def save(self,
name='new_model',
path=None,
global_step=None,
replace=False):
if not hasattr(self, '_saver'):
self._saver = tfnn.NetworkSaver()
self._saver.save(self, name, path, global_step, replace=replace)
def compare_real(data_path):
load_data = pd.read_pickle(data_path)
s, f = 10700, 10900
xs = load_data.iloc[s:f, 1:]
ys = load_data.a[s:f]
network_saver = tfnn.NetworkSaver()
restore_path = '/tmp/'
network = network_saver.restore(restore_path)
prediction = network.predict(network.normalizer.fit_transform(xs))
plt.plot(np.arange(xs.shape[0]), prediction, 'r--', label='predicted')
plt.plot(np.arange(xs.shape[0]), ys, 'k-', label='real')
plt.legend(loc='best')
plt.show()
network.sess.close()
def save(self, path='/tmp/'):
saver = tfnn.NetworkSaver()
saver.save(self, path)
network = tfnn.ClfNetwork(input_size=mnist.train.images.shape[1],
output_size=mnist.train.labels.shape[1])
# add hidden layer
network.add_hidden_layer(n_neurons=100, activator=tfnn.nn.relu)
# add output layer
network.add_output_layer(activator=None)
# set optimizer. Default GradientDescent
network.set_optimizer()
# set evaluator for compute the accuracy, loss etc.
evaluator = tfnn.Evaluator(network)
# similar to sklearn, we have fit function
network.fit(mnist.train.images, mnist.train.labels, steps=2000)
print('trained network predict:')
print(network.predict(mnist.test.images[10:20]))
print('real data value:')
print(mnist.test.labels[10:20].argmax(axis=1))
# save network
network.save(name='model', path='tmp')
# reload network
saver = tfnn.NetworkSaver()
network2 = saver.restore(name='model', path='tmp')
print('\nLoaded network predict:')
print(network2.predict(mnist.test.images[10:20]))
def test():
network_saver = tfnn.NetworkSaver()
restore_path = '/tmp/'
network = network_saver.restore(restore_path)
test_time = 60
cars = []
for i in range(8):
cars.append(Car(i * -15))
for i in range(test_time * 10):
for j in range(len(cars)):
if j == 0:
if i < 1 * 10:
a = 0
elif 1 * 10 <= i < 15 * 10:
a = 2
elif 15 * 10 <= i < 20 * 10:
a = 0
elif 20 * 10 <= i < 28 * 10:
a = -3
elif 28 * 10 <= i < 30 * 10:
a = 0
elif 30 * 10 <= i < 35 * 10:
a = 3
elif 35 * 10 <= i < 37 * 10:
a = 0
elif 37 * 10 <= i < 45 * 10:
a = -1
elif 45 * 10 <= i < 50 * 10:
a = 2
else:
a = 0
cars[0].ps.append(cars[0].ps[-1] + cars[0].vs[-1] * 0.1 +
1 / 2 * cars[0].acs[-1] * 0.1**2)
v = cars[0].vs[-1] + 0.1 * a
if v < 0:
v = 0
cars[0].vs.append(v)
cars[0].acs.append(a)
else:
if i <= 1 * 10:
a = 0
else:
ss_data = cars[j].ss[-10:]
vs_data = cars[j].vs[-10:]
vs_l_data = cars[j - 1].vs[-10:]
xs_data = np.array(ss_data + vs_data + vs_l_data)
a = network.predict(
network.normalizer.fit_transform(xs_data))
cars[j].ps.append(cars[j].ps[-1] + cars[j].vs[-1] * 0.1 +
1 / 2 * cars[j].acs[-1] * 0.1**2)
v = cars[j].vs[-1] + 0.1 * a
if v < 0:
v = 0
cars[j].vs.append(v)
cars[j].acs.append(a)
cars[j].ss.append(cars[j - 1].ps[-1] - cars[j].ps[-1])
xs = list(range(test_time * 10 + 1))
plt.figure(1)
plt.subplot(411)
for i in range(len(cars)):
if i == 0:
plt.plot(xs, cars[i].ps, 'k-')
else:
plt.plot(xs, cars[i].ps, 'r--')
plt.ylabel('p (m)')
# plt.legend(loc='best')
plt.grid()
plt.subplot(412)
for i in range(len(cars)):
if i == 0:
plt.plot(xs, cars[i].acs, 'k-')
else:
plt.plot(xs, cars[i].acs, 'r--')
plt.ylabel('a (m/s^2)')
# plt.legend(loc='best')
plt.grid()
plt.subplot(413)
for i in range(len(cars)):
if i == 0:
plt.plot(xs, cars[i].vs, 'k-')
else:
plt.plot(xs, cars[i].vs, 'r--')
plt.ylabel('v (m/s)')
# plt.legend(loc='best')
plt.grid()
plt.subplot(414)
for i in range(1, len(cars)):
plt.plot(xs, cars[i].ss, 'r--')
plt.ylabel('space (m)')
# plt.legend(loc='best')
plt.grid()
plt.show()