Tensorflow2：最小化损失时变量不存在梯度

Question

感谢您在以下方面的帮助。为方便起见，请考虑一个模拟问题，其中子类模型接受输入（None，3,3,1）并返回第四元标签0-3。但是，假设算法首先需要将批处理数据细分为两个互斥的集合（例如，具有标签0-1和标签2-3的那些），然后使用更精细的网络处理每个集合并将其组合回（无，4）形状。根据下面给出的代码，我使用了一个嵌套的交换层来处理批处理中仅包含i）0-1，ii）2-3和iii）0-1和2-3的标签的情况。 。

该模型确实给出了适当的摘要，也可以按预期进行预测，但是由于不存在渐变值，因此我无法为其拟合数据。如果您能提供任何帮助，我将不胜感激。

import tensorflow as tf

class CustomModel(tf.keras.Model):
    def __init__(self, nClasses = 2):
        super(CustomModel, self).__init__()
        self.nClasses = nClasses
        self.Net = tf.keras.Sequential()
        self.Net.add(tf.keras.layers.Conv2D(filters=4, kernel_size=3, activation="relu", padding="same"))
        self.Net.add(tf.keras.layers.Dense(5, activation='relu'))
        self.Net.add(tf.keras.layers.Flatten())
        self.Net.add(tf.keras.layers.Dense(self.nClasses, activation='softmax'))
    def call(self, x):
        y = self.Net(x)
        return y


class CustomModelSwitch(tf.keras.Model):
    def __init__(self):
        super(CustomModelSwitch, self).__init__()
        self.Net0 = CustomModel(2) # this governs intermediate binary lables (0-1)
        self.Net1 = CustomModel(4) # this governs final quaternary lables (0-1-2-3)
        self.Net1.build((None, 3, 3, 1)) # Gives shape errors without this  
    def call(self, x):
        def MyGetSubInput(x,Index,Val) : # Returns a subset of x with the lables "Val"
            y = tf.keras.backend.squeeze(tf.where(tf.keras.backend.equal(Index,Val)),axis=1)
            y = tf.keras.backend.gather(x, y)
            return y                    
        def MyIsSubBatchInputEmpty(x,Argmax,Val):
            SubZeroOne = tf.keras.layers.Lambda(MyGetSubInput,arguments={'Index': Argmax, 'Val': Val})(x)
            return tf.keras.backend.equal(tf.keras.backend.shape(SubZeroOne)[0],0)
        def MyLayer01(x):
            Sub0 = self.Net1(x) 
            Sub0 = tf.keras.layers.Lambda(lambda t : 2*t)(Sub0)  # This is just a mock lambda layer to be replaced with an intended functionality.
            return Sub0 # shape(None,4)
        def MyLayer23(x):
            Sub1 = self.Net1(x) 
            Sub1 = tf.keras.layers.Lambda(lambda t : 5*t)(Sub1) # This is just a mock lambda layer ...
            return Sub1 # shape(None,4)
        def MyLayer0123(x,Argmax):
            Sub0 = tf.keras.layers.Lambda(MyGetSubInput,arguments={'Index': Argmax, 'Val': 0})(x) # this extracts 0 lables
            Sub1 = tf.keras.layers.Lambda(MyGetSubInput,arguments={'Index': Argmax, 'Val': 1})(x) # this extracts 1 lables      

            Sub0 = self.Net1(Sub0) 
            Sub0 = tf.keras.layers.Lambda(lambda t : 2*t)(Sub0)  # This is just a mock lambda layer ...

            Sub1 = self.Net1(Sub1) 
            Sub1 = tf.keras.layers.Lambda(lambda t : 5*t)(Sub1) # This is just a mock lambda layer ...

            Sub01 = tf.keras.layers.Lambda(lambda t: tf.concat(t,axis=0))([Sub0, Sub1])
            return Sub01 # shape(None,4)
        y = self.Net0(x) # this governs intermediate binary lables (0-1)
        Argmax = tf.keras.layers.Lambda(lambda t: tf.keras.backend.cast(tf.keras.backend.argmax(t),'float32'))(y)

        # The following is a nested switch: i) If Sub0 is empty or not and ii) if Sub1 is empty or not.
        SubFinal = tf.keras.layers.Lambda(lambda t: tf.keras.backend.switch(MyIsSubBatchInputEmpty(t,Argmax,0),
        MyLayer23(t),
        tf.keras.backend.switch(MyIsSubBatchInputEmpty(t,Argmax,1),MyLayer01(t), MyLayer0123(t,Argmax))))(x)

        return SubFinal # shape(None,4)
#---------------------------------------------------------------------------------------
def MyDefineData(DataCount):
    import numpy as np
    a = np.empty((DataCount,3,3,1),dtype='uint8')
    b = np.zeros((DataCount),dtype='uint8')
    for i in range(DataCount) :
        a[i,:,:,:] = np.random.randint(low = 0, high = 2, size = 9).reshape((3,3,1))
    from keras.utils import to_categorical
    a = a.astype('float32') / 2
    b[0] = 0
    b[1] = 1
    b[2] = 2
    b[3] = 3
    b[4] = 3
    b[5] = 2    
    b = to_categorical(b,dtype=np.float32)
    return a,b 
inputData,inputLables = MyDefineData(6)

#------------------------------------------------------------------------------------

model = CustomModelSwitch()
model.build((None, 3, 3, 1))

model.summary() 

model.compile(optimizer='adamax', loss=['categorical_crossentropy'], loss_weights=[1], metrics=['accuracy'])
model.fit(inputData, inputLables, epochs=3, batch_size=3,  verbose=1)

<< img src =“ https://image.soinside.com/eyJ1cmwiOiAiaHR0cHM6Ly9pLnN0YWNrLmltZ3VyLmNvbS9wZzFTRy5qcGcifQ==” alt =“在此处输入图像描述”>

Answer 1

我不确定这里是否百分之一百，但在我看来，问题可能出在以下情况：您首先使用一个似乎要训练的网络，以便从x（y = net0 （X））。然后，对结果使用不可微操作（argmax）决定进一步的计算。但是，这意味着net0参数的微小变化可能导致发生完全不同的计算，因此，给定输入x作为其参数的函数的模型结果似乎不是连续的，更不用说是可区分的了。