def test_save_loss_graphs_no_class_weight(self):
x = np.random.uniform(-1, 1, self.x_shape)
x = Variable(x.astype(np.float32))
t = np.random.randint(
0, 12, (self.x_shape[0], self.x_shape[2], self.x_shape[3]))
t = Variable(t.astype(np.int32))
for depth in six.moves.range(1, self.n_encdec + 1):
model = segnet.SegNet(n_encdec=self.n_encdec,
n_classes=12,
in_channel=self.x_shape[1])
model = segnet.SegNetLoss(model,
class_weight=None,
train_depth=depth)
y = model(x, t)
cg = build_computational_graph([y],
variable_style=_var_style,
function_style=_func_style).dump()
for e in range(1, self.n_encdec + 1):
self.assertTrue(
'encdec{}'.format(e) in model.predictor._children)
fn = 'tests/SegNet_xt_depth-{}_{}.dot'.format(self.n_encdec, depth)
if os.path.exists(fn):
continue
with open(fn, 'w') as f:
f.write(cg)
subprocess.call('dot -Tpng {} -o {}'.format(
fn, fn.replace('.dot', '.png')),
shell=True)
def get_xt(self):
x = np.random.uniform(-1, 1, self.x_shape)
x = Variable(x.astype(np.float32))
t = np.random.randint(
0, self.n_classes,
(self.x_shape[0], self.x_shape[2], self.x_shape[3]))
t = Variable(t.astype(np.int32))
return x, t
def test():
args = parse_args()
cfg = Config.from_file(args.config)
data_path = cfg.test.source_data
out = cfg.test.out
model_path = cfg.test.gen_path
print('GPU: {}'.format(args.gpu))
# number of gesture classes and users
n_gesture = cfg.train.n_gesture
n_user = len(cfg.train.dataset_dirs)
n_style = n_gesture if cfg.style == 'gesture' else n_user
## Import generator model
gen = getattr(models, cfg.train.generator.model)(cfg.train.generator,
n_style=n_style)
serializers.load_npz(model_path, gen)
print('')
print(f'loading generator weight from {model_path}')
## Set GPU
if args.gpu >= 0:
# Make a specified GPU current
chainer.backends.cuda.get_device_from_id(args.gpu).use()
gen.to_gpu()
test_data = data_load(data_path, ges_class=cfg.test.ges_class)
test_data = np.expand_dims(test_data, axis=1)
style_label = np.zeros((test_data.shape[0], n_style * 2, 1, 1))
source = cfg.test.ges_class if cfg.style == 'gesture' else cfg.source_user
target = cfg.test.target_style
style_label[:, source] += 1
style_label[:, target + n_style] += 1
test_data = Variable(cuda.to_gpu(test_data.astype(np.float32)))
style_label = Variable(cuda.to_gpu(style_label.astype(np.float32)))
with chainer.using_config('train', False), chainer.using_config(
'enable_backprop', False):
gen_data = gen(test_data, style_label)
gen_data.to_cpu()
if cfg.style == 'gesture':
save_path = f'./{out}/user{cfg.source_user}/ges{target}_from_ges{source}'
else:
save_path = f'./{out}/user{target}/ges{cfg.test.ges_class}_from_user{source}'
if not os.path.exists(os.path.dirname(save_path)):
os.makedirs(os.path.dirname(save_path))
print('saving generated data to ' + save_path + '.npy')
np.save(save_path, gen_data.data)
def real_batch(self, iterator):
batch = iterator.next()
x, y = self.converter(batch, self.device)
c = x[:, -3:, :, :]
x = x[:, :-3, :, :]
#cast 16bit -> 32bit (cannot use tensor core)
x = Variable(x.astype('float32'))
y = Variable(y.astype('float32'))
c = Variable(c.astype('float32'))
return x, y, c
def update(self, trajs):
obs = np.concatenate([traj['observations'] for traj in trajs], axis=0)
if self.concat_time:
ts = np.concatenate([np.arange(len(traj['observations'])) / self.env_spec.timestep_limit for traj in trajs],
axis=0)
obs = np.concatenate([obs, ts[:, None]], axis=-1)
returns = np.concatenate([traj['returns'] for traj in trajs], axis=0)
baselines = np.concatenate([traj['baselines']
for traj in trajs], axis=0)
# regress to a mixture of current and past predictions
targets = returns * (1. - self.mixture_fraction) + \
baselines * self.mixture_fraction
# use lbfgs to perform the update
cur_params = get_flat_params(self)
obs = Variable(obs)
targets = Variable(targets.astype(np.float32))
def f_loss_grad(x):
set_flat_params(self, x)
self.cleargrads()
values = self.compute_baselines(obs)
loss = F.mean(F.square(values - targets))
loss.backward()
flat_grad = get_flat_grad(self)
return loss.data.astype(np.float64), flat_grad.astype(np.float64)
new_params = scipy.optimize.fmin_l_bfgs_b(
f_loss_grad, cur_params, maxiter=10)[0]
set_flat_params(self, new_params)
def real_batch(self, iter_key='main'):
batch = self.get_iterator(iter_key).next()
batch = self.converter(batch, self.device)
if isinstance(batch, tuple) or isinstance(batch, list):
x, t = batch
#16bit -> 32bit (not use tensor core)
x = Variable(x.astype('float32'))
t = Variable(t.astype('float32'))
return x, t
x = Variable(batch.astype('float32'))
return x, None
def compute_loss(self, state, action, reward, next_state, episode_ends):
batchsize = state.shape[0]
xp = self.dqn.model.xp
with chainer.using_config("train", True):
q = self.dqn.compute_q_value(state)
with chainer.no_backprop_mode():
max_target_q_data = self.dqn.compute_target_q_value(
next_state).data
max_target_q_data = xp.amax(max_target_q_data, axis=1)
t = reward + (1 -
episode_ends) * self.discount_factor * max_target_q_data
t = Variable(xp.reshape(t.astype(xp.float32), (-1, 1)))
y = functions.reshape(functions.select_item(q, action), (-1, 1))
if self.clip_loss:
loss = functions.huber_loss(t, y, delta=1.0)
else:
loss = functions.mean_squared_error(t, y) / 2
loss = functions.sum(loss)
# check NaN
loss_value = float(loss.data)
if loss_value != loss_value:
import pdb
pdb.set_trace()
return loss
def make_img(self, x, l, num_lm, random=0):
if random == 0:
lm = Variable(xp.clip(l.data, 0, 1))
else:
eps = xp.random.normal(0, 1, size=l.data.shape).astype(xp.float32)
lm = xp.clip(l.data + eps * xp.sqrt(self.vars), 0, 1)
lm = Variable(lm.astype(xp.float32))
if self.use_gpu:
xm = make_sampled_image.generate_xm_const_size_gpu(
lm.data,
self.gsize,
x.data,
num_lm,
g_size=self.gsize,
img_size=self.img_size)
else:
xm = make_sampled_image.generate_xm_const_size(
lm.data,
self.gsize,
x.data,
num_lm,
g_size=self.gsize,
img_size=self.img_size)
return xm, lm
def make_img(self, x, l, s, num_lm, random=0):
if random == 0:
lm = Variable(xp.clip(l.data, 0, 1))
sm = Variable(xp.clip(s.data, 0, 1))
else:
eps = xp.random.normal(0, 1, size=l.data.shape).astype(xp.float32)
epss = xp.random.normal(0, 1, size=s.data.shape).astype(xp.float32)
sm = xp.clip((s.data + xp.sqrt(self.var) * epss), 0,
1).astype(xp.float32)
lm = xp.clip(l.data + eps * xp.sqrt(self.vars), 0, 1)
sm = Variable(sm)
lm = Variable(lm.astype(xp.float32))
if self.use_gpu:
xm = make_sampled_image.generate_xm_rgb_gpu(lm.data,
sm.data,
x,
num_lm,
g_size=self.gsize)
else:
xm = make_sampled_image.generate_xm_rgb(lm.data,
sm.data,
x,
num_lm,
g_size=self.gsize)
return xm, lm, sm
def pool1(self, x):
x = Variable(
x.astype(cp.float32).reshape(-1, 1, self.atomsize, self.lensize))
h = F.leaky_relu(self.bn1(self.conv1(x))) # 1st conv
h = F.average_pooling_2d(h, (self.k2, 1),
stride=self.s2,
pad=(self.k2 // 2, 0)) # 1st pooling
return h.data
def fingerprint(self, x):
x = Variable(
x.astype(cp.float32).reshape(-1, 1, self.atomsize, self.lensize))
h = F.leaky_relu(self.bn1(self.conv1(x))) # 1st conv
h = F.average_pooling_2d(h, (self.k2, 1),
stride=self.s2,
pad=(self.k2 // 2, 0)) # 1st pooling
h = F.leaky_relu(self.bn2(self.conv2(h))) # 2nd conv
h = F.average_pooling_2d(h, (self.k3, 1),
stride=self.s3,
pad=(self.k3 // 2, 0)) # 2nd pooling
h = F.max_pooling_2d(h,
(self.l4, 1)) # grobal max pooling, fingerprint
return h.data
def test(gen_infogan,gen_cgan,n_tests,n_z_infogan, n_z_cgan, n_continuous):
# cinfogan
z_infogan = np.random.uniform(-2, 2, (n_tests, n_z_infogan+n_continuous))
z_infogan = Variable(np.array(z_infogan, dtype=np.float32))
# cgan
z_cgan = np.random.uniform(-2, 2, (n_tests, n_z_cgan))
z_cgan = Variable(np.array(z_cgan, dtype=np.float32))
x = np.random.uniform(0.0, 1.0, (n_tests,12));
x = Variable(x.astype(np.float32))
infogan_xi= gen_infogan(z_infogan,x)
cgan_xi = gen_cgan(z_cgan,x)
return collisions_measure(x.data,cgan_xi.data,infogan_xi.data)
def make_img(self, x, l, num_lm, random=0):
s = xp.log10(xp.ones((1, 1)) * self.gsize / self.img_size) + 1
sm = xp.repeat(s, num_lm, axis=0)
if random == 0:
lm = Variable(xp.clip(l.data, 0, 1))
else:
eps = xp.random.normal(0, 1, size=l.data.shape).astype(xp.float32)
lm = xp.clip(l.data + eps * xp.sqrt(self.vars), 0, 1)
lm = Variable(lm.astype(xp.float32))
if self.use_gpu:
xm = make_sampled_image.generate_xm_rgb_gpu(lm.data, sm, x, num_lm, g_size=self.gsize)
else:
xm = make_sampled_image.generate_xm_rgb(lm.data, sm, x, num_lm, g_size=self.gsize)
return xm, lm
def sim():
zeta0 = (np.random.uniform(-10.0, 10.0, (1, 1, H, W)).astype(np.float32))
zeta0 = Variable(chainer.cuda.to_gpu(zeta0.astype(np.float32)),
volatile=True)
for it in range(100):
zeta0 += 0.1 * lap.forward(zeta0)
zeta = 0.0 + zeta0
psi = poisson_jacobi(zeta, num_iter=1000)
rho = Variable(rho0, volatile=True)
for i in range(10000):
psi = poisson_jacobi(zeta, x0=psi)
dpdx = FTCS_X().forward(psi) # -vy
dpdy = FTCS_Y().forward(psi) # vx
dzdx = upwind(Kawamura_X().forward(zeta), dpdy)
dzdy = upwind(Kawamura_Y().forward(zeta), -dpdx)
lapz = lap.forward(zeta)
rho_ = rho - 0.5 * dt * (
dpdy * upwind(Kawamura_X().forward(rho), dpdy) - dpdx * upwind(
Kawamura_Y().forward(rho), -dpdx) - 0.000 * lap.forward(rho))
rho_.data[0, 0, :, 0] = rho_.data[0, 0, :, 499]
sum_rho = chainer.functions.sum(rho_)
rho_ = rho_ / (xp.zeros_like(rho.data) + sum_rho)
dzdt = dpdx * dzdy - dpdy * dzdx + nu * lapz
zeta_ = zeta + 0.5 * dt * dzdt
psi = poisson_jacobi(zeta_, x0=psi)
dpdx = FTCS_X().forward(psi) # -vy
dpdy = FTCS_Y().forward(psi) # vx
dzdx = upwind(Kawamura_X().forward(zeta_), dpdy)
dzdy = upwind(Kawamura_Y().forward(zeta_), -dpdx)
lapz = lap.forward(zeta_)
rho = rho - dt * (dpdy * upwind(Kawamura_X().forward(rho_), dpdy) -
dpdx * upwind(Kawamura_Y().forward(rho_), -dpdx) -
0.000 * lap.forward(rho_))
rho.data[0, 0, :, 0] = rho.data[0, 0, :, 499]
sum_rho = chainer.functions.sum(rho)
rho = rho / (xp.zeros_like(rho.data) + sum_rho)
dzdt = dpdx * dzdy - dpdy * dzdx + nu * lapz
zeta = zeta + dt * dzdt
if i % 10 == 0:
yield zeta, psi, rho, i
def sim():
zeta0 = (np.random.uniform(-10.0, 10.0, (1,1,H,W)).astype(np.float32))
zeta0 = Variable(chainer.cuda.to_gpu(zeta0.astype(np.float32)), volatile=True)
for it in range(100):
zeta0 += 0.1*lap.forward(zeta0)
zeta = 0.0 + zeta0
psi = poisson_jacobi(zeta, num_iter=1000)
rho = Variable(rho0, volatile=True)
for i in range(10000):
psi = poisson_jacobi(zeta, x0=psi)
dpdx = FTCS_X().forward(psi) # -vy
dpdy = FTCS_Y().forward(psi) # vx
dzdx = upwind(Kawamura_X().forward(zeta), dpdy)
dzdy = upwind(Kawamura_Y().forward(zeta), -dpdx)
lapz = lap.forward(zeta)
rho_ = rho-0.5*dt*(dpdy*upwind(Kawamura_X().forward(rho), dpdy)-dpdx*upwind(Kawamura_Y().forward(rho), -dpdx) - 0.000*lap.forward(rho))
rho_.data[0,0,:,0] = rho_.data[0,0,:,499]
sum_rho = chainer.functions.sum(rho_)
rho_ = rho_/(xp.zeros_like(rho.data)+sum_rho)
dzdt = dpdx*dzdy - dpdy*dzdx + nu*lapz
zeta_ = zeta+0.5*dt * dzdt
psi = poisson_jacobi(zeta_, x0=psi)
dpdx = FTCS_X().forward(psi) # -vy
dpdy = FTCS_Y().forward(psi) # vx
dzdx = upwind(Kawamura_X().forward(zeta_), dpdy)
dzdy = upwind(Kawamura_Y().forward(zeta_), -dpdx)
lapz = lap.forward(zeta_)
rho = rho - dt*(dpdy*upwind(Kawamura_X().forward(rho_), dpdy)-dpdx*upwind(Kawamura_Y().forward(rho_), -dpdx) - 0.000*lap.forward(rho_))
rho.data[0,0,:,0] = rho.data[0,0,:,499]
sum_rho = chainer.functions.sum(rho)
rho = rho/(xp.zeros_like(rho.data)+sum_rho)
dzdt = dpdx*dzdy - dpdy*dzdx + nu*lapz
zeta = zeta + dt * dzdt
if i%10==0:
yield zeta, psi, rho, i
def update_core(self):
batch = self._iterators['main'].next()
in_arrays = self.converter(batch, self.device)
true_x = in_arrays # mnist (200, 1, 28, 28)
# create input z as random
batchsize = true_x.shape[0]
if self.device == 0:
z = cuda.cupy.random.normal(size=(batchsize, self.z_dim, 1, 1), dtype=np.float32)
z = Variable(z)
else:
z = np.random.uniform(-1, 1, (batchsize, self.z_dim, 1, 1))
z = z.astype(dtype=np.float32)
z = Variable(z)
# G -> x1 -> y of gen
# + -> X -> D -> split
# Truedata -> x2 -> y of true data
gen_output = self.gen(z) # gen_output (200, 1, 28, 28)
x = F.concat((gen_output, true_x), 0) # gen_output + true_data = (400, 1, 28, 28)
dis_output = self.dis(x)
y_gen, y_data = F.split_axis(dis_output, 2, 0) # 0~1 value (200, 1, 1, 1)
# DがGの生成物を1(間違い), TrueDataを0(正しい)と判定するように学習させる
# sigmoid_cross_entropy(x, 0) == softplus(x)
# sigmoid_cross_entropy(x, 1) == softplus(-x)
loss_gen = F.sum(F.softplus(-y_gen))
loss_data = F.sum(F.softplus(y_data))
loss = (loss_gen + loss_data) / batchsize
for optimizer in self._optimizers.values():
optimizer.target.cleargrads()
loss.backward()
for optimizer in self._optimizers.values():
optimizer.update()
reporter.report({'loss':loss, 'gen/loss':loss_gen / batchsize, 'dis/loss':loss_data / batchsize})
save_image(gen_output, self.epoch, self.device)
def test_save_normal_graphs(self):
x = np.random.uniform(-1, 1, self.x_shape)
x = Variable(x.astype(np.float32))
for depth in six.moves.range(1, self.n_encdec + 1):
model = segnet.SegNet(n_encdec=self.n_encdec,
in_channel=self.x_shape[1])
y = model(x, depth)
cg = build_computational_graph([y],
variable_style=_var_style,
function_style=_func_style).dump()
for e in range(1, self.n_encdec + 1):
self.assertTrue('encdec{}'.format(e) in model._children)
fn = 'tests/SegNet_x_depth-{}_{}.dot'.format(self.n_encdec, depth)
if os.path.exists(fn):
continue
with open(fn, 'w') as f:
f.write(cg)
subprocess.call('dot -Tpng {} -o {}'.format(
fn, fn.replace('.dot', '.png')),
shell=True)
def __call__(self, X):
# generate random values
R = np.random.randn(X.data.shape[0], self.rand_sz)
R = Variable(R.astype("float32"))
# attach random to the inputs
h = F.concat([R, X])
#h = R
h = self.ipt(h)
#h = F.dropout(h)
y = self.out(h)
# prior knowledge: environment observation is one - hot vector
obs = F.softmax(y[:, :-2])
# prior knowledge: reward is in [0,1]
rew = F.sigmoid(y[:, [-2]])
fin = F.sigmoid(y[:, [-1]])
y = F.concat([obs, rew, fin])
return y
def __call__(self, x, t):
x = Variable(x.astype(np.float32).reshape(x.shape[0], 4))
t = Variable(t.astype("i"))
y = self.forward(x)
return F.softmax_cross_entropy(y, t), F.accuracy(y, t)
def predict(self, x):
x = Variable(x.astype(np.float32).reshape(x.shape[0],1))
y = self.forward(x)
return y.data
def main():
# Set the number of epochs
parser = argparse.ArgumentParser(description='IaGo:')
parser.add_argument('--epoch', '-e', type=int, default=20, help='Number of sweeps over the dataset to train')
parser.add_argument('--gpu', '-g', type=int, default=0, help='GPU ID to be used')
args = parser.parse_args()
# Model definition
model = network.Value()
optimizer = optimizers.Adam()
optimizer.setup(model)
optimizer.add_hook(chainer.optimizer_hooks.WeightDecay(5e-4))
cuda.get_device(args.gpu).use()
test_x = np.load('./value_data/npy/states_test.npy')
test_y = np.load('./value_data/npy/results_test.npy')
test_x = np.stack([test_x==1, test_x==2], axis=0).astype(np.float32)
test_x = Variable(cuda.to_gpu(test_x.transpose(1,0,2,3)))
test_y = Variable(cuda.to_gpu(test_y.astype(np.float32)))
# Load train dataset
train_x = np.load('./value_data/npy/states.npy')
train_y = np.load('./value_data/npy/results.npy')
train_size = train_y.shape[0]
minibatch_size = 4096 # 2**12
# Learing loop
for epoch in tqdm(range(args.epoch)):
model.to_gpu(args.gpu)
# Should be unnecessary...
#chainer.config.train = True
#chainer.config.enable_backprop = True
# Shuffle train dataset
rands = np.random.choice(train_size, train_size, replace=False)
train_x = train_x[rands,:,:]
train_y = train_y[rands]
# Minibatch learning
for idx in tqdm(range(0, train_size, minibatch_size)):
x = train_x[idx:min(idx+minibatch_size, train_size), :, :]
x = np.stack([x==1, x==2], axis=0).astype(np.float32)
x = Variable(cuda.to_gpu(x.transpose(1,0,2,3)))
y = train_y[idx:min(idx+minibatch_size, train_size)]
y = Variable(cuda.to_gpu(y.astype(np.float32)))
train_pred = model(x)
train_loss = mean_squared_error(train_pred, y)
model.cleargrads()
train_loss.backward()
optimizer.update()
# Calculate loss
with chainer.using_config('train', False):
with chainer.using_config('enable_backprop', False):
test_pred = model(test_x)
test_loss = mean_squared_error(test_pred, test_y)
print('\nepoch :', epoch, ' loss :', test_loss)
# Log
with open("./log_value.txt", "a") as f:
f.write(str(test_loss)[9:15]+", \n")
# Save models
model.to_cpu()
serializers.save_npz('./models/value_model.npz', model)
serializers.save_npz('./models/value_optimizer.npz', optimizer)
def __call__(self, x_data, y_data, train=True, n_patches=32):
if not isinstance(x_data, Variable):
x = Variable(x_data, volatile=not train)
else:
x = x_data
x_data = x.data
#self.n_images = y_data.shape[0]
self.n_images = 1
self.n_patches = x_data.shape[0]
# print x_data.shape[0]
# print y_data.shape[0]
self.n_patches_per_image = self.n_patches / self.n_images
#self.n_patches_per_image = 32
h = F.relu(self.conv1(x))
h = F.relu(self.conv2(h))
h = F.max_pooling_2d(h, 2)
h = F.relu(self.conv3(h))
h = F.relu(self.conv4(h))
h = F.max_pooling_2d(h, 2)
h = F.relu(self.conv5(h))
h = F.relu(self.conv6(h))
h = F.max_pooling_2d(h, 2)
h = F.relu(self.conv7(h))
h = F.relu(self.conv8(h))
h = F.max_pooling_2d(h, 2)
h = F.relu(self.conv9(h))
h = F.relu(self.conv10(h))
h = F.max_pooling_2d(h, 2)
h_ = h
self.h = h_
h = F.dropout(F.relu(self.fc1(h_)), train=train, ratio=0.5)
h = self.fc2(h)
if self.top == "weighted":
a = F.dropout(F.relu(self.fc1_a(h_)), train=train, ratio=0.5)
a = F.relu(self.fc2_a(a)) + 0.000001
t = Variable(y_data, volatile=not train)
self.weighted_loss(h, a, t)
elif self.top == "patchwise":
a = xp.ones_like(h.data)
#y_data.data, to make y_data do not be a Variable
t = xp.repeat(y_data.data, 1)
t = xp.array(t.astype(np.float32))
#print h.data
#print len(t)
#print t
self.patchwise_loss(h, a, t)
if train:
reporter.report({'loss': self.loss}, self)
#print 'self.lose:', self.loss.data
return self.loss
else:
return self.loss, self.y
total_loss, acc = model(x, y)
return total_loss
# model.save_model("./model.npz",save_format="npz")
model.load_model("./model.npz", load_format="npz")
serializers.load_npz("./opt.npz", opt)
step = 0
while True:
# x = Variable(data=np.zeros(shape=[10, 3, 60, 160], dtype=np.float32))
# y = Variable(data=np.zeros(shape=[10, 5], dtype=np.int32))
x, y = yzmData.nextBatch(testOrTrain="train", batch_size=64)
x = Variable(x.astype(np.float32))
y = Variable(y.astype(np.int32))
x.to_gpu(0)
y.to_gpu(0)
opt.update(loss, x.data, y.data)
print("step:%d\ttotal_loss:%f\terro_rate:%f" % (step, model.loss.data, model.acc.data))
if step % 100 == 0:
model.save_model("./model.npz", save_format="npz")
serializers.save_npz("./opt.npz", opt)
test_x, test_t = yzmData.nextBatch(testOrTrain="test", batch_size=40)
test_x = Variable(test_x.astype(np.float32))
test_t = Variable(test_t.astype(np.int32))
test_x.to_gpu(0)
T, F = s_x.shape
Y = np.zeros((8, T, F), dtype=np.complex64)
Y[ch-1, :, :] = s_x
s_x_abs = 20 * log_sp(np.abs(s_x))
s_x_abs = stack_features(s_x_abs.astype(np.float32), 5)
s_x_abs = Variable(s_x_abs)
if args.gpu >= 0:
s_x_abs.to_gpu(args.gpu)
s_x_abs_list.append(s_x_abs)
elif args.single == 1:
audio_data = read_audio(cur_line)
fx, tx, s_x = signal.stft(audio_data, fs=16000, nperseg=512,
noverlap=512-128, nfft=512)
s_x = np.transpose(s_x)
s_x_abs = 20 * log_sp(np.abs(s_x))
s_x_abs = stack_features(s_x_abs.astype(np.float32), 5)
s_x_abs = Variable(s_x_abs)
if args.gpu >= 0:
s_x_abs.to_gpu(args.gpu)
s_x_abs_list.append(s_x_abs)
with chainer.no_backprop_mode():
X = model.predict(s_x_abs_list)
if args.gpu >= 0:
X.to_cpu()
if args.single == 0:
xs = X.data
xs = np.reshape(xs, (8, xs.shape[0] // 8, xs.shape[1]))
taps = 10
xs_power = np.square(exp_sp(xs/20))
Y_hat = np.copy(Y)
def __call__(self, x, t):
x = Variable(x.astype(np.float32).reshape(x.shape[0],1))
t = Variable(t.astype(np.float32).reshape(t.shape[0],1))
return F.mean_squared_error(self.forward(x), t)
def layer1(self, x):
x = Variable(
x.astype(cp.float32).reshape(-1, 1, self.atomsize, self.lensize))
h = self.bn1(self.conv1(x)) # 1st conv
return h.data