def create_ui(self):
self.progress = gtk.ProgressBar()
self.exec_btn = gtk.Button(stock=gtk.STOCK_EXECUTE)
self.exec_btn.connect('clicked', self.__on_start_training)
self.fonts = AdaptableWidget('Font:', gtk.FontButton)
self.pack_start(self.fonts)
self.pack_start(self.progress, False, False)
self.pack_end(self.exec_btn, False, False)
self.thread = None
self.status = None
def create_ui(self):
self.progress = gtk.ProgressBar()
self.exec_btn = gtk.Button(stock=gtk.STOCK_EXECUTE)
self.exec_btn.connect("clicked", self.__on_start_training)
self.fonts = AdaptableWidget("Font:", gtk.FontButton)
self.pack_start(self.fonts)
self.pack_start(self.progress, False, False)
self.pack_end(self.exec_btn, False, False)
self.thread = None
self.status = None
class TrainingPage(GenericPage):
title = 'Auto training'
image = gtk.STOCK_ABOUT
def create_ui(self):
self.progress = gtk.ProgressBar()
self.exec_btn = gtk.Button(stock=gtk.STOCK_EXECUTE)
self.exec_btn.connect('clicked', self.__on_start_training)
self.fonts = AdaptableWidget('Font:', gtk.FontButton)
self.pack_start(self.fonts)
self.pack_start(self.progress, False, False)
self.pack_end(self.exec_btn, False, False)
self.thread = None
self.status = None
def lock_widgets(self, lock):
GenericPage.lock_widgets(self, lock)
lock = not lock
self.fonts.set_sensitive(lock)
self.exec_btn.set_sensitive(lock)
def __on_start_training(self, widget):
self.lock_widgets(True)
# Initialize the cairo surface and relative context
surface = cairo.ImageSurface(cairo.FORMAT_RGB24, IMG_PIXELS,
IMG_PIXELS)
ctx = cairo.Context(surface)
# Disable antialias for font rendering
fo = cairo.FontOptions()
fo.set_antialias(cairo.ANTIALIAS_NONE)
fo.set_hint_style(cairo.HINT_STYLE_NONE)
ctx.set_font_options(fo)
train_dict = defaultdict(list)
for fntbtn in self.fonts.foreach():
font = fntbtn.get_font_name()
# Select new font face
ctx.select_font_face(font, cairo.FONT_SLANT_NORMAL,
cairo.FONT_WEIGHT_NORMAL)
ctx.set_font_size(IMG_PIXELS)
for char in self.get_alphabet():
ctx.set_source_rgb(1, 1, 1)
ctx.rectangle(0, 0, IMG_PIXELS, IMG_PIXELS)
ctx.fill()
ctx.set_source_rgb(0, 0, 0)
x_bearing, y_bearing, width, height = ctx.text_extents(
char)[:4]
fx = IMG_PIXELS / 2
fy = IMG_PIXELS / 2
fx = fx - width / 2 - x_bearing
fy = fy - height / 2 - y_bearing
ctx.move_to(fx, fy)
ctx.show_text(char)
train_dict[char].append(scaled_from_rawdata(
surface.get_data()))
#if DEBUG:
# surface.write_to_png("images/" + str(ord(char)) + ".png")
self.thread = Thread(target=self.__on_train, args=(train_dict, ))
self.thread.setDaemon(True)
self.thread.start()
gobject.timeout_add(500, self.__update_percentage)
def __update_percentage(self):
text, percentage = self.status
self.progress.set_fraction(percentage)
self.progress.set_text(text)
working = (self.thread and self.thread.isAlive() or False)
if not working:
self.lock_widgets(False)
return working
def __on_train(self, tdict):
"""
Learning thread function
@param tdict: a dictionary containing the scaled images
"""
correct = 0
previous = []
keys = tdict.keys()
alp_idx = self.get_alphabet_index()
network = self.owner.get_network(alp_idx)
for i in xrange(100000):
target = choice(keys)
image = choice(tdict[target])
receptor_states = check_receptor(image)
guess = guess_letter(network, receptor_states, alp_idx)
e = network.train(receptor_states, find_target(target, alp_idx))
hit = (guess == target)
if len(previous) == 1000:
previous.pop(0)
previous.append(hit)
success = \
float(len([a for a in previous if a])) / len(previous) * 100
if DEBUG:
sys.stdout.write(
'\rLetter was %s, guessed %s (err: %2.6f) - %s - %.1f%%' % \
(target, guess, e, hit and 'right' or 'wrong', success)
)
correct = hit and correct + 1 or 0
if success > 98 and len(previous) == 1000:
print 'Greater than 98% success over 1000 guesses'
break
if i % 5 == 0:
self.status = (
'%.1f%% - Letter was %s, guessed %s (err: %2.6f) - %s' %
(success, target, guess, e, hit and 'right' or 'wrong'),
success / 100.0)
if i % 100 == 0:
self.owner.save_network(alp_idx)
class TrainingPage(GenericPage):
title = "Auto training"
image = gtk.STOCK_ABOUT
def create_ui(self):
self.progress = gtk.ProgressBar()
self.exec_btn = gtk.Button(stock=gtk.STOCK_EXECUTE)
self.exec_btn.connect("clicked", self.__on_start_training)
self.fonts = AdaptableWidget("Font:", gtk.FontButton)
self.pack_start(self.fonts)
self.pack_start(self.progress, False, False)
self.pack_end(self.exec_btn, False, False)
self.thread = None
self.status = None
def lock_widgets(self, lock):
GenericPage.lock_widgets(self, lock)
lock = not lock
self.fonts.set_sensitive(lock)
self.exec_btn.set_sensitive(lock)
def __on_start_training(self, widget):
self.lock_widgets(True)
# Initialize the cairo surface and relative context
surface = cairo.ImageSurface(cairo.FORMAT_RGB24, IMG_PIXELS, IMG_PIXELS)
ctx = cairo.Context(surface)
# Disable antialias for font rendering
fo = cairo.FontOptions()
fo.set_antialias(cairo.ANTIALIAS_NONE)
fo.set_hint_style(cairo.HINT_STYLE_NONE)
ctx.set_font_options(fo)
train_dict = defaultdict(list)
for fntbtn in self.fonts.foreach():
font = fntbtn.get_font_name()
# Select new font face
ctx.select_font_face(font, cairo.FONT_SLANT_NORMAL, cairo.FONT_WEIGHT_NORMAL)
ctx.set_font_size(IMG_PIXELS)
for char in self.get_alphabet():
ctx.set_source_rgb(1, 1, 1)
ctx.rectangle(0, 0, IMG_PIXELS, IMG_PIXELS)
ctx.fill()
ctx.set_source_rgb(0, 0, 0)
x_bearing, y_bearing, width, height = ctx.text_extents(char)[:4]
fx = IMG_PIXELS / 2
fy = IMG_PIXELS / 2
fx = fx - width / 2 - x_bearing
fy = fy - height / 2 - y_bearing
ctx.move_to(fx, fy)
ctx.show_text(char)
train_dict[char].append(scaled_from_rawdata(surface.get_data()))
# if DEBUG:
# surface.write_to_png("images/" + str(ord(char)) + ".png")
self.thread = Thread(target=self.__on_train, args=(train_dict,))
self.thread.setDaemon(True)
self.thread.start()
gobject.timeout_add(500, self.__update_percentage)
def __update_percentage(self):
text, percentage = self.status
self.progress.set_fraction(percentage)
self.progress.set_text(text)
working = self.thread and self.thread.isAlive() or False
if not working:
self.lock_widgets(False)
return working
def __on_train(self, tdict):
"""
Learning thread function
@param tdict: a dictionary containing the scaled images
"""
correct = 0
previous = []
keys = tdict.keys()
alp_idx = self.get_alphabet_index()
network = self.owner.get_network(alp_idx)
for i in xrange(100000):
target = choice(keys)
image = choice(tdict[target])
receptor_states = check_receptor(image)
guess = guess_letter(network, receptor_states, alp_idx)
e = network.train(receptor_states, find_target(target, alp_idx))
hit = guess == target
if len(previous) == 1000:
previous.pop(0)
previous.append(hit)
success = float(len([a for a in previous if a])) / len(previous) * 100
if DEBUG:
sys.stdout.write(
"\rLetter was %s, guessed %s (err: %2.6f) - %s - %.1f%%"
% (target, guess, e, hit and "right" or "wrong", success)
)
correct = hit and correct + 1 or 0
if success > 98 and len(previous) == 1000:
print "Greater than 98% success over 1000 guesses"
break
if i % 5 == 0:
self.status = (
"%.1f%% - Letter was %s, guessed %s (err: %2.6f) - %s"
% (success, target, guess, e, hit and "right" or "wrong"),
success / 100.0,
)
if i % 100 == 0:
self.owner.save_network(alp_idx)
