def load_data(training=True):
"""Adapted from http://g.sweyla.com/blog/2012/mnist-numpy/"""
path = "./data"
if training:
fname_img = os.path.join(path, 'train-images-idx3-ubyte')
fname_lbl = os.path.join(path, 'train-labels-idx1-ubyte')
else:
fname_img = os.path.join(path, 't10k-images-idx3-ubyte')
fname_lbl = os.path.join(path, 't10k-labels-idx1-ubyte')
# Inputs
fimg = open(fname_img, 'rb')
magic_nr, size, rows, cols = struct.unpack(">IIII", fimg.read(16))
imgs = pyarray("B", fimg.read())
fimg.close()
imgs = [imgs[n:n+784] for n in xrange(0, len(imgs), 784)]
inputs = []
for img in imgs:
V = Vol(28, 28, 1, 0.0)
V.w = [ (px / 255.0) for px in img ]
inputs.append(V)
# Outputs
flbl = open(fname_lbl, 'rb')
magic_nr, size = struct.unpack(">II", flbl.read(8))
labels = pyarray("b", flbl.read())
flbl.close()
return zip(inputs, labels)
def load_data(training=True):
"""Adapted from http://g.sweyla.com/blog/2012/mnist-numpy/"""
path = './data'
if training:
fname_img = os.path.join(path, 'train-images-idx3-ubyte')
fname_lbl = os.path.join(path, 'train-labels-idx1-ubyte')
else:
fname_img = os.path.join(path, 't10k-images-idx3-ubyte')
fname_lbl = os.path.join(path, 't10k-labels-idx1-ubyte')
# Inputs
fimg = open(fname_img, 'rb')
magic_nr, size, rows, cols = struct.unpack(">IIII", fimg.read(16))
imgs = pyarray("B", fimg.read())
fimg.close()
imgs = [imgs[n:n + 784] for n in xrange(0, len(imgs), 784)]
inputs = []
V = Vol(28, 28, 1, 0.0)
for img in imgs:
V.w = [(px / 255.0) for px in img]
inputs.append(augment(V, 24))
# Outputs
flbl = open(fname_lbl, 'rb')
magic_nr, size = struct.unpack(">II", flbl.read(8))
labels = pyarray("b", flbl.read())
flbl.close()
return zip(inputs, labels)
def backward(self, reward):
self.latest_reward = reward
self.average_reward_window.add(reward)
self.reward_window.pop(0)
self.reward_window.append(reward)
if not self.learning:
return
self.age += 1
#it is time t+1 and we have to store (s_t, a_t, r_t, s_{t+1}) as new experience
#(given that an appropriate number of state measurements already exist, of course)
if self.forward_passes > self.temporal_window + 1:
n = self.window_size
e = Experience(
self.net_window[n - 2],
self.action_window[n - 2],
self.reward_window[n - 2],
self.net_window[n - 1]
)
if len(self.experience) < self.experience_size:
self.experience.append(e)
else:
ri = randi(0, self.experience_size)
self.experience[ri] = e
#learn based on experience, once we have some samples to go on
#this is where the magic happens...
if len(self.experience) > self.start_learn_threshold:
avcost = 0.0
for k in range(self.tdtrainer.batch_size):
re = randi(0, len(self.experience))
e = self.experience[re]
x = Vol(1, 1, self.net_inputs)
x.w = e.state0
maxact = self.policy(e.state1)
r = e.reward0 + self.gamma * maxact['value']
ystruct = {'dim': e.action0, 'val': r}
stats = self.tdtrainer.train(x, ystruct)
# print(stats['accuracy'])
avcost += stats['loss']
avcost /= self.tdtrainer.batch_size
self.average_loss_window.add(avcost)
def forward(self, V, is_training):
self.in_act = V
A = Vol(1, 1, self.out_depth, 0.0)
# max activation
max_act = max(V.w)
# compute exponentials (carefully to not blow up)
# normalize
exps = [exp(w - max_act) for w in V.w]
exps_sum = float(sum(exps))
exps_norm = [elem / exps_sum for elem in exps]
self.es = exps_norm
A.w = exps_norm
self.out_act = A
return self.out_act
def forward(self, V, is_training):
self.in_act = V
A = Vol(1, 1, self.out_depth, 0.0)
# max activation
max_act = max(V.w)
# compute exponentials (carefully to not blow up)
# normalize
exps = [ exp(w - max_act) for w in V.w ]
exps_sum = float(sum(exps))
exps_norm = [ elem / exps_sum for elem in exps ]
self.es = exps_norm
A.w = exps_norm
self.out_act = A
return self.out_act
def backward(self, reward):
self.latest_reward = reward
self.average_reward_window.add(reward)
self.reward_window.pop(0)
self.reward_window.append(reward)
if not self.learning:
return
self.age += 1
#it is time t+1 and we have to store (s_t, a_t, r_t, s_{t+1}) as new experience
#(given that an appropriate number of state measurements already exist, of course)
if self.forward_passes > self.temporal_window + 1:
n = self.window_size
e = Experience(self.net_window[n - 2], self.action_window[n - 2],
self.reward_window[n - 2], self.net_window[n - 1])
if len(self.experience) < self.experience_size:
self.experience.append(e)
else:
ri = randi(0, self.experience_size)
self.experience[ri] = e
#learn based on experience, once we have some samples to go on
#this is where the magic happens...
if len(self.experience) > self.start_learn_threshold:
avcost = 0.0
for k in xrange(self.tdtrainer.batch_size):
re = randi(0, len(self.experience))
e = self.experience[re]
x = Vol(1, 1, self.net_inputs)
x.w = e.state0
maxact = self.policy(e.state1)
r = e.reward0 + self.gamma * maxact['value']
ystruct = {'dim': e.action0, 'val': r}
stats = self.tdtrainer.train(x, ystruct)
avcost += stats['loss']
avcost /= self.tdtrainer.batch_size
print avcost
self.average_loss_window.add(avcost)
def load_data():
global training_data, testing_data
lfw_people = fetch_lfw_people(min_faces_per_person=70, resize=0.4)
xs = lfw_people.data
ys = lfw_people.target
inputs = []
labels = list(ys)
for face in xs:
V = Vol(50, 37, 1, 0.0)
V.w = list(face)
inputs.append(augment(V, 30))
x_tr, x_te, y_tr, y_te = train_test_split(inputs, labels, test_size=0.25)
training_data = zip(x_tr, y_tr)
testing_data = zip(x_te, y_te)
print 'Dataset made...'
def load_data():
global training_data, testing_data
lfw_people = fetch_lfw_people(min_faces_per_person=70, resize=0.4)
xs = lfw_people.data
ys = lfw_people.target
inputs = []
labels = list(ys)
for face in xs:
V = Vol(50, 37, 1, 0.0)
V.w = list(face)
inputs.append(augment(V, 30))
x_tr, x_te, y_tr, y_te = train_test_split(inputs, labels, test_size=0.25)
training_data = zip(x_tr, y_tr)
testing_data = zip(x_te, y_te)
print 'Dataset made...'
def load_data(training=True):
"""Adapted from http://g.sweyla.com/blog/2012/mnist-numpy/"""
path = "./data"
if training:
fname_img = os.path.join(path, 'train-images-idx3-ubyte')
fname_lbl = os.path.join(path, 'train-labels-idx1-ubyte')
else:
fname_img = os.path.join(path, 't10k-images-idx3-ubyte')
fname_lbl = os.path.join(path, 't10k-labels-idx1-ubyte')
# Inputs
fimg = open(fname_img, 'rb')
magic_nr, size, rows, cols = struct.unpack(">IIII", fimg.read(16))
imgs = pyarray("B", fimg.read())
fimg.close()
imgs = [imgs[n:n + 784] for n in xrange(0, len(imgs), 784)]
inputs = []
for img in imgs:
V = Vol(28, 28, 1, 0.0)
pxs = [0.0] * 784
for n in xrange(784):
if n % 28 == 0:
pxs[n] = 0
else:
pxs[n] = img[n - 1]
V.w = pxs
inputs.append(V)
# Outputs
flbl = open(fname_lbl, 'rb')
magic_nr, size = struct.unpack(">II", flbl.read(8))
labels = pyarray("b", flbl.read())
flbl.close()
return zip(inputs, labels)