def __init__(self, zone2queue, alpha = .1, beta = .01, save = 10000):
self.counter = 0
self.alpha = alpha
self.beta = beta
self.file_out = 'MT_zoneB.csv'
self.file_name = 'MT_zoneB_backup.csv'
#self.MT = genfromtxt(self.file_name, delimiter=',')
self.xi = [30, 30]
self.MT = create_MEGA_THETA(self.xi)
self.MTDIM = shape(self.MT)
self.DELTA = zeros([self.MTDIM[0],self.MTDIM[1]])
self.zone2queue = zone2queue
self.save_when = save
def __init__(self, inlayer, outlayer, alpha = 1, beta = 0.4, mangle_upper = 0, save = 10, threads = 40): '''initiate queues and primary variables''' self.zone2_queue = Queue() self.DELTA_queue = Queue() self.errors_queue = Queue() # initiate NN variables self.counter = 0 self.mangle_upper = mangle_upper self.file_out = 'MT_zoneA.csv' self.file_name = 'MT_zoneA_backup.csv' #self.MT = genfromtxt(self.file_name, delimiter=',') self.xi = [inlayer, outlayer] self.MT = create_MEGA_THETA(self.xi) self.MTDIM = shape(self.MT) self.DELTA = zeros([self.MTDIM[0],self.MTDIM[1]]) self.alpha = alpha self.beta = beta self.save_when = save self.num_threads = threads self.threads = []
# get data from csv file into array
data = genfromtxt("training_data.csv", delimiter=",")
# get a subset for testing
num_cpus = 32
N = 100 * num_cpus # batch size of data subset
y_vals = data[:, 0] # outputs
x_vals = data[:, 1:].T # inputs
x_vals = (x_vals > 0) * x_vals # make inputs 0/1
data_size = size(y_vals)
# define some NN layers
xi = array([100, 10])
# randomly initiate NN values
MT = create_MEGA_THETA(xi)
MTDIM = shape(MT)
DELTA = zeros([MTDIM[0], MTDIM[1]])
# learning rate
alpha = 100 * N
# create some Queues for DELTA, time, good
time_queue = Queue()
good_queue = Queue()
DELTA_queue = Queue()
# initiate queues
time_queue.put(1.0)
good_queue.put(0.0)
def optimise_network(self, xi, all_x, all_y):
'''run backpropagation for digit = self.digit and save results'''
self.MT = create_MEGA_THETA(xi)
self.MTDIM = shape(self.MT)
self.DELTA = zeros([self.MTDIM[0],self.MTDIM[1]])
self.data_size = size(all_y)
# execute backprop procedure until accuracy is high
while self.success < .99999: # success rate goal
self.selection = random.random_integers(0,self.data_size-1,self.N) # stochastic gradient descent
# get some pixels' labels to mangle
self.mangle0 = random.randint(0,self.mangle_upper)
self.mangle1 = random.randint(0,self.mangle_upper)
self.random_mangle0 = random.random_integers(0,self.base_in, self.mangle0)
self.random_mangle1 = random.random_integers(0,self.base_in, self.mangle1)
self.batch_x_vals = all_x[:,self.selection] # SGD batch
self.batch_y_vals = all_y[self.selection] # SGD batch
# mangle pixels, some --> 0 some --> 1
self.batch_x_vals[self.random_mangle0,:] = 0
self.batch_x_vals[self.random_mangle1,:] = 1
self.DELTA_queue.put(self.DELTA)
self.DELTA_queue.task_done()
# run backprop optimisation w/ multi-threading
self.threads = []
# set NN input/output layers
for i in range(0,self.N):
# input vector
self.x = self.batch_x_vals[:,i]
self.x = kron(self.x, self.x)
self.y = zeros(2) # output of NN (row-less)
if self.batch_y_vals[i] == self.digit:
self.y[0] = 1 # indicate digit
else:
self.y[1] = 1 # indicate NOT digit
self.thr = threading.Thread(target=forwardBackward,
args=(xi, self.x, self.y, self.MT, self.time_queue, self.good_queue, self.DELTA_queue))
self.threads.append(self.thr)
# start all threads
for i in range(0,self.N):
self.threads[i].start()
# spin until all threads finish
for i in range(0,self.N):
self.threads[i].join()
self.DELTA_queue.join()
self.good = self.good_queue.get()
self.time = self.time_queue.get()
self.success = self.good/self.time
print 'Success['+ str(self.digit) +']: ', self.success
self.good_queue.put(self.good)
self.good_queue.task_done()
self.time_queue.put(self.time)
self.time_queue.task_done()
self.DELTA = self.DELTA_queue.get()
self.MT = self.MT - (self.alpha/self.N)*self.DELTA - self.beta*self.MT
self.DELTA = 0*self.DELTA
if self.counter == 50:
savetxt(self.file_out, self.MT, delimiter=',')
print 'Success ['+ str(self.digit) +']: ', self.success
print '### SAVED DATA to '+ self.file_out +' ###'
self.counter = 0 # reset
self.counter += 1
savetxt(self.file_out, self.MT, delimiter=',')
