Question

我想创建一个能够独立玩赛车游戏的CNN。我用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和d键（单热编码）但我的模型预测这不是预测三个概率：

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

我的模型结构中存在问题，还是需要更大的数据集？

我的代码是：

训练

    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')

模型

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

测试

        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的神经网络来解决。

神经网络模型预测意外结果

1 个答案: