def __call__(self, x):
h = cf.relu(self.linear1_bn(self.linear1(x)))
h = cf.relu(self.linear2_bn(self.linear2(h)))
bias = cf.reshape(self.linear_bias(h), (-1, self.num_vertices, 3))
bias *= self.scaling
base = self.vertices_base
base = cf.broadcast_to(base[None, :, :], bias.shape)
vertices = base + bias
if self.symmetric:
xy = vertices[:, :, :2] # [bs, nv, 2]
z = cf.absolute(vertices[:, :, 2:3]) # [bs, nv, 1]
vertices = cf.concat((xy, z), axis=2)
vertices = cf.transpose(
cf.tensordot(vertices, self.symmetric_matrix, axes=(1, 0)),
(0, 2, 1))
xy = vertices[:, :, :2] # [bs, nv, 2]
z = vertices[:, :, 2:3] # [bs, nv, 1]
z = z * self.z_sign[None, :, None]
vertices = cf.concat((xy, z), axis=2)
vertices = cf.tanh(vertices) * self.tanh_scale
return vertices, self.faces
def create_2d_window(window_size, channel, xp):
_1D_window = F.reshape(gaussian(window_size, 1.5, xp), (-1, 1))
_2D_window = F.reshape(
F.tensordot(_1D_window, _1D_window.transpose(), axes=1),
(1, 1, -1, window_size))
window = F.repeat(_2D_window, channel, axis=0)
return window
def setUp(self):
self.a = self._setup_tensor(.5, 1, self.a_shape, self.a_dtype)
self.b = self._setup_tensor(.5, 1, self.b_shape, self.b_dtype)
ret_dtype = numpy.result_type(self.a_dtype, self.b_dtype)
self.gc = self._setup_tensor(-1, 1, self.gc_shape, ret_dtype)
self.gga = self._setup_tensor(.5, 1, self.a_shape, self.a_dtype)
self.ggb = self._setup_tensor(.5, 1, self.b_shape, self.b_dtype)
self.op = lambda a, b: F.tensordot(a, b, axes=self.axes)
self.forward_answer = numpy.tensordot(self.a, self.b, self.axes)
def __call__(self, x, is_prob=False):
if is_prob:
x_emb = F.tensordot(x, self.embedding.embed.W, [[2], [0]])
else:
x_emb = self.embedding(x)
x_emb = F.expand_dims(x_emb, 1)
x_emb = normalizing(x_emb, 1)
H = self.encoder(x_emb)
logits = self.disc(H)
return logits, H
def test_invalid_shape(self):
a_data = numpy.zeros((4, 3, 2), dtype=numpy.float32)
b_data = numpy.zeros((2, 3, 5), dtype=numpy.float32)
a = chainer.Variable(a_data)
b = chainer.Variable(b_data)
with self.assertRaises(ValueError):
F.tensordot(a, b)
with self.assertRaises(ValueError):
F.tensordot(a, b, axes=((1, 2), (0, 1)))
with self.assertRaises(ValueError):
F.tensordot(a, b, axes=((0), (0)))
with self.assertRaises(ValueError):
F.tensordot(a, b, axes=((2), (2)))
def __call__(self, x):
batch_size = x.shape[0]
hs = []
for i in range(6):
linear1 = getattr(self, 'linear_p%d_1' % i)
conv1 = getattr(self, 'conv_p%d_1' % i)
conv2 = getattr(self, 'conv_p%d_2' % i)
conv3 = getattr(self, 'conv_p%d_3' % i)
conv4 = getattr(self, 'conv_p%d_4' % i)
linear1_bn = getattr(self, 'linear_p%d_1_bn' % i)
conv1_bn = getattr(self, 'conv_p%d_1_bn' % i)
conv2_bn = getattr(self, 'conv_p%d_2_bn' % i)
conv3_bn = getattr(self, 'conv_p%d_3_bn' % i)
h = linear1(x)
h = h.reshape((h.shape[0], -1, 2, 2))
h = cf.relu(linear1_bn(h))
if not self.use_up_sampling:
h = cf.relu(conv1_bn(conv1(h)))
h = cf.relu(conv2_bn(conv2(h)))
h = cf.relu(conv3_bn(conv3(h)))
else:
h = cf.relu(conv1_bn(conv1(up_sample(h))))
h = cf.relu(conv2_bn(conv2(up_sample(h))))
h = cf.relu(conv3_bn(conv3(up_sample(h))))
h = conv4(h) # [bs, 3, 16, 16]
h = cf.transpose(h, (0, 2, 3, 1)) # [bs, 16, 16, 3]
h = cf.reshape(h, (batch_size, -1, 3)) # [bs, 16 ** 2, 3]
hs.append(h)
h = cf.concat(hs, axis=1) # [bs, 6 * 16 ** 2, 3]
vm = self.xp.tile(self.vertices_matrix[None, :, :], (batch_size, 1, 1))
bias = cf.matmul(vm, h, transa=True)
bias *= self.scaling
base = self.vertices_base
base = cf.broadcast_to(base[None, :, :], bias.shape)
vertices = base + bias
if self.symmetric:
xy = vertices[:, :, :2] # [bs, nv, 2]
z = cf.absolute(vertices[:, :, 2:3]) # [bs, nv, 1]
vertices = cf.concat((xy, z), axis=2)
vertices = cf.transpose(
cf.tensordot(vertices, self.symmetric_matrix, axes=(1, 0)),
(0, 2, 1))
xy = vertices[:, :, :2] # [bs, nv, 2]
z = vertices[:, :, 2:3] # [bs, nv, 1]
z = z * self.z_sign[None, :, None]
vertices = cf.concat((xy, z), axis=2)
vertices = cf.tanh(vertices) * self.tanh_scale
return vertices, self.faces
def __call__(self, x_data, u_h):
"""
Perform one forward step computation given the input x_data,
and the current states of the input neurons (u_io) and the
hidden context layer neurons (u_h).
If u_h is None, the initial states are used instead.
"""
from chainer.backends import cuda
xp = cuda.get_array_module(self.initial_states.W)
x = chainer.Variable(x_data) #.reshape((1,-1)))
if not self.current_y.array is None:
# get mean and variance for prior distribution
pred_mean = self.current_y
# prediction variance is taken from current variance estimation which expresses ability to predict the variance in the training data
# hyp_prior multiplication of factor H
pred_var = self.current_v * (self.hyp_prior)
if not xp.any(pred_var.array):
# It should not happen, but just in case check for zero prediction variance
print("Predicted variance is zero!!!!")
pred_var.array = xp.tile(np.float32(0.00001),
pred_var.array.shape)
# get mean and variance for input distribution
input_mean = x
# if external_signal_variance is not available, set equal to pred_var
if self.external_signal_variance is None:
input_var = pred_var
else:
input_var = chainer.Variable(
xp.tile(
xp.asarray(xp.float32([self.external_signal_variance
])), (x.shape[0], x.shape[1])))
### Bayesian inference ###
if xp.any(xp.isinf(input_var.array)):
sigma_BI = xp.sqrt(pred_var.array)
mu_BI = pred_mean.array
else:
# standard deviation of BI signal
sigma_BI = xp.sqrt(
xp.divide(xp.multiply(input_var.array, pred_var.array),
(input_var.array + pred_var.array)))
# mean of BI signal
mu_BI = xp.power(sigma_BI, 2) * (
xp.divide(pred_mean.array, pred_var.array) +
xp.divide(input_mean.array, input_var.array))
# sample from posterior distribution to get new input for the network
if self.add_BI_variance:
if cuda.get_array_module(self.initial_states.W) == np:
x.array = mu_BI + sigma_BI * xp.float32(
xp.random.randn(sigma_BI.shape[0], sigma_BI.shape[1]))
else:
x.array = mu_BI + sigma_BI * (xp.random.randn(
sigma_BI.shape[0], sigma_BI.shape[1],
dtype=xp.float32))
else:
x.array = mu_BI
self.current_x = x
# in the first time step no u_h is given => set initial states
if u_h is None:
u_h = self.initial_states()[self.classes]
# forward context mapping
# update recurrent connections
h = F.tanh(u_h)
if self.bias_learning:
recurrent_connections = F.transpose(
F.tensordot(self.h_to_h.W, F.transpose(h), axes=1)) + F.tile(
self.h_to_h.b, (h.shape[0], 1))
u_h = chainer.functions.scale(
self.x_to_h(x) + recurrent_connections,
1 / self.tau_c()) + chainer.functions.scale(
u_h, (1 - 1 / self.tau_c()))
else:
u_h = chainer.functions.scale(
self.x_to_h_bias(self.x_to_h(x)) +
self.h_to_h_bias(self.h_to_h(h)),
1 / self.tau_c()) + chainer.functions.scale(
u_h, (1 - 1 / self.tau_c()))
# forward output mapping
u_out = self.h_to_y(h)
y = F.tanh(u_out)
# forward variance mapping
u_v = self.h_to_v(h)
v = F.exp(u_v + self.aberrant_sensory_precision
) + 0.00001 # [Idei et al. 2017]
self.current_y = y
self.current_v = v
return u_h, y, v
'-m',
type=str,
default="model/ResNet-50-model.npz")
parser.add_argument('--image', '-i', type=str, default="data/cat.jpeg")
args = parser.parse_args()
img = utils.read_image(args.image, color=True)
img = F.expand_dims(img, 0)
img = F.resize_images(img, (224, 224))
ax = vis_image(F.squeeze(img, 0).data)
model = ResNetCAM()
chainer.serializers.load_npz(args.model, model)
with chainer.using_config('train', False):
conv_outputs, preds = model(img)
resized_conv_outputs = F.resize_images(conv_outputs, (224, 224))
to_multiply_tensor = model.fc6.W[2].reshape(1, 2048)
heatmap = np.flip(
F.tensordot(to_multiply_tensor, resized_conv_outputs).data, 1)
ax.imshow(heatmap, cmap='jet', alpha=.4)
ax.get_xaxis().set_visible(False)
ax.get_yaxis().set_visible(False)
ax.get_xaxis().set_ticks([])
ax.get_yaxis().set_ticks([])
plt.title("Predicted: " + CLASS_ID[int(F.argmax(preds).data)])
plt.savefig("predicted/result.png")
plt.show()
def forward(self, inputs, device):
a, b = inputs
y = F.tensordot(a, b, axes=self.axes)
return y,
