pub_vel.angular.z = 0.8
elif predict_action[1] < -0.8:
pub_vel.angular.z = -0.8
else:
pub_vel.angular.z = predict_action[1]
pub_vel.linear.x = pub_vel.linear.x*0.5
pub_vel.angular.z = pub_vel.angular.z * 0.5
pub.publish(pub_vel)
#print("\r"+str(pub_vel) + " ",end="")
print(str(pub_vel))
#print(predict_action)
step_count += 1
x.append(step_count)
y.append(pub_vel.angular.z*4)
if(len(x)>150):
x.pop(0)
y.pop(0)
plt.xlim(x[0],x[0]+150)
line, = ax.plot(x, y, color='blue')
plt.pause(0.01)
# グラフをクリア
line.remove()
r.sleep()
def create(self):
input = Input(shape=self.input_shape)
def image_channels_slice(image, channel, before):
if before:
return image[:, :, :, :channel]
else:
return image[:, :, :, channel:]
origin = Lambda(image_channels_slice, arguments={'channel': 3, 'before': True})(input)
superpixel = Lambda(image_channels_slice, arguments={'channel': 3, 'before': False})(input)
#####################
# First Convolution #
#####################
tiramisu = Conv2D(self.n_first_filters, kernel_size=3, padding='same',
input_shape=self.origin_shape,
kernel_initializer=self.kernel_initializer,
kernel_regularizer=self.regularizer)(origin)
#####################
# Downsampling path #
#####################
m = 0
skip_connections = []
atrous_rate = 1
for i in range(self.n_downsamples):
if self.use_atrous[i]:
atrous_rate *= 2
for j in range(self.n_layers_per_block[i]):
l = self.AtrousLayer(self.growth_rate, rate=atrous_rate)(tiramisu)
tiramisu = Concatenate()([tiramisu, l])
m += self.growth_rate
skip_connections.append(None)
tiramisu = Concatenate()([tiramisu, self.AtrousLayer(self.growth_rate, rate=atrous_rate)(tiramisu)])
else:
for j in range(self.n_layers_per_block[i]):
l = self.Layer(self.growth_rate)(tiramisu)
tiramisu = Concatenate()([tiramisu, l])
m += self.growth_rate
skip_connections.append(Conv2D(int(m / 2), kernel_size=1, kernel_initializer=self.kernel_initializer)(tiramisu))
# skip_connections.append(tiramisu)
tiramisu = self.TransitionDown(m)(tiramisu)
if skip_connections[0] == None:
skip_connections[0] = superpixel
else:
skip_connections[0] = Concatenate()([skip_connections[0], superpixel])
skip_connections = skip_connections[::-1]
self.use_atrous = self.use_atrous[::-1]
#####################
# Bottleneck #
#####################
upsample_tiramisu = []
for j in range(self.n_layers_per_block[self.n_downsamples]):
l = self.Layer(self.growth_rate)(tiramisu)
upsample_tiramisu.append(l)
tiramisu = Concatenate()([tiramisu, l])
#######################
# Upsampling path #
#######################
for i in range(self.n_downsamples):
if self.use_atrous[i]:
assert(skip_connections[i] == None)
assert(not self.use_bilinear_interplolation[i])
else:
upsample_tiramisu = Concatenate()(upsample_tiramisu)
n_keep_filters = self.growth_rate * self.n_layers_per_block[self.n_downsamples + i]
if self.use_bilinear_interplolation[i]:
tiramisu = self.BilinearUp(2)(upsample_tiramisu, skip_connections[i])
else:
tiramisu = self.TransitionUp(n_keep_filters)(upsample_tiramisu, skip_connections[i])
upsample_tiramisu = []
for j in range(self.n_layers_per_block[self.n_downsamples + i + 1]):
l = self.Layer(self.growth_rate)(tiramisu)
upsample_tiramisu.append(l)
tiramisu = Concatenate()([tiramisu, l])
#####################
# Softmax #
#####################
tiramisu = Conv2D(self.n_classes, kernel_size=1, padding='same',
kernel_initializer=self.kernel_initializer,
kernel_regularizer=self.regularizer)(tiramisu)
tiramisu = Reshape((self.input_shape[0] * self.input_shape[1], self.n_classes))(tiramisu)
tiramisu = Activation('softmax')(tiramisu)
self.model = Model(inputs=input, outputs=tiramisu)
self.model.summary()
with open('tiramisu_fc_dense_model.json', 'w') as outfile:
outfile.write(json.dumps(json.loads(self.model.to_json()), indent=3))