class EncoderTrainer(ApplicationSession):
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.batch_size = 2
self.needStop = False
self.codec = Codec()
self.model = Model('train')
@asyncio.coroutine
def train(self):
self.publish('semanticaio.encoder.trainer.started')
print('[emit] semanticaio.encoder.trainer.started')
input_dataset = np.zeros((self.batch_size, self.codec.seq_len, self.codec.n_chars), dtype = np.bool)
output_dataset = np.zeros((self.batch_size, self.codec.seq_len, self.codec.n_chars), dtype = np.bool)
while not self.needStop :
yield from asyncio.sleep(0.1)
batch = yield from self.call('semanticaio.db.batch.get', size = self.batch_size)
for i, question in enumerate(batch) :
self.codec.encode(question['sentence'], input_dataset[i])
if question['correctFormId'] == None :
self.codec.encode(question['sentence'], output_dataset[i])
else :
correctQuestion = yield from self.call('semanticaio.db.get', id = question['correctFormId'])
self.codec.encode(correctQuestion['sentence'], output_dataset[i])
(loss, accuracy) = self.model.train(input_dataset, output_dataset)
print('training:', loss, accuracy)
self.needStop = False
self.publish('semanticaio.encoder.trainer.stopped')
print('[emit] semanticaio.encoder.trainer.stopped')
def load(self, *args, **kwargs):
print('[call] semanticaio.encoder.trainer.load')
try :
self.model.load()
except :
print('[error] semanticaio.encoder.trainer.load')
self.model.compile()
def save(self, *args, **kwargs):
print('[call] semanticaio.encoder.trainer.save')
self.model.save()
@asyncio.coroutine
def start(self, *args, **kwargs):
print('[event received] semanticaio.encoder.trainer.start')
yield from self.train()
def stop(self, *args, **kwargs):
print('[event received] semanticaio.encoder.trainer.stop')
self.needStop = True
@asyncio.coroutine
def onJoin(self, details):
yield from self.register(self.load, 'semanticaio.encoder.trainer.load')
yield from self.register(self.save, 'semanticaio.encoder.trainer.save')
yield from self.subscribe(self.start, 'semanticaio.encoder.trainer.start')
yield from self.subscribe(self.stop, 'semanticaio.encoder.trainer.stop')
print('[encoder-trainer started]')
def main(opts):
cap = cv2.VideoCapture(opts.input)
model = Model(66, opts.size)
model.load(opts.weights)
fourcc = cv2.VideoWriter_fourcc(*'mp4v')
out = cv2.VideoWriter('data/res.mp4', fourcc, 10, (640, 480))
while True:
_, img = cap.read()
# Предикт лица (по дефолту тут HOG)
inp_img = img[:, :, ::-1]
face_locations = face_recognition.face_locations(inp_img)
for (top, right, bottom, left) in face_locations:
# Расширяем ббокс и смещаем вертикально
bbox_width = abs(bottom - top)
bbox_height = abs(right - left)
left -= int(2 * bbox_width / 4)
right += int(2 * bbox_width / 4)
top -= int(3 * bbox_height / 4)
bottom += int(bbox_height / 4)
# Выход за пределы
top = max(top, 0)
left = max(left, 0)
bottom = min(img.shape[0], bottom)
right = min(img.shape[1], right)
crop = img[top:bottom, left:right]
# Ресайз по меньшей стороне и кроп от центра
crop = resize_center_crop(crop, opts.size)
# Нормализация
normed_img = normalize(crop)
imgs = []
imgs.append(normed_img)
# Предикт
res = model.test_online(imgs)
# Отрисовка
img = draw_axis(img,
*res,
tdx=(left + right) / 2,
tdy=(top + bottom) / 2,
size=100)
cv2.rectangle(img, (left, top), (right, bottom), (255, 0, 0), 1)
out.write(img)
cv2.imshow('img', img)
if cv2.waitKey(1) == 27:
break
out.release()
def main(opts):
model = Model(66, opts.size)
# Если есть предобучение, то возьмем
if opts.pretrain is not None:
print(f'Initial weights from {opts.pretrain}')
model.load(opts.pretrain)
train_list, val_list = split(opts.data)
# Обучающий лоадер с аугментациями, но там их не сильно много
train_dataset = AFLW2000(train_list,
augment=True,
batch_size=opts.bs,
input_size=opts.size)
val_dataset = AFLW2000(val_list, batch_size=opts.bs, input_size=opts.size)
# Учим
chkpt_name = f'model_size{opts.size}_e{opts.epoch}_lr{opts.lr:.01E}.h5'
model.train(chkpt_name, train_dataset, val_dataset, opts.epoch)
def main(opts):
model = Model(66, opts.size)
model.model.summary()
model.load(opts.weights)
train_list, val_list = split(opts.data)
val_dataset = AFLW2000(val_list, batch_size=1, input_size=opts.size)
err, times = [], []
for idx, (x, y) in enumerate(val_dataset.data_generator()):
print(f'{idx}/{val_dataset.epoch_steps}')
t1 = time()
res = model.test_online(x)
times.append(time() - t1)
ypr = np.array(y)[:, 0, 1]
err.append(abs(ypr - res))
print(f'YPR: {np.mean(np.array(err), axis=0)}')
print(f'TIME: {np.mean(times)}')
if idx == val_dataset.epoch_steps:
break
class Encoder(ApplicationSession):
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.codec = Codec()
self.model = Model('encode')
def load(self, *args, **kwargs):
print('[call] semanticaio.encoder.load')
try :
self.model.load()
except :
print('[error] semanticaio.encoder.load')
self.model.compile()
def _encode(self, question) :
coded_question = np.zeros((self.codec.seq_len, self.codec.n_chars), dtype = np.bool)
self.codec.encode(question, coded_question)
return self.model.encode(coded_question).tolist()
def encode(self, *args, **kwargs):
print('[call] semanticaio.encoder.encode:', kwargs)
result = {}
if 'question' in kwargs :
result['encoded'] = self._encode(kwargs['question'])
elif 'questions' in kwargs :
result['encoded'] = []
for question in kwargs['questions'] :
result['encoded'].append(self._encode(question))
return result
@coroutine
def onJoin(self, details):
yield from self.register(self.load, 'semanticaio.encoder.load')
yield from self.register(self.encode, 'semanticaio.encoder.encode')
print('[encoder started]')
parser.add_argument("--bank", help="--bank 名称(jianhang|renfa|nonghang)")
parser.add_argument("--image",
help="--image 图片名称(位于data/validate/<bank代号>/)")
parser.add_argument("--test", type=int, help="--test 需要测试的图片数量")
args = parser.parse_args()
config = Config()
if args.bank == None:
parser.print_help()
exit()
conf = config.get(args.bank)
_model = Model()
model = _model.load(conf)
if (model == None):
log.error("模型文件[%s]不存在", _model.model_path())
if args.test == None and args.image == None:
parser.print_help()
exit()
if args.image != None:
image_path = "data/validate/" + args.bank + "/" + args.image
image_data = cv2.imread(image_path)
print("预测图片为：" + image_path)
print("预测结果为：" + predict(image_data, args.image, conf, model))
exit()
