def optimise_network(self, all_x, all_y): '''run backpropagation for digit = self.digit and save results''' self.zone2init = fk_zone2(self.zone2_queue) self.z2thread = threading.Thread(target=self.zone2init.start) #self.z2thread.start() self.data_size = shape(all_x)[1] # execute backprop procedure forever while True: # get some pixels' labels to mangle self.selection = random.random_integers(0,self.data_size-1, self.num_threads) #self.mangle0 = random.randint(0,self.mangle_upper) #self.mangle1 = random.randint(0,self.mangle_upper) #self.random_mangle0 = random.random_integers(0,self.xi[0]-1, self.mangle0) #self.random_mangle1 = random.random_integers(0,self.xi[0]-1, self.mangle1) # mangle pixels, some --> 0 some --> 255 #self.x[self.random_mangle0] = 0 #self.x[self.random_mangle1] = 255 self.x = all_x[:,self.selection] self.y = all_y[:,self.selection] # spin up some threads for i in range(0,self.num_threads): th = threading.Thread(target=forwardBackward, args=(self.xi, self.x[:,i], self.y[:,i], self.MT, self.DELTA_queue, self.errors_queue, self.zone2_queue)) self.threads.append(th) print 'spinning' # pause until finished for i in range(0, self.num_threads): self.threads[i].start() for i in range(0, self.num_threads): self.threads[i].join() print 'joining' # debugging: # print self.y # print self.zone1out print amax(self.MT) # print amin(self.MT) # print amax(self.x) # print amin(self.y) # print self.A[-30:] self.errors = self.errors_queue.get()/self.num_threads # run back propagation self.DELTA = back_propagation(self.y, self.A, self.MT, self.xi) self.MT = self.MT - self.alpha*self.DELTA/self.num_threads - self.beta*self.MT if self.counter == self.save_when: savetxt(self.file_out, self.MT, delimiter=',') print 'saved zone1 data to ' + self.file_out + '\t', print 'errors zone1: ' + str(sum(self.errors)) + '\tmax error arg: ' + str(argmax(self.errors)) + '\t MEAE Z1: ' + str(int(amax(self.errors))) + '\tmin error arg: ' + str(argmin(self.errors)) + '\t MinEAE Z1: ' + str(int(amin(self.errors))) self.counter = 0 # reset self.counter += 1 self.DELTA = 0*self.DELTA self.errors = 0*self.errors
def entire_network_back_prop(network, input_layer, cur_model_output, cur_desired_output): # cur_model_output is the current model outputs
print("entire_network_back_prop: Begin")
cost_ini = cost_func(cur_model_output, cur_desired_output)[2]
print("entire_network_back_prop: finished calling cost_func()")
new_cost = []
if len(network.hidden_layers) == 1:
print("entire_network_back_prop: len(network.hidden_layers) == 1")
new_cost = back_propagation(network.hidden_layers[0], input_layer, cost_ini)
else:
for i in range(len(network.hidden_layers)):
print("entire_network_back_prop: i == " + str(i))
if i == 0:
new_cost = back_propagation(network.hidden_layers[len(network.hidden_layers) - 1], network.hidden_layers[len(network.hidden_layers) - 2], cost_ini)
elif i == len(network.hidden_layers) - 1:
new_cost = back_propagation(network.hidden_layers[len(network.hidden_layers) - (i + 1)], input_layer, new_cost)
else:
new_cost = back_propagation(network.hidden_layers[len(network.hidden_layers) - (i + 1)], network.hidden_layers[len(network.hidden_layers) - (i + 2)], new_cost)
print("entire_network_back_prop: Done")
return cost_ini
def forwardBackward(xi, x, y, MT, time_queue, good_queue, DELTA_queue): A = neural_forward(xi, x, MT) check = argmax(A[-xi[-1]:]) # send back some progress statistic if y[check]-1 == 0: good = good_queue.get() good += 1 good_queue.put(good) good_queue.task_done() time = time_queue.get() time += 1 time_queue.put(time) time_queue.task_done() DELTA = DELTA_queue.get() DELTA = DELTA + back_propagation(y, A, MT, xi) DELTA_queue.put(DELTA) DELTA_queue.task_done()
def start(self):
'''start learning process for zone2 of facial keypoints analysis'''
while False:
self.data = self.zone2queue.get()
self.zone2queue.task_done()
self.x = self.data[:,0]
self.y = self.data[:,1]
self.A = neural_forward(self.xi, self.x, self.MT)
self.DELTA = back_propagation(self.y, self.A, self.MT, self.xi)
self.errors = sqrt((self.A[-self.xi[1]:]-self.y)**2)
self.MT = self.MT - self.alpha*self.DELTA - self.beta*self.MT
self.DELTA = 0*self.DELTA
if self.counter == self.save_when:
savetxt(self.file_out, self.MT, delimiter=',')
print 'saved zone2 data to ' + self.file_out + '\t',
print 'errors zone2: ' + str(sum(self.errors)) + '\tmax error arg: ' + str(argmax(self.errors)) + '\t MEAE Z2: ' + str(int(amax(self.errors))) + '\tmin error arg: ' + str(argmin(self.errors)) + '\t MinEAE Z2: ' + str(int(amin(self.errors)))
self.counter = 0 # reset
self.counter += 1
