如何解决tensorflow 2.2.0中的CNN模型拟合问题？

Question

我想用图像数据训练一个 CNN 模型。我有 2 类（面具和不面具）。我通过以下代码导入并保存数据：

data_path='/train/'
categories=os.listdir(data_path)
labels=[i for i in range(len(categories))]
label_dict=dict(zip(categories,labels))
data=[]
target=[]
for category in categories:
    folder_path=os.path.join(data_path,category)
    img_names=os.listdir(folder_path)
    for img_name in img_names:
        img_path=os.path.join(folder_path,img_name)
        img=cv2.imread(img_path)
        try:
            gray=cv2.cvtColor(img,cv2.COLOR_BGR2GRAY) 
            resized=cv2.resize(gray,(500, 500))#dataset
            data.append(resized)
            target.append(label_dict[category])
        except Exception as e:
            print('Exception:',e)
data=np.array(data)/255.0
data=np.reshape(data,(data.shape[0],500, 500,1))
target=np.array(target)
new_target=np_utils.to_categorical(target)
#np.save('data',data)
#np.save('target',new_target)

我这样构建模型：

model=tf.keras.models.Sequential([
    Conv2D(32, 1, activation='relu', input_shape=(500, 500, 1)),
    MaxPooling2D(2,2),
    Conv2D(64, 1, activation='relu'),
    MaxPooling2D(2,2),
    Conv2D(128, 1, padding='same', activation='relu'),
    MaxPooling2D(2,2),
    Flatten(),
    Dropout(0.5), 
    Dense(256, activation='relu'),
    Dense(2, activation='softmax') # dense layer has a shape of 2 as we have only 2 classes 
])
model.compile(loss='binary_crossentropy',optimizer='adam',metrics=['accuracy'])

model.summary 给我以下结果：

________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d (Conv2D)              (None, 500, 500, 32)      64        
_________________________________________________________________
max_pooling2d (MaxPooling2D) (None, 250, 250, 32)      0         
_________________________________________________________________
conv2d_1 (Conv2D)            (None, 250, 250, 64)      2112      
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 125, 125, 64)      0         
_________________________________________________________________
conv2d_2 (Conv2D)            (None, 125, 125, 128)     8320      
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 62, 62, 128)       0         
_________________________________________________________________
flatten (Flatten)            (None, 492032)            0         
_________________________________________________________________
dropout (Dropout)            (None, 492032)            0         
_________________________________________________________________
dense (Dense)                (None, 256)               125960448 
_________________________________________________________________
dense_1 (Dense)              (None, 2)                 514       
=================================================================
Total params: 125,971,458
Trainable params: 125,971,458
Non-trainable params: 0

然后我拟合模型但内核停止。我的拟合代码是：

history=model.fit(data, target, epochs=10, batch_size=128, validation_data=data_val)

我的 tensorflow 版本是 2.2.0。为什么不运行我的模型？

Answer 1

由于线程占用了太多资源，您的内核似乎正在死亡（被杀死）。似乎您通过添加太多连接和可训练参数来制作不必要的复杂模型。事实上，单个密集层实际上负责 99.991% 的所有可训练参数 (125,960,448 / 125,971,458)。

问题是您的计算资源（主要是 RAM）已用完。只是为了给你一个背景，以下是一些最有影响力的基于 CNN 的架构，其中大部分已经在 power GPU 上进行了 DAYS 训练。

LeNet-5 - 60,000 parameters
AlexNet - 60M paramters
VGG-16 - 138M paramters
Inception-v1 - 5M parameters
Inception-v3 - 24M parameters
ResNet-50 - 26M parameters
Xception - 23M parameters
Inception-v4 - 43M parameters
Inception-ResNet-V2 - 56M parameters
ResNeXt-50 - 25M parameters

Your basic 2 CNN stack model - 125M parameters!

---

这是你可以做的 -

flatten (Flatten)            (None, 492032)            0         
_________________________________________________________________
dropout (Dropout)            (None, 492032)            0         
_________________________________________________________________
dense (Dense)                (None, 256)               125960448 <---!!!!
_________________________________________________________________

您正在将 62x62x128 张量展平为 492,000 长度的向量！相反，要么尝试添加更多 CNN，以使前 2 个暗淡更易于管理，并且/或者增加以前 CNN 中内核的大小。

这里的目标是在到达 Dense 层之前拥有一个可管理大小的张量。另外，尽量减少密集层的节点数。

对于初学者尝试这样的事情，你的设备可以在不杀死内核的情况下实际处理的事情，比如使用 68k 参数（你应该更简单，以后会增加复杂性。）

model=tf.keras.models.Sequential([
    Conv2D(32, 3, activation='relu', input_shape=(500, 500, 1)),
    MaxPooling2D(3,3),
    Conv2D(64, 3, activation='relu'),
    MaxPooling2D(3,3),
    Conv2D(128, 3, padding='same', activation='relu'),
    MaxPooling2D(3,3),
    Conv2D(256, 3, padding='same', activation='relu'),
    MaxPooling2D(3,3),
    Flatten(),
    Dropout(0.5), 
    Dense(32, activation='relu'),
    Dense(2, activation='softmax') # dense layer has a shape of 2 as we have only 2 classes 
])

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d_19 (Conv2D)           (None, 498, 498, 32)      320       
_________________________________________________________________
max_pooling2d_18 (MaxPooling (None, 166, 166, 32)      0         
_________________________________________________________________
conv2d_20 (Conv2D)           (None, 164, 164, 64)      18496     
_________________________________________________________________
max_pooling2d_19 (MaxPooling (None, 54, 54, 64)        0         
_________________________________________________________________
conv2d_21 (Conv2D)           (None, 54, 54, 128)       73856     
_________________________________________________________________
max_pooling2d_20 (MaxPooling (None, 18, 18, 128)       0         
_________________________________________________________________
conv2d_22 (Conv2D)           (None, 18, 18, 256)       295168    
_________________________________________________________________
max_pooling2d_21 (MaxPooling (None, 6, 6, 256)         0         
_________________________________________________________________
flatten_5 (Flatten)          (None, 9216)              0         
_________________________________________________________________
dropout_5 (Dropout)          (None, 9216)              0         
_________________________________________________________________
dense_10 (Dense)             (None, 32)                294944    
_________________________________________________________________
dense_11 (Dense)             (None, 2)                 66        
=================================================================
Total params: 682,850
Trainable params: 682,850
Non-trainable params: 0
_________________________________________________________________