.max, tmp.data)) # max_a Q(s',a)
max_q_prime = np.asanyarray(tmp, dtype=np.float32)
target = np.asanyarray(copy.deepcopy(q.data.get()), dtype=np.float32)
for i in range(self.replay_size):
if episode_end[i][0] is True:
tmp_ = np.sign(reward[i])
else:
# The sign of reward is used as the reward of DQN!
tmp_ = np.sign(reward[i]) + self.gamma * max_q_prime[i]
target[i, action[i]] = tmp_
# TD-error clipping
td = Variable(cuda.to_gpu(target)) - q # TD error
td_tmp = td.data + 1000.0 * (abs(td.data) <= 1) # Avoid zero division
td_clip = td * (abs(td.data) <= 1) + td/abs(td_tmp) * (abs(td.data) > 1)
zero_val = Variable(cuda.to_gpu(np.zeros((self.replay_size, self.n_act), dtype=np.float32)))
loss = F.mean_squared_error(td_clip, zero_val)
return loss, q
def __call__(self, x_data, y_data):
x = chainer.Variable(x_data)
y = chainer.Variable(y_data)
# q(z|x,y)
rh1 = F.relu(self.recog1(x))
rh2 = F.relu(self.recog2(rh1))
recog_mean = self.recog_mean(rh2)
#recog_log_sigma = 0.5 * self.recog_log_sigma(rh2)
recog_log_sigma = self.recog_log_sigma(rh2)
eps = np.random.normal(0, 1, (x.data.shape[0], nz)).astype(np.float32)
eps = chainer.Variable(eps)
# z = mu + sigma + epsilon
z = recog_mean + F.exp(0.5 * recog_log_sigma) * eps
#z = recog_mean + F.exp(recog_log_sigma) * eps
gh1 = F.relu(self.gen1(z))
gh2 = F.relu(self.gen2(gh1))
gen_mean = self.gen_mean(gh2)
output = F.sigmoid(gen_mean)
loss = F.mean_squared_error(output, y)
kld = -0.5 * F.sum(1 + recog_log_sigma - recog_mean**2 - F.exp(recog_log_sigma)) / (x_data.shape[0] * x_data.shape[1])
return loss, kld, output
def predict(self, x_data, y_data, state):
x ,t = Variable(x_data,volatile=False),Variable(y_data,volatile=False)
h1_in = self.l1_x(x) + self.l1_h(state['h1'])
c1, h1 = F.lstm(state['c1'], h1_in)
y = self.l6(h1)
state = {'c1': c1, 'h1': h1}
return state,F.mean_squared_error(y,t)
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)
y = self.forward_super(x, train=train)
return F.mean_squared_error(y, t)
def calc_loss(self, x, t,layer,train=True): self.clear() #print(x.data.shape) h = F.relu(model.conv3_32(x)) h = F.relu(model.conv32_3(h)) h = F.relu(self.norm_ch3(h, test=not train)) h = F.relu(model.conv3_32(h)) h = F.relu(model.conv32_3(h)) h = F.relu(self.norm_ch3(h, test=not train)) if layer > 0: h = F.relu(model.conv3_32(h)) h = F.relu(model.conv32_3(h)) h = F.relu(self.norm_ch3(h, test=not train)) if layer > 1: h = F.relu(model.conv3_32(h)) h = F.relu(model.conv32_3(h)) h = F.relu(self.norm_ch3(h, test=not train)) if layer > 2: h = F.relu(model.conv3_32(h)) h = F.relu(model.conv32_3(h)) loss = F.mean_squared_error(h, t) return loss
def __call__(self, x, t):
h = F.relu(self.bn1_1(self.conv1_1(x), test=not self.train))
h = F.relu(self.bn1_2(self.conv1_2(h), test=not self.train))
h = F.max_pooling_2d(h, 2, stride=2)
h = F.relu(self.bn2_1(self.conv2_1(h), test=not self.train))
h = F.relu(self.bn2_2(self.conv2_2(h), test=not self.train))
h = F.max_pooling_2d(h, 2, stride=2)
h = F.relu(self.bn3_1(self.conv3_1(h), test=not self.train))
h = F.relu(self.bn3_2(self.conv3_2(h), test=not self.train))
h = F.relu(self.bn3_3(self.conv3_3(h), test=not self.train))
h = F.max_pooling_2d(h, 2, stride=2)
h = F.relu(self.bn4_1(self.conv4_1(h), test=not self.train))
h = F.relu(self.bn4_2(self.conv4_2(h), test=not self.train))
h = F.relu(self.bn4_3(self.conv4_3(h), test=not self.train))
h = F.max_pooling_2d(h, 2, stride=2)
h = F.relu(self.bn5_1(self.conv5_1(h), test=not self.train))
h = F.relu(self.bn5_2(self.conv5_2(h), test=not self.train))
h = F.relu(self.bn5_3(self.conv5_3(h), test=not self.train))
h = F.max_pooling_2d(h, 2, stride=2)
h = F.dropout(F.relu(self.fc6(h)), train=self.train, ratio=0.6)
h = F.dropout(F.relu(self.fc7(h)), train=self.train, ratio=0.6)
self.pred = self.fc8(h)
if t is not None:
self.loss = F.mean_squared_error(self.pred, t)
return self.loss
else:
return self.pred
def _fit(self, X):
self.optimizer.zero_grads()
y_var = self._forward(X, train=True)
loss = F.mean_squared_error(y_var, Variable(X.copy()))
loss.backward()
self.optimizer.update()
return loss
def __call__(self, x, t=None):
self.clear()
h1 = F.leaky_relu(self.conv1(x), slope=0.1)
h1 = F.leaky_relu(self.conv2(h1), slope=0.1)
h1 = F.leaky_relu(self.conv3(h1), slope=0.1)
h2 = self.seranet_v1_crbm(x)
# Fusion
h12 = F.concat((h1, h2), axis=1)
lu = F.leaky_relu(self.convlu6(h12), slope=0.1)
lu = F.leaky_relu(self.convlu7(lu), slope=0.1)
lu = F.leaky_relu(self.convlu8(lu), slope=0.1)
ru = F.leaky_relu(self.convru6(h12), slope=0.1)
ru = F.leaky_relu(self.convru7(ru), slope=0.1)
ru = F.leaky_relu(self.convru8(ru), slope=0.1)
ld = F.leaky_relu(self.convld6(h12), slope=0.1)
ld = F.leaky_relu(self.convld7(ld), slope=0.1)
ld = F.leaky_relu(self.convld8(ld), slope=0.1)
rd = F.leaky_relu(self.convrd6(h12), slope=0.1)
rd = F.leaky_relu(self.convrd7(rd), slope=0.1)
rd = F.leaky_relu(self.convrd8(rd), slope=0.1)
# Splice
h = CF.splice(lu, ru, ld, rd)
h = F.leaky_relu(self.conv9(h), slope=0.1)
h = F.leaky_relu(self.conv10(h), slope=0.1)
h = F.leaky_relu(self.conv11(h), slope=0.1)
h = F.clipped_relu(self.conv12(h), z=1.0)
if self.train:
self.loss = F.mean_squared_error(h, t)
return self.loss
else:
return h
def forward(self, x_data, y_data, train=True):
#print y_data
batchsize = len(x_data)
csize = self.channel
x, t = chainer.Variable(x_data,volatile=not train), chainer.Variable(y_data,volatile=not train)
x = F.reshape(x,(batchsize,csize,-1))
h = F.reshape(x,(batchsize,csize,-1,1))
h = self.conv1(h)
h = F.reshape(h,(batchsize,10,-1))
h = F.tanh(h)
h = F.reshape(h,(batchsize,10,-1,1))
h = self.conv2(h)
h = F.reshape(h,(batchsize,10,-1))
h = F.tanh(h)
h = F.reshape(h,(batchsize,10,-1,1))
h = self.conv3(h)
h = F.reshape(h,(batchsize,100,-1))
h = F.tanh(h)
h = F.reshape(h,(batchsize,100,-1,1))
h = self.conv4(h)
h = F.reshape(h,(batchsize,100,-1))
h = F.tanh(h)
h = F.dropout(F.tanh(self.fc5(h)), train=train)
y = self.fc6(h)
return F.mean_squared_error(y, t)
def test(self, x_data):
x_data = self.procInput(x_data)
x = Variable(x_data)
t = Variable(x_data)
h = self.encode(t)
y = self.decode(h)
return self.procOutput(F.mean_squared_error(y, x))
def sample_3():
# create random input and output data
x = Variable(X.copy())
y = Variable(Y.copy())
# create a network
model = TwoLayerNet(INPUT_SIZE, HIDDEN_SIZE, OUTPUT_SIZE)
for t in range(EPOCHS):
# forward
y_pred = model(x)
# compute and print loss
loss = F.mean_squared_error(y_pred, y)
print(loss.data)
# zero the gradients
model.cleargrads()
# backward
loss.backward()
# update weights
model.linear1.W.data -= LEARNING_RATE * model.linear1.W.grad
model.linear2.W.data -= LEARNING_RATE * model.linear2.W.grad
def forward_one_step(self, x_data, y_data, state, train=True,dropout_ratio=0.0):
x ,t = Variable(x_data,volatile=not train),Variable(y_data,volatile=not train)
h1_in = self.l1_x(F.dropout(x, ratio=dropout_ratio, train=train)) + self.l1_h(state['h1'])
c1, h1 = F.lstm(state['c1'], h1_in)
y = self.l6(F.dropout(h1, ratio=dropout_ratio, train=train))
state = {'c1': c1, 'h1': h1}
return state, F.mean_squared_error(y, t)
def calculate_loss(self, q_model, teacher, states, actions, next_states, rewards, dones):
indices = np.random.permutation(len(states)) # random sampling
shuffle = lambda x : np.array(x)[indices]
states = shuffle(states)
actions = shuffle(actions)
next_states = shuffle(next_states)
rewards = shuffle(rewards)
dones = shuffle(dones)
#print("states:{}".format(states))
#print("actions:{}".format(actions))
#print("next_states:{}".format(next_states))
#print("rewards:{}".format(rewards))
#print("dones:{}".format(dones))
v_states = chainer.Variable(states)
v_next_states = chainer.Variable(next_states)
qv = q_model.forward(v_states)
#print("qv:{}".format(qv.data))
max_qv_next = np.max(teacher.forward(v_next_states).data, axis=1)
target = qv.data.copy()
#teacher_qv = np.sign(rewards)
teacher_qv = rewards
for i, action in enumerate(actions):
if dones[i] == False:
teacher_qv[i] += self.gamma * max_qv_next[i]
target[i, action] = teacher_qv[i]
#print("target:{}".format(target))
td = chainer.Variable(target) - qv
#print("td:{}".format(td.data))
zeros = chainer.Variable(np.zeros(td.data.shape, dtype=np.float32))
loss = F.mean_squared_error(td, zeros)
return loss
def _forward(self, batch, test=False):
# TrainingSetのEncodeとDecode
encoded, means, ln_vars = self._encode(batch, test=test)
rec = self._decode(encoded, test=test)
normer = reduce(lambda x, y: x*y, means.data.shape) # データ数
kl_loss = F.gaussian_kl_divergence(means, ln_vars)/normer
#print 'means={}'.format(means.data.shape)
#print 'ln_vars={}'.format(ln_vars.data.shape)
#print 'kl_loss={}, normer={}'.format(kl_loss.data, normer)
# zのサンプル
samp_p = np.random.standard_normal(means.data.shape).astype('float32')
z_p = chainer.Variable(samp_p)
if self.flag_gpu:
z_p.to_gpu()
rec_p = self._decode(z_p)
disc_rec, conv_layer_rec = self.disc(rec, test=test, dropout_ratio=self.dropout_ratio)
disc_batch, conv_layer_batch = self.disc(batch, test=test, dropout_ratio=self.dropout_ratio)
disc_x_p, conv_layer_x_p = self.disc(rec_p, test=test, dropout_ratio=self.dropout_ratio)
dif_l = F.mean_squared_error(conv_layer_rec, conv_layer_batch)
return kl_loss, dif_l, disc_rec, disc_batch, disc_x_p
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.conv1_1(x))
h = F.relu(self.conv1_2(h))
h = F.max_pooling_2d(h, 2, stride=2)
h = F.relu(self.conv2_1(h))
h = F.relu(self.conv2_2(h))
h = F.max_pooling_2d(h, 2, stride=2)
h = F.relu(self.conv3_1(h))
h = F.relu(self.conv3_2(h))
h = F.relu(self.conv3_3(h))
h = F.max_pooling_2d(h, 2, stride=2)
h = F.relu(self.conv4_1(h))
h = F.relu(self.conv4_2(h))
h = F.relu(self.conv4_3(h))
h = F.max_pooling_2d(h, 2, stride=2)
h = F.relu(self.conv5_1(h))
h = F.relu(self.conv5_2(h))
h = F.relu(self.conv5_3(h))
h = F.max_pooling_2d(h, 2, stride=2)
h = F.dropout(F.relu(self.fc6(h)), train=train, ratio=0.6)
h = F.dropout(F.relu(self.fc7(h)), train=train, ratio=0.6)
h = self.fc8(h)
loss = F.mean_squared_error(h, t)
return loss, h
def main(self):
rospy.init_node('q_update_client_dqn')
rospy.Subscriber("/state_observation_flag", Int64,
self.state_observation_flag_callback)
rospy.Subscriber("/joint1_state", Float64, self.joint1_state_callback)
rospy.Subscriber("/joint2_state", Float64, self.joint2_state_callback)
rospy.Subscriber("/joint3_state", Float64, self.joint3_state_callback)
rospy.Subscriber("/joint4_state", Float64, self.joint4_state_callback)
rospy.Subscriber("/joint5_state", Float64, self.joint5_state_callback)
pub_1 = rospy.Publisher("/joint1_pose", Float64, queue_size=1)
pub_2 = rospy.Publisher("/joint2_pose", Float64, queue_size=1)
pub_3 = rospy.Publisher("/joint3_pose", Float64, queue_size=1)
pub_4 = rospy.Publisher("/joint4_pose", Float64, queue_size=1)
pub_5 = rospy.Publisher("/joint5_pose", Float64, queue_size=1)
pub_6 = rospy.Publisher("/action_num", Int64, queue_size=1)
pub_7 = rospy.Publisher("/num_step", Int64, queue_size=1)
pub_8 = rospy.Publisher("/num_episode", Int64, queue_size=1)
loop_rate = rospy.Rate(100)
filename_result = "/home/amsl/ros_catkin_ws/src/arm_q_learning/dqn_results/test_result.txt"
count = 0
count_temp = 0
self.joint1 = self.init_joint1
self.joint2 = self.init_joint2
self.joint3 = self.init_joint3
self.joint4 = self.init_joint4
self.joint5 = self.init_joint5
print "joint1 : ", self.joint1
print "joint2 : ", self.joint2
print "joint3 : ", self.joint3
print "joint4 : ", self.joint4
print "joint5 : ", self.joint5
self.next_joint1 = self.init_joint1
self.next_joint2 = self.init_joint2
self.next_joint3 = self.init_joint3
self.next_joint4 = self.init_joint4
self.next_joint5 = self.init_joint5
print "next joint1 : ", self.next_joint1
print "next joint2 : ", self.next_joint2
print "next joint3 : ", self.next_joint3
print "next joint4 : ", self.next_joint4
print "next joint5 : ", self.next_joint5
step_count = 0
episode_count = 0
episode_now = 0
episode_past = 0
temp_count = 0
loss_list = []
print "Q Learning Start!!"
while not rospy.is_shutdown():
if self.wait_flag:
print "wait 1 seconds!!"
count += 1
if count == 100:
self.wait_flag = False
self.select_action_flag = False
self.q_update_flag = False
self.state_observation_flag = True
self.state_observation_flag1 = True
self.state_observation_flag2 = True
self.state_observation_flag3 = True
self.state_observation_flag4 = True
self.state_observation_flag5 = True
count = 0
if count == 10:
self.action_num = 0
self.joint1 = self.init_next_joint1
self.joint2 = self.init_next_joint2
self.joint3 = self.init_next_joint3
self.joint4 = self.init_next_joint4
self.joint5 = self.init_next_joint5
self.reward = 0.0
pub_1.publish(self.joint1)
pub_2.publish(self.joint2)
pub_3.publish(self.joint3)
pub_4.publish(self.joint4)
pub_5.publish(self.joint5)
pub_6.publish(self.action_num)
else:
if self.select_action_flag:
self.action = self.epsilon_greedy(self.joint1, self.joint2,
self.joint3, self.joint4,
self.joint5)
self.action_num = self.action
print "self.action_num : ", self.action_num
pub_1.publish(self.joint1)
pub_2.publish(self.joint2)
pub_3.publish(self.joint3)
pub_4.publish(self.joint4)
pub_5.publish(self.joint5)
pub_6.publish(self.action_num)
self.select_action_flag = False
if self.state_observation_flag and self.state_observation_flag1 and self.state_observation_flag2 and self.state_observation_flag3 and self.state_observation_flag4 and self.state_observation_flag5:
print "self.joint_state[0] : ", self.joint_state[0]
print "self.joint_state[1] : ", self.joint_state[1]
print "self.joint_state[2] : ", self.joint_state[2]
print "self.joint_state[3] : ", self.joint_state[3]
print "self.joint_state[4] : ", self.joint_state[4]
print "now joint1 : ", self.joint1
print "now joint2 : ", self.joint2
print "now joint3 : ", self.joint3
print "now joint4 : ", self.joint4
print "now joint5 : ", self.joint5
self.next_joint1 = self.joint_state[0]
self.next_joint2 = self.joint_state[1]
self.next_joint3 = self.joint_state[2]
self.next_joint4 = self.joint_state[3]
self.next_joint5 = self.joint_state[4]
print "next joint1 : ", self.next_joint1
print "next joint2 : ", self.next_joint2
print "next joint3 : ", self.next_joint3
print "next joint4 : ", self.next_joint4
print "next joint5 : ", self.next_joint5
self.reward = self.reward_calculation_client(step_count)
print "reward : ", self.reward
# self.select_action_flag = True
self.q_update_flag = True
self.state_observation_flag = False
self.state_observation_flag1 = False
self.state_observation_flag2 = False
self.state_observation_flag3 = False
self.state_observation_flag4 = False
self.state_observation_flag5 = False
if self.q_update_flag:
target_val = self.reward + self.GAMMA * np.max(
self.forward(self.next_joint1, self.next_joint2,
self.next_joint3, self.next_joint4,
self.next_joint5).data)
self.optimizer.zero_grads()
tt = xp.copy(self.q_list.data)
tt[0][self.action] = target_val
target = chainer.Variable(tt)
# loss = 0.5 * (target - self.q_list) ** 2
loss = F.mean_squared_error(target, self.q_list)
# self.ALPHA = float(self.ALPHA)
# loss.grad = xp.array([[self.ALPHA, self.ALPHA, self.ALPHA, self.ALPHA, self.ALPHA, self.ALPHA, self.ALPHA, self.ALPHA, self.ALPHA, self.ALPHA, self.ALPHA, self.ALPHA, self.ALPHA, self.ALPHA, self.ALPHA, self.ALPHA, self.ALPHA, self.ALPHA, self.ALPHA, self.ALPHA, self.ALPHA, self.ALPHA, self.ALPHA, self.ALPHA, self.ALPHA, self.ALPHA, self.ALPHA]], dtype=xp.float32)
loss_list.append(np.max(loss.data))
loss.backward()
self.optimizer.update()
self.select_action_flag = True
self.q_update_flag = False
step_count += 1
print "episode : %d " % episode_count,
print "step : %d " % step_count,
print "now joint1 : %d " % self.joint1,
print "now joint2 : %d " % self.joint2,
print "now joint3 : %d " % self.joint3,
print "now joint4 : %d " % self.joint4,
print "now joint5 : %d " % self.joint5,
print "now action : %d" % self.action,
print "loss : ", np.max(loss.data),
print "reward : %.1f " % self.reward,
print "EPSILON : %.5f " % self.EPSILON
print ""
# print ""
if self.reward >= 1:
print "episode : %d " % episode_count,
print "step : %d " % step_count,
print "now joint1 : %d " % self.joint1,
print "now joint2 : %d " % self.joint2,
print "now joint3 : %d " % self.joint3,
print "now joint4 : %d " % self.joint4,
print "now joint5 : %d " % self.joint5,
print "now action : %d" % self.action,
print "loss average : %.3f " % (sum(loss_list) /
len(loss_list)),
print "reward : %.1f " % self.reward,
print "EPSILON : %.5f " % self.EPSILON,
print "succsess!!"
print ""
temp_result = np.array(([[episode_count, step_count]]),
dtype=np.int32)
if episode_count == 0:
test_result = temp_result
else:
test_result = np.r_[test_result, temp_result]
while 1:
rand_joint1 = randint(0, 1 - 1) + self.init_joint1
rand_joint2 = randint(0, 11 - 1) + self.init_joint2
rand_joint3 = randint(-10, 1) + self.init_joint3
rand_joint4 = randint(0, 1 - 1) + self.init_joint4
rand_joint5 = randint(0, 1 - 1) + self.init_joint5
if rand_joint2 <= -57 and rand_joint3 <= 75:
print "one more!"
else:
self.init_next_joint1 = rand_joint1
self.init_next_joint2 = rand_joint2
self.init_next_joint3 = rand_joint3
self.init_next_joint4 = rand_joint4
self.init_next_joint5 = rand_joint5
break
step_count = 0
episode_count += 1
episode_now = episode_count
self.action_num = 0
self.joint1 = self.init_next_joint1
self.joint2 = self.init_next_joint2
self.joint3 = self.init_next_joint3
self.joint4 = self.init_next_joint4
self.joint5 = self.init_next_joint5
pub_1.publish(self.joint1)
pub_2.publish(self.joint2)
pub_3.publish(self.joint3)
pub_4.publish(self.joint4)
pub_5.publish(self.joint5)
pub_6.publish(self.action_num)
loss_list = []
self.wait_flag = True
else:
if step_count < 300:
self.joint1 = self.next_joint1
self.joint2 = self.next_joint2
self.joint3 = self.next_joint3
self.joint4 = self.next_joint4
self.joint5 = self.next_joint5
episode_past = episode_now
else:
print "episode : %d " % episode_count,
print "step : %d " % step_count,
print "now joint1 : %d " % self.joint1,
print "now joint2 : %d " % self.joint2,
print "now joint3 : %d " % self.joint3,
print "now joint4 : %d " % self.joint4,
print "now joint5 : %d " % self.joint5,
print "now action : %d" % self.action,
print "loss average : %.3f " % (sum(loss_list) /
len(loss_list)),
print "reward : %.1f " % self.reward,
print "EPSILON : %.5f " % self.EPSILON,
print "failuer!!"
print ""
temp_result = np.array(([[episode_count, step_count]]),
dtype=np.int32)
if episode_count == 0:
test_result = temp_result
else:
test_result = np.r_[test_result, temp_result]
while 1:
rand_joint1 = randint(0, 1 - 1) + self.init_joint1
rand_joint2 = randint(0, 11 - 1) + self.init_joint2
rand_joint3 = randint(-10, 1) + self.init_joint3
rand_joint4 = randint(0, 1 - 1) + self.init_joint4
rand_joint5 = randint(0, 1 - 1) + self.init_joint5
if rand_joint2 <= -57 and rand_joint3 <= 75:
print "one more!"
else:
self.init_next_joint1 = rand_joint1
self.init_next_joint2 = rand_joint2
self.init_next_joint3 = rand_joint3
self.init_next_joint4 = rand_joint4
self.init_next_joint5 = rand_joint5
break
step_count = 0
episode_count += 1
episode_now = episode_count
self.action_num = 0
self.joint1 = self.init_next_joint1
self.joint2 = self.init_next_joint2
self.joint3 = self.init_next_joint3
self.joint4 = self.init_next_joint4
self.joint5 = self.init_next_joint5
pub_1.publish(self.joint1)
pub_2.publish(self.joint2)
pub_3.publish(self.joint3)
pub_4.publish(self.joint4)
pub_5.publish(self.joint5)
pub_6.publish(self.action_num)
loss_list = []
self.wait_flag = True
if math.fabs(episode_now - episode_past) > 1e-6:
if self.EPSILON > 0.3000:
self.EPSILON -= 0.0002
self.num_step = step_count
pub_7.publish(self.num_step)
self.num_episode = episode_count
pub_8.publish(self.num_episode)
if episode_count % 50 == 0:
model_filename = "/home/amsl/ros_catkin_ws/src/arm_q_learning/dqn_model/dqn_arm_model_%d.dat" % episode_count
f = open(model_filename, 'w')
pickle.dump(self.model, f)
if episode_count > 5000:
np.savetxt(filename_result,
test_result,
fmt="%d",
delimiter=",")
# f = open('/home/amsl/ros_catkin_ws/src/arm_q_learning/dqn_model/dqn_arm_model.dat', 'w')
# pickle.dump(self.model, f)
print "Finish!!!"
break
loop_rate.sleep()
def loss_enc(self, enc, x_real, x_fake):
loss = F.mean_squared_error(x_real, x_fake)
chainer.report({'loss': loss}, enc)
return loss
def train_dqn(env):
class Q_Network(chainer.Chain):
def __init__(self, input_size, hidden_size, output_size):
super(Q_Network,
self).__init__(fc1=L.Linear(input_size, hidden_size),
fc2=L.Linear(hidden_size, hidden_size),
fc3=L.Linear(hidden_size, output_size))
def __call__(self, x):
h = F.relu(self.fc1(x))
h = F.relu(self.fc2(h))
y = self.fc3(h)
return y
def reset(self):
self.cleargrads()
Q = Q_Network(input_size=env.history_t + 1, hidden_size=100, output_size=3)
Q_ast = copy.deepcopy(Q)
optimizer = chainer.optimizers.Adam()
optimizer.setup(Q)
epoch_num = 50
step_max = len(env.data) - 1
memory_size = 200
batch_size = 20
epsilon = 1.0
epsilon_decrease = 1e-3
epsilon_min = 0.1
start_reduce_epsilon = 200
train_freq = 10
update_q_freq = 20
gamma = 0.97
show_log_freq = 5
memory = []
total_step = 0
total_rewards = []
total_losses = []
start = time.time()
for epoch in range(epoch_num):
pobs = env.reset()
step = 0
done = False
total_reward = 0
total_loss = 0
while not done and step < step_max:
# select act
pact = np.random.randint(3)
if np.random.rand() > epsilon:
pact = Q(np.array(pobs, dtype=np.float32).reshape(1, -1))
pact = np.argmax(pact.data)
# act
obs, reward, done = env.step(pact)
# add memory
memory.append((pobs, pact, reward, obs, done))
if len(memory) > memory_size:
memory.pop(0)
# train or update q
if len(memory) == memory_size:
if total_step % train_freq == 0:
shuffled_memory = np.random.permutation(memory)
memory_idx = range(len(shuffled_memory))
for i in memory_idx[::batch_size]:
batch = np.array(shuffled_memory[i:i + batch_size])
b_pobs = np.array(batch[:, 0].tolist(),
dtype=np.float32).reshape(
batch_size, -1)
b_pact = np.array(batch[:, 1].tolist(), dtype=np.int32)
b_reward = np.array(batch[:, 2].tolist(),
dtype=np.int32)
b_obs = np.array(batch[:, 3].tolist(),
dtype=np.float32).reshape(
batch_size, -1)
b_done = np.array(batch[:, 4].tolist(), dtype=np.bool)
q = Q(b_pobs)
maxq = np.max(Q_ast(b_obs).data, axis=1)
target = copy.deepcopy(q.data)
for j in range(batch_size):
#MODIFY
target[j, b_pact[j]] = b_reward[
j] + gamma * maxq[j] * (not b_done[j])
Q.reset()
loss = F.mean_squared_error(q, target)
total_loss += loss.data
loss.backward()
optimizer.update()
if total_step % update_q_freq == 0:
Q_ast = copy.deepcopy(Q)
# epsilon
if epsilon > epsilon_min and total_step > start_reduce_epsilon:
epsilon -= epsilon_decrease
# next step
total_reward += reward
pobs = obs
step += 1
total_step += 1
total_rewards.append(total_reward)
total_losses.append(total_loss)
if (epoch + 1) % show_log_freq == 0:
log_reward = sum(total_rewards[(
(epoch + 1) - show_log_freq):]) / show_log_freq
log_loss = sum(total_losses[(
(epoch + 1) - show_log_freq):]) / show_log_freq
elapsed_time = time.time() - start
print('\t'.join(
map(str, [
epoch + 1, epsilon, total_step, log_reward, log_loss,
elapsed_time
])))
start = time.time()
return Q, total_losses, total_rewards
serializers.load_npz(DIR_NAME + '/my.model', model)
swingDataList = []
for i in range(prop['SWING_NUM']):
rootDir = '{}/learning/IMAGES/{}/{}/{}/{}' \
.format(ROOT, prop['IMG_DIR'], prop['DATA_TYPE'], 'swing', i)
dataList = mylib.image.createInputSwingDataList(rootDir)
swingDataList.append(np.array(dataList).astype(np.float32))
errors = []
for swingNum in range(prop['SWING_NUM']):
x = Variable(mylib.NN.cupyArray(swingDataList[swingNum], False))
t = Variable(x.data)
y = model(x)
loss = F.mean_squared_error(y, t)
errors.append(loss.data)
mylib.util.mkdir('{}/output/swing'.format(DIR_NAME))
for i in range(len(swingDataList[swingNum])):
mylib.util.mkdir('{}/output/swing/swing{}'.format(DIR_NAME, i))
img = mylib.image.makeOutputData(y.data[i], prop['IMG_HEIGHT'], prop['IMG_WIDTH'])
cv2.imwrite('{}/output/swing/swing{}/img{}.png'.format(DIR_NAME, swingNum, i), img)
# エラーを記録
errors = np.array(errors)
fError = open(DIR_NAME + '/output/swing/error.dat', 'w')
fError.write('# swingNum\t' + 'error\n')
for swingNum in range(prop['SWING_NUM']):
fError.write('\t'.join([str(swingNum), str(errors[swingNum])]) + '\n')
fError.close()
def update_core(self):
opt_enc_x = self.get_optimizer('opt_enc_x')
opt_dec_x = self.get_optimizer('opt_dec_x')
opt_enc_y = self.get_optimizer('opt_enc_y')
opt_dec_y = self.get_optimizer('opt_dec_y')
opt_x = self.get_optimizer('opt_x')
opt_y = self.get_optimizer('opt_y')
opt_z = self.get_optimizer('opt_z')
# get mini-batch
batch_x = self.get_iterator('main').next()
batch_y = self.get_iterator('train_B').next()
x = Variable(self.converter(batch_x, self.args.gpu[0]))
y = Variable(self.converter(batch_y, self.args.gpu[0]))
# encode to latent (X,Y => Z)
x_z = self.enc_x(losses.add_noise(x, sigma=self.args.noise))
y_z = self.enc_y(losses.add_noise(y, sigma=self.args.noise))
loss_gen = 0
## regularisation on the latent space
if self.args.lambda_reg > 0:
loss_reg_enc_y = losses.loss_func_reg(y_z[-1], 'l2')
loss_reg_enc_x = losses.loss_func_reg(x_z[-1], 'l2')
loss_gen = loss_gen + self.args.lambda_reg * (loss_reg_enc_x +
loss_reg_enc_y)
chainer.report({'loss_reg': loss_reg_enc_x}, self.enc_x)
chainer.report({'loss_reg': loss_reg_enc_y}, self.enc_y)
## discriminator for the latent space: distribution of image of enc_x should look same as that of enc_y
# since z is a list (for u-net), we use only the output of the last layer
if self.args.lambda_dis_z > 0:
if self.args.dis_wgan:
loss_enc_x_adv = -F.average(self.dis_z(x_z[-1]))
loss_enc_y_adv = F.average(self.dis_z(y_z[-1]))
else:
loss_enc_x_adv = losses.loss_func_comp(self.dis_z(x_z[-1]),
1.0)
loss_enc_y_adv = losses.loss_func_comp(self.dis_z(y_z[-1]),
0.0)
loss_gen = loss_gen + self.args.lambda_dis_z * (loss_enc_x_adv +
loss_enc_y_adv)
chainer.report({'loss_adv': loss_enc_x_adv}, self.enc_x)
chainer.report({'loss_adv': loss_enc_y_adv}, self.enc_y)
# cycle for X=>Z=>X (Autoencoder)
x_x = self.dec_x(x_z)
loss_cycle_xzx = F.mean_absolute_error(x_x, x)
chainer.report({'loss_cycle': loss_cycle_xzx}, self.enc_x)
# cycle for Y=>Z=>Y (Autoencoder)
y_y = self.dec_y(y_z)
loss_cycle_yzy = F.mean_absolute_error(y_y, y)
chainer.report({'loss_cycle': loss_cycle_yzy}, self.enc_y)
loss_gen = loss_gen + self.args.lambda_Az * loss_cycle_xzx + self.args.lambda_Bz * loss_cycle_yzy
## decode from latent Z => Y,X
x_y = self.dec_y(x_z)
y_x = self.dec_x(y_z)
# cycle for X=>Z=>Y=>Z=>X (Z=>Y=>Z does not work well)
x_y_x = self.dec_x(self.enc_y(x_y))
loss_cycle_x = F.mean_absolute_error(x_y_x, x)
chainer.report({'loss_cycle': loss_cycle_x}, self.dec_x)
# cycle for Y=>Z=>X=>Z=>Y
y_x_y = self.dec_y(self.enc_x(y_x))
loss_cycle_y = F.mean_absolute_error(y_x_y, y)
chainer.report({'loss_cycle': loss_cycle_y}, self.dec_y)
loss_gen = loss_gen + self.args.lambda_A * loss_cycle_x + self.args.lambda_B * loss_cycle_y
## adversarial for Y
if self.args.lambda_dis_y > 0:
x_y_copy = Variable(self._buffer_y.query(x_y.data))
if self.args.dis_wgan:
loss_dec_y_adv = -F.average(self.dis_y(x_y))
else:
loss_dec_y_adv = losses.loss_func_comp(self.dis_y(x_y), 1.0)
loss_gen = loss_gen + self.args.lambda_dis_y * loss_dec_y_adv
chainer.report({'loss_adv': loss_dec_y_adv}, self.dec_y)
## adversarial for X
if self.args.lambda_dis_x > 0:
y_x_copy = Variable(self._buffer_x.query(y_x.data))
if self.args.dis_wgan:
loss_dec_x_adv = -F.average(self.dis_x(y_x))
else:
loss_dec_x_adv = losses.loss_func_comp(self.dis_x(y_x), 1.0)
loss_gen = loss_gen + self.args.lambda_dis_x * loss_dec_x_adv
chainer.report({'loss_adv': loss_dec_x_adv}, self.dec_x)
## idempotence
if self.args.lambda_idempotence > 0:
loss_idem_x = F.mean_absolute_error(y_x,
self.dec_x(self.enc_y(y_x)))
loss_idem_y = F.mean_absolute_error(x_y,
self.dec_y(self.enc_x(x_y)))
loss_gen = loss_gen + self.args.lambda_idempotence * (loss_idem_x +
loss_idem_y)
chainer.report({'loss_idem': loss_idem_x}, self.dec_x)
chainer.report({'loss_idem': loss_idem_y}, self.dec_y)
# Y => X shouldn't change X
if self.args.lambda_domain > 0:
loss_dom_x = F.mean_absolute_error(x, self.dec_x(self.enc_y(x)))
loss_dom_y = F.mean_absolute_error(y, self.dec_y(self.enc_x(y)))
loss_gen = loss_gen + self.args.lambda_domain * (loss_dom_x +
loss_dom_y)
chainer.report({'loss_dom': loss_dom_x}, self.dec_x)
chainer.report({'loss_dom': loss_dom_y}, self.dec_y)
## images before/after conversion should look similar in terms of perceptual loss
if self.args.lambda_identity_x > 0:
loss_id_x = losses.loss_perceptual(
x,
x_y,
self.vgg,
layer=self.args.perceptual_layer,
grey=self.args.grey)
loss_gen = loss_gen + self.args.lambda_identity_x * loss_id_x
chainer.report({'loss_id': 1e-3 * loss_id_x}, self.enc_x)
if self.args.lambda_identity_y > 0:
loss_id_y = losses.loss_perceptual(
y,
y_x,
self.vgg,
layer=self.args.perceptual_layer,
grey=self.args.grey)
loss_gen = loss_gen + self.args.lambda_identity_y * loss_id_y
chainer.report({'loss_id': 1e-3 * loss_id_y}, self.enc_y)
## background (pixels with value -1) should be preserved
if self.args.lambda_air > 0:
loss_air_x = losses.loss_comp_low(x,
x_y,
self.args.air_threshold,
norm='l2')
loss_air_y = losses.loss_comp_low(y,
y_x,
self.args.air_threshold,
norm='l2')
loss_gen = loss_gen + self.args.lambda_air * (loss_air_x +
loss_air_y)
chainer.report({'loss_air': loss_air_x}, self.dec_y)
chainer.report({'loss_air': loss_air_y}, self.dec_x)
## images before/after conversion should look similar in the gradient domain
if self.args.lambda_grad > 0:
loss_grad_x = losses.loss_grad(x, x_y)
loss_grad_y = losses.loss_grad(y, y_x)
loss_gen = loss_gen + self.args.lambda_grad * (loss_grad_x +
loss_grad_y)
chainer.report({'loss_grad': loss_grad_x}, self.dec_y)
chainer.report({'loss_grad': loss_grad_y}, self.dec_x)
## total variation (only for X -> Y)
if self.args.lambda_tv > 0:
loss_tv = losses.total_variation(x_y,
tau=self.args.tv_tau,
method=self.args.tv_method)
if self.args.imgtype == "dcm" and self.args.num_slices > 1:
loss_tv += losses.total_variation_ch(x_y)
loss_gen = loss_gen + self.args.lambda_tv * loss_tv
chainer.report({'loss_tv': loss_tv}, self.dec_y)
## back propagate
self.enc_x.cleargrads()
self.dec_x.cleargrads()
self.enc_y.cleargrads()
self.dec_y.cleargrads()
loss_gen.backward()
opt_enc_x.update(loss=loss_gen)
opt_dec_x.update(loss=loss_gen)
if not self.args.single_encoder:
opt_enc_y.update(loss=loss_gen)
opt_dec_y.update(loss=loss_gen)
##########################################
## discriminator for Y
if self.args.dis_wgan: ## synthesised -, real +
eps = self.xp.random.uniform(0, 1, size=len(batch_y)).astype(
self.xp.float32)[:, None, None, None]
if self.args.lambda_dis_y > 0:
## discriminator for X=>Y
loss_dis_y = F.average(self.dis_y(x_y_copy) - self.dis_y(y))
y_mid = eps * y + (1.0 - eps) * x_y_copy
# gradient penalty
gd_y, = chainer.grad([self.dis_y(y_mid)], [y_mid],
enable_double_backprop=True)
gd_y = F.sqrt(F.batch_l2_norm_squared(gd_y) + 1e-6)
loss_dis_y_gp = F.mean_squared_error(
gd_y, self.xp.ones_like(gd_y.data))
chainer.report({'loss_dis': loss_dis_y}, self.dis_y)
chainer.report(
{'loss_gp': self.args.lambda_wgan_gp * loss_dis_y_gp},
self.dis_y)
loss_dis_y = loss_dis_y + self.args.lambda_wgan_gp * loss_dis_y_gp
self.dis_y.cleargrads()
loss_dis_y.backward()
opt_y.update(loss=loss_dis_y)
if self.args.lambda_dis_x > 0:
## discriminator for B=>A
loss_dis_x = F.average(self.dis_x(y_x_copy) - self.dis_x(x))
x_mid = eps * x + (1.0 - eps) * y_x_copy
# gradient penalty
gd_x, = chainer.grad([self.dis_x(x_mid)], [x_mid],
enable_double_backprop=True)
gd_x = F.sqrt(F.batch_l2_norm_squared(gd_x) + 1e-6)
loss_dis_x_gp = F.mean_squared_error(
gd_x, self.xp.ones_like(gd_x.data))
chainer.report({'loss_dis': loss_dis_x}, self.dis_x)
chainer.report(
{'loss_gp': self.args.lambda_wgan_gp * loss_dis_x_gp},
self.dis_x)
loss_dis_x = loss_dis_x + self.args.lambda_wgan_gp * loss_dis_x_gp
self.dis_x.cleargrads()
loss_dis_x.backward()
opt_x.update(loss=loss_dis_x)
## discriminator for latent: X -> Z is - while Y -> Z is +
if self.args.lambda_dis_z > 0 and t == 0:
loss_dis_z = F.average(
self.dis_z(x_z[-1]) - self.dis_z(y_z[-1]))
z_mid = eps * x_z[-1] + (1.0 - eps) * y_z[-1]
# gradient penalty
gd_z, = chainer.grad([self.dis_z(z_mid)], [z_mid],
enable_double_backprop=True)
gd_z = F.sqrt(F.batch_l2_norm_squared(gd_z) + 1e-6)
loss_dis_z_gp = F.mean_squared_error(
gd_z, self.xp.ones_like(gd_z.data))
chainer.report({'loss_dis': loss_dis_z}, self.dis_z)
chainer.report(
{'loss_gp': self.args.lambda_wgan_gp * loss_dis_y_gp},
self.dis_y)
loss_dis_z = loss_dis_z + self.args.lambda_wgan_gp * loss_dis_z_gp
self.dis_z.cleargrads()
loss_dis_z.backward()
opt_z.update(loss=loss_dis_z)
else: ## LSGAN
if self.args.lambda_dis_y > 0:
## discriminator for A=>B (real:1, fake:0)
disy_fake = self.dis_y(x_y_copy)
loss_dis_y_fake = losses.loss_func_comp(
disy_fake, 0.0, self.args.dis_jitter)
disy_real = self.dis_y(y)
loss_dis_y_real = losses.loss_func_comp(
disy_real, 1.0, self.args.dis_jitter)
if self.args.dis_reg_weighting > 0: ## regularization
loss_dis_y_reg = (
F.average(F.absolute(disy_real[:, 1, :, :])) +
F.average(F.absolute(disy_fake[:, 1, :, :])))
else:
loss_dis_y_reg = 0
chainer.report({'loss_reg': loss_dis_y_reg}, self.dis_y)
loss_dis_y_gp = 0
chainer.report({'loss_fake': loss_dis_y_fake}, self.dis_y)
chainer.report({'loss_real': loss_dis_y_real}, self.dis_y)
loss_dis_y = (
loss_dis_y_fake + loss_dis_y_real
) * 0.5 + self.args.dis_reg_weighting * loss_dis_y_reg + self.args.lambda_wgan_gp * loss_dis_y_gp
self.dis_y.cleargrads()
loss_dis_y.backward()
opt_y.update(loss=loss_dis_y)
if self.args.lambda_dis_x > 0:
## discriminator for B=>A
disx_fake = self.dis_x(y_x_copy)
loss_dis_x_fake = losses.loss_func_comp(
disx_fake, 0.0, self.args.dis_jitter)
disx_real = self.dis_x(x)
loss_dis_x_real = losses.loss_func_comp(
disx_real, 1.0, self.args.dis_jitter)
if self.args.dis_reg_weighting > 0: ## regularization
loss_dis_x_reg = (
F.average(F.absolute(disx_fake[:, 1, :, :])) +
F.average(F.absolute(disx_real[:, 1, :, :])))
else:
loss_dis_x_reg = 0
chainer.report({'loss_reg': loss_dis_x_reg}, self.dis_x)
loss_dis_x_gp = 0
chainer.report({'loss_fake': loss_dis_x_fake}, self.dis_x)
chainer.report({'loss_real': loss_dis_x_real}, self.dis_x)
loss_dis_x = (
loss_dis_x_fake + loss_dis_x_real
) * 0.5 + self.args.dis_reg_weighting * loss_dis_x_reg + self.args.lambda_wgan_gp * loss_dis_x_gp
self.dis_x.cleargrads()
loss_dis_x.backward()
opt_x.update(loss=loss_dis_x)
## discriminator for latent: X -> Z is 0.0 while Y -> Z is 1.0
if self.args.lambda_dis_z > 0:
disz_xz = self.dis_z(x_z[-1])
loss_dis_z_x = losses.loss_func_comp(disz_xz, 0.0,
self.args.dis_jitter)
disz_yz = self.dis_z(y_z[-1])
loss_dis_z_y = losses.loss_func_comp(disz_yz, 1.0,
self.args.dis_jitter)
if self.args.dis_reg_weighting > 0: ## regularization
loss_dis_z_reg = (
F.average(F.absolute(disz_xz[:, 1, :, :])) +
F.average(F.absolute(disz_yz[:, 1, :, :])))
else:
loss_dis_z_reg = 0
chainer.report({'loss_x': loss_dis_z_x}, self.dis_z)
chainer.report({'loss_y': loss_dis_z_y}, self.dis_z)
chainer.report({'loss_reg': loss_dis_z_reg}, self.dis_z)
loss_dis_z = (
loss_dis_z_x + loss_dis_z_y
) * 0.5 + self.args.dis_reg_weighting * loss_dis_z_reg
self.dis_z.cleargrads()
loss_dis_z.backward()
opt_z.update(loss=loss_dis_z)
def evaluate(self):
domain = ['in', 'truth', 'out']
if self.eval_hook:
self.eval_hook(self)
for k, dataset in enumerate(['test', 'train']):
batch = self._iterators[dataset].next()
x_in, t_out = chainer.dataset.concat_examples(batch, self.device)
x_in = Variable(x_in) # original image
t_out = Variable(
t_out) # corresponding translated image (ground truth)
with chainer.using_config(
'train', False), chainer.function.no_backprop_mode():
x_out = self._targets['dec_y'](
self._targets['enc_x'](x_in)) # translated image by NN
## unfold stack and apply softmax
if self.args.class_num > 0 and self.args.stack > 0:
x_in = x_in.reshape(x_in.shape[0] * self.args.stack,
x_in.shape[1] // self.args.stack,
x_in.shape[2], x_in.shape[3])
x_out = F.softmax(
x_out.reshape(x_out.shape[0] * self.args.stack,
x_out.shape[1] // self.args.stack,
x_out.shape[2], x_out.shape[3]))
t_out = t_out.reshape(t_out.shape[0] * self.args.stack,
t_out.shape[1] // self.args.stack,
t_out.shape[2], t_out.shape[3])
#print(x_out.shape, t_out.shape)
# select middle slices
x_in = x_in[(self.args.stack // 2)::self.args.stack]
x_out = x_out[(self.args.stack // 2)::self.args.stack]
t_out = t_out[(self.args.stack // 2)::self.args.stack]
if dataset == 'test': # for test dataset, compute some statistics
fig = plt.figure(figsize=(12, 6 * len(x_out)))
gs = gridspec.GridSpec(2 * len(x_out),
4,
wspace=0.1,
hspace=0.1)
loss_rec_L1 = F.mean_absolute_error(x_out, t_out)
loss_rec_L2 = F.mean_squared_error(x_out, t_out)
loss_rec_CE = softmax_focalloss(x_out,
t_out,
gamma=self.args.focal_gamma,
class_weight=self.class_weight)
result = {
"myval/loss_L1": loss_rec_L1,
"myval/loss_L2": loss_rec_L2,
"myval/loss_CE": loss_rec_CE
}
## iterate over batch
for i, var in enumerate([x_in, t_out, x_out]):
if i % 3 != 0 and self.args.class_num > 0: # t_out, x_out
imgs = var2unit_img(var, 0, 1) # softmax
#imgs[:,:,:,0] = 0 # class 0 => black ######
#imgs = np.roll(imgs,1,axis=3)[:,:,:,:3] ## R0B, show only 3 classes (-1,0,1)
else:
imgs = var2unit_img(var) # tanh
# print(imgs.shape,np.min(imgs),np.max(imgs))
for j in range(len(imgs)):
ax = fig.add_subplot(gs[j + k * len(x_out), i])
ax.set_title(dataset + "_" + domain[i], fontsize=8)
if (imgs[j].shape[2] == 3): ## RGB
ax.imshow(imgs[j],
interpolation='none',
vmin=0,
vmax=1)
elif (imgs[j].shape[2] >= 4): ## categorical
cols = ['k', 'b', 'c', 'g', 'y', 'r', 'm', 'w'] * 5
cmap = colors.ListedColormap(cols)
im = np.argmax(imgs[j], axis=2)
norm = colors.BoundaryNorm(list(range(len(cols) + 1)),
cmap.N)
ax.imshow(im,
interpolation='none',
cmap=cmap,
norm=norm)
else:
ax.imshow(imgs[j][:, :, -1],
interpolation='none',
cmap='gray',
vmin=0,
vmax=1)
ax.set_xticks([])
ax.set_yticks([])
## difference image
if (x_out.shape[1] >= 4): ## categorical
eps = 1e-7
p = F.clip(
x_out, x_min=eps, x_max=1 -
eps) ## we assume the input is already applied softmax
q = -F.clip(t_out, x_min=eps, x_max=1 - eps) * F.log(p)
diff = F.sum(q * ((1 - p)**2), axis=1, keepdims=True)
vmin = -1
vmax = 1
else:
diff = (x_out - t_out)
vmin = -0.1
vmax = 0.1
diff = diff.data.get().transpose(0, 2, 3, 1)
for j in range(len(diff)):
ax = fig.add_subplot(gs[j + k * len(x_out), 3])
ax.imshow(diff[j][:, :, 0],
interpolation='none',
cmap='coolwarm',
vmin=vmin,
vmax=vmax)
ax.set_xticks([])
ax.set_yticks([])
gs.tight_layout(fig)
plt.savefig(os.path.join(self.vis_out,
'count{:0>4}.jpg'.format(self.count)),
dpi=200)
self.count += 1
plt.close()
return result
nbatch]]
t_batch_2 = train_2_label[inds[icount *
nbatch:(icount + 1) *
nbatch]]
x_batch_3 = train_3[inds[icount * nbatch:(icount + 1) *
nbatch]]
t_batch_3 = train_3_label[inds[icount *
nbatch:(icount + 1) *
nbatch]]
# FORWARD PROP.
y_1 = model_1(x_batch_1)
y_2 = model_2(x_batch_2)
y_3 = model_3(x_batch_3)
if LOSS == 'RMSE':
loss_1 = F.mean_squared_error(y_1, t_batch_1)
loss_2 = F.mean_squared_error(y_2, t_batch_2)
loss_3 = F.mean_squared_error(y_3, t_batch_3)
elif LOSS == 'MAE':
loss_1 = F.mean_absolute_error(y_1, t_batch_1)
loss_2 = F.mean_absolute_error(y_2, t_batch_2)
loss_3 = F.mean_absolute_error(y_3, t_batch_3)
# BACK PROP.
model_1.cleargrads()
loss_1.backward()
model_2.cleargrads()
loss_2.backward()
model_3.cleargrads()
def evaluate(self):
batch_x = self._iterators['main'].next()
batch_y = self._iterators['testB'].next()
models = self._targets
if self.eval_hook:
self.eval_hook(self)
fig = plt.figure(figsize=(9, 3 * (len(batch_x) + len(batch_y))))
gs = gridspec.GridSpec(len(batch_x) + len(batch_y),
3,
wspace=0.1,
hspace=0.1)
x = Variable(self.converter(batch_x, self.device))
y = Variable(self.converter(batch_y, self.device))
with chainer.using_config('train', False):
with chainer.function.no_backprop_mode():
if len(models) > 2:
x_y = models['dec_y'](models['enc_x'](x))
if self.single_encoder:
x_y_x = models['dec_x'](models['enc_x'](x_y))
else:
x_y_x = models['dec_x'](
models['enc_x'](x)) ## autoencoder
#x_y_x = models['dec_x'](models['enc_y'](x_y))
else:
x_y = models['gen_g'](x)
x_y_x = models['gen_f'](x_y)
# for i, var in enumerate([x, x_y]):
for i, var in enumerate([x, x_y, x_y_x]):
imgs = postprocess(var).astype(np.float32)
for j in range(len(imgs)):
ax = fig.add_subplot(gs[j, i])
if imgs[j].shape[2] == 1:
ax.imshow(imgs[j, :, :, 0],
interpolation='none',
cmap='gray',
vmin=0,
vmax=1)
else:
ax.imshow(imgs[j], interpolation='none', vmin=0, vmax=1)
ax.set_xticks([])
ax.set_yticks([])
with chainer.using_config('train', False):
with chainer.function.no_backprop_mode():
if len(models) > 2:
if self.single_encoder:
y_x = models['dec_x'](models['enc_x'](y))
else:
y_x = models['dec_x'](models['enc_y'](y))
# y_x_y = models['dec_y'](models['enc_y'](y)) ## autoencoder
y_x_y = models['dec_y'](models['enc_x'](y_x))
else: # (gen_g, gen_f)
y_x = models['gen_f'](y)
y_x_y = models['gen_g'](y_x)
# for i, var in enumerate([y, y_y]):
for i, var in enumerate([y, y_x, y_x_y]):
imgs = postprocess(var).astype(np.float32)
for j in range(len(imgs)):
ax = fig.add_subplot(gs[j + len(batch_x), i])
if imgs[j].shape[2] == 1:
ax.imshow(imgs[j, :, :, 0],
interpolation='none',
cmap='gray',
vmin=0,
vmax=1)
else:
ax.imshow(imgs[j], interpolation='none', vmin=0, vmax=1)
ax.set_xticks([])
ax.set_yticks([])
gs.tight_layout(fig)
plt.savefig(os.path.join(self.vis_out,
'count{:0>4}.jpg'.format(self.count)),
dpi=200)
self.count += 1
plt.close()
cycle_y_l1 = F.mean_absolute_error(y, y_x_y)
cycle_y_l2 = F.mean_squared_error(y, y_x_y)
cycle_x_l1 = F.mean_absolute_error(x, x_y_x)
id_xy_grad = losses.loss_grad(x, x_y)
id_xy_l1 = F.mean_absolute_error(x, x_y)
result = {
"myval/cycle_y_l1": cycle_y_l1,
"myval/cycle_y_l2": cycle_y_l2,
"myval/cycle_x_l1": cycle_x_l1,
"myval/id_xy_grad": id_xy_grad,
"myval/id_xy_l1": id_xy_l1
}
return result
def __call__(self, x_data, y_data):
x = Variable(x_data.astype(np.float32).reshape(len(x_data),
1)) # Variableオブジェクトに変換
y = Variable(y_data.astype(np.float32).reshape(len(y_data),
1)) # Variableオブジェクトに変換
return F.mean_squared_error(self.predict(x), y)
def __call__(self, x, y):
y_p = self.model(x)
loss = F.mean_squared_error(y_p, y)
reporter.report({'loss': loss}, self)
return loss
def __call__(self, X, Y):
predY = self.predict(X)
loss = F.mean_squared_error(predY, Y)
return loss
def __call__(self, x, y):
return F.mean_squared_error(self.fwd(x), y)
def __call__(self, x, t):
y = self.model(x)
loss = F.mean_squared_error(y, t)
return loss
def loss_gen_fm(self, gen, real_fm_expected, fake_fm_expected):
fm_loss = F.mean_squared_error(real_fm_expected, fake_fm_expected)
chainer.report({'loss': fm_loss}, gen) # gen/loss でアクセス可能
return fm_loss
m_list.append(m)
noise_list.append(noise)
y = generator(const, m_list, stage, noise_list, alpha)
y_dis = discriminator(y, stage, alpha)
x_dis = discriminator(x_down, stage, alpha)
dis_loss = F.mean(F.softplus(-x_dis)) + F.mean(F.softplus(y_dis))
eps = xp.random.uniform(0,1,size = batchsize).astype(xp.float32)[:,None,None,None]
x_mid = eps * y + (1.0 - eps) * x_down
y_mid = F.sum(discriminator(x_mid, stage, alpha))
grad, = chainer.grad([y_mid], [x_mid], enable_double_backprop=True)
grad = F.sqrt(F.sum(grad*grad, axis=(1,2,3)))
loss_gp = lambda_gp * F.mean_squared_error(grad, xp.ones_like(grad.data))
y.unchain_backward()
dis_loss += loss_gp
discriminator.cleargrads()
dis_loss.backward()
dis_opt.update()
dis_loss.unchain_backward()
m_list = []
noise_list = []
z1 = chainer.as_variable(xp.random.normal(size=(batchsize, 512)).astype(xp.float32))
z2 = chainer.as_variable(xp.random.normal(size=(batchsize, 512)).astype(xp.float32))
switch_point = np.random.randint(1, 2*stage)
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 mse(self, x, y, undo_norm):
y = Variable(np.array(y, dtype=np.float32))
pred = undo_norm(self.prediction(x))
return F.mean_squared_error(pred, y)
def __call__(self, t, sw):
hf0 = self.fx0(vgg.h)
hf1 = self.fx1(hf0)
fusionG = xp.tile(hf1.data,(14,14,1))
fusionG = fusionG.transpose(2,0,1)
fusionG = fusionG[np.newaxis,:,:,:]
fusionL = self.convf_0(vgg.h5)
fusionL = fusionL.data
fusion = xp.concatenate([fusionG, fusionL], axis=1)
h0 = F.relu(self.conv5_1(self.bn0(Variable(fusion))))
h0 = self.deconv5_1(h0)
h1 = self.bn1(vgg.h4)
h2 = h0 + h1
h2 = self.conv4_1(h2)
h2 = F.relu(self.bn2(h2))
h2 = self.deconv4_1(h2)
h3 = self.bn2(vgg.h3)
h4 = h2 + h3
h4 = self.conv4_2(h4)
h4 = F.relu(self.bn3(h4))
h4 = self.deconv3_1(h4)
h5 = self.bn4(vgg.h2)
h5 = h4 + h5
h5 = self.conv3_1(h5)
h5 = F.relu(self.bn5(h5))
h5 = self.deconv2_1(h5)
h6 = self.bn6(vgg.h1)
h6 = h5 + h6
h6 = self.conv2_1(h6)
h6 = F.relu(self.bn7(h6))
h7 = self.bn8(vgg.x)
h8 = h6 + h7
h8 = self.conv1_1(h8)
h8 = F.relu(self.bn9(h8))
h8 = self.conv0_5(h8)
zx81 = F.split_axis(h8,1,0)
zx82 = F.split_axis(zx81,2,1)
if sw == 1:
t1 = F.reshape(t,(1, 224*224))
x = F.reshape(zx82[0],(1, 224*224))
self.loss = F.mean_squared_error(x, t1)
return self.loss
elif sw == 2:
t1 = F.reshape(t,(1, 224*224))
x = F.reshape(zx82[1],(1, 224*224))
self.loss = F.mean_squared_error(x, t1)
return self.loss
else:
return h8
def __call__(self, x, y):
y = Variable(np.array(y, dtype=np.float32))
pred = self.prediction(x)
return F.mean_squared_error(pred, y)
def __call__(self, x_recon, x):
bs = x.shape[0]
d = np.prod(x.shape[1:])
self.loss = F.mean_squared_error(x_recon, x) / d
return self.loss
for i in range(2): #six.moves.range(jump*n_epoch):
print i
x_batch = np.array([[goes_flux[(jump * j + i) % whole_len]]
for j in six.moves.range(batchsize)],
dtype=np.float32)
truth_data = np.array([[goes_max[(jump * j + i) % whole_len]]
for j in six.moves.range(batchsize)],
dtype=np.float32)
if args.gpu >= 0:
truth_data = cuda.to_gpu(y_batch)
truth = chainer.Variable(truth_data, volatile=False)
state, y = forward_one_step(x_batch, state)
loss_i = F.mean_squared_error(y, truth)
accum_loss += loss_i
cur_log_perp += loss_i.data.reshape(())
optimizer.clip_grads(grad_clip)
optimizer.update()
if (i + 1) % bprop_len == 0: # Run truncated BPTT
print i, jump * n_epoch
optimizer.zero_grads()
print accum_loss.data
accum_loss.backward()
accum_loss.unchain_backward() # truncate
accum_loss = chainer.Variable(mod.zeros((), dtype=np.float32))
optimizer.clip_grads(grad_clip)
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)
x_real, real_label = zip(*batch)
x_real = self.xp.asarray(x_real).astype("f")
x_real_std = self.xp.std(x_real, axis=0, keepdims=True)
x_rnd = self.xp.random.uniform(0, 1, x_real.shape).astype("f")
one_hot_label, x_real, real_label_image = self.make_label_infomation(
real_label, x_real, batchsize)
real_label_image = Variable(real_label_image)
x_real = Variable(x_real)
y_real = self.dis(x_real)
z = self.gen.make_hidden(batchsize)
z = Variable(self.xp.concatenate([z, one_hot_label], axis=1))
x_fake = self.gen(z)
x_fake = F.concat([x_fake, real_label_image])
y_fake = self.dis(x_fake)
if i == 0:
loss_gen = F.sigmoid_cross_entropy(
y_fake, self.xp.ones_like(y_fake.data))
self.gen.cleargrads()
loss_gen.backward()
gen_optimizer.update()
chainer.reporter.report({'gen/loss': loss_gen})
x_fake.unchain_backward()
alpha = self.xp.random.uniform(0, 1,
size=batchsize).astype("f")[:, None,
None,
None]
x_perturb = x_real + alpha * self.perturb_range * x_real_std * x_rnd
grad, = chainer.grad([self.dis(x_perturb)], [x_perturb],
enable_double_backprop=True)
grad = F.sqrt(F.batch_l2_norm_squared(grad))
loss_gp = self.lam * \
F.mean_squared_error(grad, xp.ones_like(grad.data))
loss_dis = F.sigmoid_cross_entropy(y_real,
self.xp.ones_like(y_fake.data))
loss_dis += F.sigmoid_cross_entropy(y_fake,
self.zeros_like(y_fake.data))
self.dis.cleargrads()
loss_dis.backward()
loss_gp.backward()
dis_optimizer.update()
chainer.reporter.report({'dis/loss': loss_dis})
chainer.reporter.report({'dis/loss_grad': loss_gp})
chainer.reporter.report({'g': F.mean(grad)})
def getLossDistill(self,x,t):
_t = chainer.Variable(t.data, volatile='off')
self.loss = F.mean_squared_error(x, _t)
return self.loss
def __call__(self, x0, x1):
if self.scaler is not None:
x0 = self.scaler.inverse_transform(x0)
x1 = self.scaler.inverse_transform(x1)
return F.sqrt(F.mean_squared_error(x0, x1))
def mse_gd_loss(x, t, eta=0.5):
mse = F.mean_squared_error(x, t)
gd = chainervr.functions.gradient_difference_error(x, t)
return mse * (1.0 - eta) + gd * eta
def __call__(self, x, t, train=True):
y = self.predictor(x, train)
self.loss = F.mean_squared_error(y, t)
report({'loss': self.loss}, self)
return self.loss
def generate_image(img_orig,
img_style,
width,
nw,
nh,
max_iter,
lr,
img_gen=None):
mid_orig = nn.forward(Variable(img_orig))
style_mats = [get_matrix(y) for y in nn.forward(Variable(img_style))]
if img_gen is None:
if args.gpu >= 0:
img_gen = xp.random.uniform(-20,
20, (1, 3, width, width),
dtype=np.float32)
else:
img_gen = np.random.uniform(-20, 20, (1, 3, width, width)).astype(
np.float32)
img_gen = chainer.links.Parameter(img_gen)
optimizer = optimizers.Adam(alpha=lr)
optimizer.setup(img_gen)
for i in range(max_iter):
img_gen.zerograds()
x = img_gen.W
y = nn.forward(x)
L = Variable(xp.zeros((), dtype=np.float32))
for l in range(len(y)):
ch = y[l].data.shape[1]
wd = y[l].data.shape[2]
gogh_y = F.reshape(y[l], (ch, wd**2))
gogh_matrix = F.matmul(gogh_y, gogh_y, transb=True) / np.float32(
ch * wd**2)
L1 = np.float32(args.lam) * np.float32(
nn.alpha[l]) * F.mean_squared_error(y[l],
Variable(mid_orig[l].data))
L2 = np.float32(nn.beta[l]) * F.mean_squared_error(
gogh_matrix, Variable(style_mats[l].data)) / np.float32(len(y))
L += L1 + L2
if i % 100 == 0:
print(i, l, L1.data, L2.data)
L.backward()
img_gen.W.grad = x.grad
optimizer.update()
tmp_shape = x.data.shape
if args.gpu >= 0:
img_gen.W.data += Clip().forward(
img_gen.W.data).reshape(tmp_shape) - img_gen.W.data
else:
def clip(x):
return -120 if x < -120 else (136 if x > 136 else x)
img_gen.W.data += np.vectorize(clip)(
img_gen.W.data).reshape(tmp_shape) - img_gen.W.data
if i % 50 == 0:
save_image(img_gen.W.data, W, nw, nh, i)
def forward_one_step(self,
x_data,
y_data,
n_layers_recog,
n_layers_gen,
nonlinear_q='softplus',
nonlinear_p='softplus',
output_f='sigmoid',
type_qx='gaussian',
type_px='gaussian',
gpu=-1):
x = Variable(x_data)
y = Variable(y_data)
# set non-linear function
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_activation = {
'sigmoid': F.sigmoid,
'identity': self.identity,
'tanh': F.tanh
}
output_a_f = output_activation[output_f]
hidden_q = [nonlinear_f_q(self.recog_x(x) + self.recog_y(y))]
# compute q(z|x, y)
for i in range(n_layers_recog - 1):
hidden_q.append(
nonlinear_f_q(getattr(self, 'recog_%i' % i)(hidden_q[-1])))
q_mean = getattr(self, 'recog_mean')(hidden_q[-1])
q_log_sigma = 0.5 * getattr(self, 'recog_log')(hidden_q[-1])
eps = np.random.normal(
0, 1,
(x.data.shape[0], q_log_sigma.data.shape[1])).astype('float32')
if gpu >= 0:
eps = cuda.to_gpu(eps)
eps = Variable(eps)
z = q_mean + F.exp(q_log_sigma) * eps
# compute q(x |y, z)
hidden_p = [nonlinear_f_p(self.gen_y(y) + self.gen_z(z))]
for i in range(n_layers_gen - 1):
hidden_p.append(
nonlinear_f_p(getattr(self, 'gen_%i' % i)(hidden_p[-1])))
hidden_p.append(output_a_f(getattr(self, 'gen_out')(hidden_p[-1])))
output = hidden_p[-1]
rec_loss = F.mean_squared_error(output, x)
KLD = -0.5 * F.sum(1 + q_log_sigma - q_mean**2 -
F.exp(q_log_sigma)) / (x_data.shape[0] *
x_data.shape[1])
return rec_loss, KLD, output
def loss(self, x):
y = self(x)
return F.mean_squared_error(x, y)
def main():
parser = argparse.ArgumentParser(
description='Second order polynomial approximation with MLP')
parser.add_argument('--n_hidden_units',
'-n',
type=int,
default=3,
help='Number of hidden units per hidden layer')
parser.add_argument('--batch_size',
'-b',
type=int,
default=25,
help='Batch size for each epoch')
parser.add_argument('--n_epochs',
'-e',
type=int,
default=1000,
help='Number of epochs to run')
args = parser.parse_args()
''' generate data '''
x = np.random.rand(1000).astype(np.float32)
y = 19 * x**3 + 10 * x**2 - 8 * x + 7
y += 6.3423 * np.random.rand(1000).astype(np.float32)
N = args.batch_size * 10
print('Number of hidden units: {}'.format(args.n_hidden_units))
print('Number of epocsh: {}'.format(args.n_epochs))
print('Batch size: {}'.format(args.batch_size))
print('N: {}'.format(N))
print('')
x_train, x_test = np.split(x, [N])
y_train, y_test = np.split(y, [N])
print('Shape of train, test data: {}, {}'.format(x_train.shape,
x_test.shape))
foo = raw_input('Enter any key to continue...')
''' instantiate the model and setup optimizer '''
model = QuadChain(args.n_hidden_units)
# optimizer = optimizers.AdaDelta(rho=0.9)
optimizer = optimizers.MomentumSGD()
# optimizer = optimizers.Adam()
optimizer.use_cleargrads()
optimizer.setup(model)
'''
-- prepare test data here
-- train data needs to be shuffled for each epoch; so we will deal with it there
'''
test_data = Variable(x_test.reshape(x_test.shape[0], -1))
test_target = Variable(y_test.reshape(y_test.shape[0], -1))
'''
- start training
- for each epoch, iterate over each mini batches and perform model update
- at the end of each mini batch, calculate test loss
'''
dt = display_train()
dt.header = str("{:^5} {:^5} {:^5} {:^5}".format('Epoch', 'Mini Batch',
'Train Loss',
'Test Loss'))
for each_epoch in xrange(1, args.n_epochs + 1):
permuted_ordering = np.random.permutation(N)
for mini_batch_index in xrange(0, N, args.batch_size):
x_batch = x_train[
permuted_ordering[mini_batch_index:mini_batch_index +
args.batch_size]]
y_batch = y_train[
permuted_ordering[mini_batch_index:mini_batch_index +
args.batch_size]]
train_data = Variable(x_batch.reshape(x_batch.shape[0], -1))
train_target = Variable(y_batch.reshape(y_batch.shape[0], -1))
train_pred = model(train_data)
train_loss = F.mean_squared_error(train_target, train_pred)
model.cleargrads()
train_loss.backward()
optimizer.update()
''' calculate test loss after this mini batch optimizer/network update '''
test_pred = model(test_data)
test_loss = F.mean_squared_error(test_target, test_pred)
logstr = str("{:4}\t{:4d}\t{:10.8} {:10.8}".format(
each_epoch, mini_batch_index, train_loss.data, test_loss.data))
dt.train_log.append(logstr)
if (len(dt.train_log) == dt.H):
dt.display()
