我的問(wèn)題是想看到模型迭代中看到圖像的渲染，這個(gè)過(guò)程需要花費(fèi)一些時(shí)間胁镐，所以要分為渲染主線程和數(shù)據(jù)計(jì)算子線程只估。然而又希望加載模型一次，下面是我嘗試成功的一個(gè)例子框架掺喻。但是想要子線程加載模型以及計(jì)算芭届，主線程只管渲染這個(gè)方法沒(méi)有嘗試成功，如果有哪位仁兄成功了感耙，可以@我一下嗎褂乍，非常感謝！
在利用qt designer 設(shè)計(jì)了界面之后即硼，會(huì)生成一個(gè)界面類(lèi)逃片；
主要包括setupUi(self, MainWindow): 和 retranslateUi(self, MainWindow): 兩個(gè)函數(shù)；按照下面方式修改類(lèi)谦絮，就可以像普通類(lèi)一樣用Ui__MainWindow了题诵。
from PyQt5.QtWidgets import QApplication , QMainWindow
class Ui_MainWindow(QMainWindow):
def __init__(self):
super().__init__()
self.setupUi(self)
def setupUi(self,MainWindow):
...
def retranslateUi(self, MainWindow):
...
if __name__ == '__main__':
app = QApplication(sys.argv)
myWin = Ui_MainWindow()
myWin.show()
sys.exit(app.exec_())
在Ui_MainWindow 類(lèi)中增加 tensorflow模型函數(shù)洁仗，因人而異的，下面僅做借鑒性锭；另外在init(self) 函數(shù)中增加loadModel 初始化赠潦；
class Ui_MainWindow(QMainWindow):
def __init__(self):
super().__init__()
self.setupUi(self)
self.loadModel(parse_arguments(sys.argv[1:]))
def loadModel(self,args):
self.phase_train_placeholder = tf.placeholder(tf.bool, name='phase_train')
self.blendshape = tf.Variable(tf.zeros([1, self.hidden_dim]), dtype=tf.float32, name="bs")
# the input for the discrimiator
self.input_Img_placeholder = tf.placeholder(tf.float32, [1, self.discri_ImgWei, self.discri_ImgWei, 3],
name="real_image")
generator = DCGAN()
self.pred = generator.g(self.blendshape, training=self.phase_train_placeholder)
pred_netfeature[args.discriminate_featuremap])
all_vars = tf.trainable_variables()
ftune_vlist = [v for v in all_vars if v.name.startswith('bs')]
self.train_op = NetUtil.train(self.loss, global_step=self.global_step, optimizer=args.optimizer,
learning_rate=self.lr, moving_average_decay=args.moving_average_decay,
update_gradient_vars=ftune_vlist)
# fixed the restore model parameters
var_to_restore = [v for v in all_vars if (v.name.startswith('g') or v.name.startswith('d'))]
gen_saver = tf.train.Saver(var_to_restore)
self.summary_op = tf.summary.merge_all()
self.sess_gen = tf.Session()
self.writer = tf.summary.FileWriter('log/train', self.sess_gen.graph)
# TODO initial the bs by different methods
self.sess_gen.run(tf.global_variables_initializer())
self.sess_gen.run(tf.local_variables_initializer())
self.gen_model = generator_model
gen_saver.restore(self.sess_gen, self.gen_model)
graph = tf.get_default_graph()
print("load the generator model success")
觸發(fā)某個(gè)事件，調(diào)用子線程草冈，傳遞所有模型的參數(shù)到子線程中;下面涉及到從子線程傳遞參數(shù)到主線程和從主線程傳遞參數(shù)到子線程她奥。其中，主線程傳遞參數(shù)到子線程怎棱，只需要在子線程init函數(shù)中傳參就可以了哩俭。從子線程傳遞到主線程就需要信號(hào)和槽機(jī)制，
- 首先 聲明 InfoSignal = QtCore.pyqtSignal(list,list,float,int)
- 在run函數(shù)中self.InfoSignal.emit(img,bsret,loss_var,i+1)
- 子線程傳回信號(hào)處理事件拳恋，在下面代碼中函數(shù) OptimEnd函數(shù)
觸發(fā)事件和收到信號(hào)代碼：
def on_click(self):
print("into on click")
self.pushButton.setDisabled(True)
# 得到總共優(yōu)化步數(shù) 開(kāi)始優(yōu)化,
optimSumStep = self.stepSumspinBox.value()
optimStep = self.stepIdxspinBox.value()
self.lr = self.lrDoubleSpinBox.value()
# 一個(gè)循環(huán)每迭代一次凡资，可視化到面板上 ，input: realImg, 和上一次得到blendshape
self.OptimThread = OptimizerThread(optimSumStep, self.sess_gen, self.RealImg,
self.last_bs, self.blendshape, self.pred,
self.train_op, self.loss, self.input_Img_placeholder,
self.phase_train_placeholder)
self.OptimThread.InfoSignal.connect(self.OptimEnd)
self.OptimThread.start()
def OptimEnd(self,gen_img,last_bs,loss,cnt):
print("loss=",loss)
self.last_bs = last_bs
print(self.last_bs)
self.gen_img = np.array(gen_img)
pix = QtGui.QPixmap(qimage2ndarray.array2qimage(self.gen_img))
self.OptimGraphicsView.setPixmap(QtGui.QPixmap(pix).scaled(580, 640))
self.stepIdxspinBox.setProperty("value", cnt)
self.pushButton.setDisabled(False)
子線程代碼：
class OptimizerThread(QThread):
# 聲明一個(gè)信號(hào),接受返回值 generator_image,bs,loss
InfoSignal = QtCore.pyqtSignal(list,list,float,int)
basic_ImgWei = 256
basic_ImgHei = 256
discri_ImgWei = 224
#構(gòu)造函數(shù)谬运，增加參數(shù)sess, realImage,bs_init
def __init__(self,optimStep,sess,realImg,bs_init,blendshape,pred,train_op,loss,input_Img_placeholder,phase_train_placeholder,parent=None):
super(OptimizerThread,self).__init__(parent)
print("into the optimizer thread ")
self.optimStep = optimStep
self.sess_gen = sess
self.realImg = realImg
self.bs_init = bs_init
self.blendshape = blendshape
self.pred = pred
self.train_op = train_op
self.loss = loss
self.input_Img_placeholder = input_Img_placeholder
self.phase_train_placeholder = phase_train_placeholder
#重寫(xiě)run函數(shù)
def run(self):
print("into thread run")
self.sess_gen.run(tf.assign(self.blendshape, self.bs_init))
print("after blendshape")
for i in range(self.optimStep):
bs_img = imresize(self.realImg, (self.discri_ImgWei, self.discri_ImgWei))
bsimg_tensor = bs_img.reshape(-1, self.discri_ImgWei, self.discri_ImgWei, 3)
gen_img, _, input_bs, loss_var = self.sess_gen.run(
[self.pred, self.train_op, self.blendshape, self.loss],
feed_dict={self.input_Img_placeholder: bsimg_tensor, self.phase_train_placeholder: True})
self.last_bs = input_bs
img = gen_img[0].tolist()
bsret = self.last_bs[0].tolist()
# self.optimStep = self.optimStep+1
self.InfoSignal.emit(img,bsret,loss_var,i+1)
print("thread end")
遺存的問(wèn)題隙赁，以上，我嘗試的方案就成功了梆暖，我的問(wèn)題是需要加載一張圖片伞访，然后做優(yōu)化。當(dāng)我點(diǎn)擊觸發(fā)優(yōu)化事件之后轰驳，再點(diǎn)擊觸發(fā)優(yōu)化事件ui就會(huì)退出厚掷，但是從新加載一張照片之后就不會(huì)產(chǎn)生這樣的情況，即使存在這樣的bug,也可以用级解。如果你發(fā)現(xiàn)如何改進(jìn)冒黑，非常感謝你能通知我，希望對(duì)初學(xué)者有用蠕趁。
