预测意外结果的神经网络模型

Question

我想创建一个能够自己玩赛车游戏的 CNN。 我已经用 30.000 个样本训练了模型，我的特征（x_test）是帧，我的标签（y_test）是 w d 键（one-hot 编码） 但我的模型是预测这个而不是预测 三个概率：

 [ 1.  0.  0.]
 [ 1.  0.  0.]
 [ 1.  0.  0.]
 [ 1.  0.  0.]
 [ 1.  0.  0.]
 [ 1.  0.  0.]

是我的模型结构有问题还是我需要更大的数据集？

我的代码是：

1. 培训

       classifier = create_model()
           train_data=np.load('data1/final_data.npy')
           for i in range(52):
               file_name = 'data2/training_data-{}.npy'.format(i)
               train_data_2=np.load(file_name)
               train_data=np.concatenate((train_data,train_data_2))    
           shuffle(train_data)
           x=train_data[:,0]
           y=train_data[:,1]
           train_data=[]
           x=x.tolist()
           x=np.array(x)
           y=y.tolist()
           y=np.array(y)
           classifier.fit(x,y, epochs=5, batch_size=32)
           classifier.save('/output/model.model') 
   

2. 型号

   def create_model():
       classifier = Sequential()
       classifier.add(Conv2D(96,(3,3),input_shape=(120,160,3),activation='relu'))
       classifier.add(MaxPooling2D(pool_size=(2,2)))
       classifier.add(Conv2D(256,(3,3),activation='relu'))
       classifier.add(MaxPooling2D(pool_size=(2,2)))
       classifier.add(Conv2D(384,(3,3),activation='relu'))
       classifier.add(Conv2D(384,(3,3),activation='relu'))  
       classifier.add(Conv2D(256,(3,3),activation='relu'))
       classifier.add(MaxPooling2D(pool_size=(2,2)))
       classifier.add(Flatten())
       classifier.add(Dense(units = 2048, activation = 'relu'))
       classifier.add(Dropout(0.5))
       classifier.add(Dense(units = 2048, activation = 'relu'))
       classifier.add(Dropout(0.5))
       classifier.add(Dense(units = 3, activation = 'softmax'))
       classifier.compile(optimizer = 'rmsprop', loss = 'categorical_crossentropy', 
       metrics = ['accuracy'])  
       return classifier
   

3. 测试员

           def straight():
               ReleaseKey(A)
               ReleaseKey(D)
               PressKey(W)
           def left():
               ReleaseKey(D)
               PressKey(W)
               PressKey(A)
   
           def right():
               ReleaseKey(A)
               PressKey(W)
               PressKey(D)
   
           def main():
               model=load_model('model.model')
   
               for i in list(range(5))[::-1]:
                   print(i+1)
                   time.sleep(1)
   
               paused = False
               while(True):
   
                   if not paused:
                       screen = grab_screen(region=(0,64,640,480))
                       screen = cv2.resize(screen, (160,120))
                       screen = cv2.cvtColor(screen, cv2.COLOR_BGR2RGB)
                       screen = np.expand_dims(screen, axis = 0)
                       prediction = model.predict(screen)[0]
                       print(prediction)
   
                       maxval=max(prediction)
   
                       if prediction[1] == maxval:
           #                straight()
                           print('straigh',maxval)
                       elif prediction[2] == maxval:
           #                right()
                           print('right',maxval)
                       elif prediction[0] == maxval:
           #                left()
                           print('left',maxval)
           #
           #        keys = key_check()
           #
           #        if 'P' in keys:
           #            if paused:
           #                paused = False
           #                time.sleep(1)
           #            else:
           #                paused = True
           #                ReleaseKey(A)
           #                ReleaseKey(W)
           #                ReleaseKey(D)
           #                time.sleep(1)
   
           main()
   

Answer 1

可以通过创建更深层次的神经网络来解决。