TF2.5: tensorflow:Gradients 在最小化损失时不存在变量 渐变胶带，keras

如何解决TF2.5: tensorflow:Gradients 在最小化损失时不存在变量 渐变胶带，keras

我正在努力建立一个深度生存模型，并从一个简单的比例风险模型开始，混合使用 keras 和非 keras tensorflow。我的问题是 GradientTape 似乎没有跟踪我使用 tf.Variable 定义的参数，只有 Keras 中的参数。

这是我的模型：

import numpy as np
import tensorflow as tf
from tensorflow.keras.models import Model
from tensorflow.keras.layers import Input,Conv1D

class phmodel(tf.Module):
    def __init__(self,lr = .0001,timesteps = 100,**kwargs):
        super().__init__(**kwargs)
        self.beta = tf.Variable(0.0) # initialize to 0 = log(1)
        self.timesteps = timesteps
        self.lr = lr
        self.makemodel()
        self.opt = tf.keras.optimizers.SGD(learning_rate = lr)

    def makemodel(self):
        inp = Input((self.timesteps,1,))
        lay = Conv1D(filters = 1,kernel_size = 1,use_bias = False)(inp) # this is equivalent to XB -- it applies the same weight to each of the 100 timesteps
        elay = tf.math.exp(lay)
        beta = tf.math.exp(self.beta) #re-exponentiate:  it's trained in logs.  This line is to make the code less confusing
        baseline_hazard = tf.expand_dims(1-tf.math.exp(-beta*tf.range(0,self.timesteps,delta=1,dtype='float32')),-1)
        cumulative_hazard = tf.math.cumsum(baseline_hazard*elay,axis = 1)
        out = 1-tf.math.exp(-cumulative_hazard)
        self.model = Model(inp,out)


    def __call__(self,x):
        return self.model(x)
       
    def train_step(self,x,y,w = None,verbose = False):
        with tf.GradientTape() as g:
            yhat = self.model(x)
            loss = tf.keras.losses.BinaryCrossentropy()(y,yhat,sample_weight = w)
            grad = g.gradient(loss,self.model.trainable_variables)
        self.opt.apply_gradients(zip(grad,self.model.trainable_variables))
        print(loss) if verbose is True else None
        

看起来像这样：

model = phmodel(100)
model.model.summary()

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_4 (InputLayer)         [(None,100,1)]          0         
_________________________________________________________________
conv1d_3 (Conv1D)            (None,1)            1         
_________________________________________________________________
tf.math.exp_5 (TFOpLambda)   (None,1)            0         
_________________________________________________________________
tf.math.multiply_2 (TFOpLamb (None,1)            0         
_________________________________________________________________
tf.math.cumsum_2 (TFOpLambda (None,1)            0         
_________________________________________________________________
tf.math.negative_2 (TFOpLamb (None,1)            0         
_________________________________________________________________
tf.math.exp_6 (TFOpLambda)   (None,1)            0         
_________________________________________________________________
tf.math.subtract_2 (TFOpLamb (None,1)            0         
=================================================================
Total params: 1
Trainable params: 1
Non-trainable params: 0
_________________________________________________________________

定义一些虚拟数据：

xtr = np.random.normal(size = 100).reshape(1,1)
ytr = np.concatenate([np.zeros(50),np.ones(50)]).reshape(1,1)

前传：

yhat = model(xtr)

看看可训练的变量：

model.trainable_variables
(<tf.Variable 'Variable:0' shape=() dtype=float32,numpy=0.0>,<tf.Variable 'conv1d_4/kernel:0' shape=(1,1) dtype=float32,numpy=array([[[1.2811402]]],dtype=float32)>)

损失似乎是有序的：

loss = tf.keras.losses.BinaryCrossentropy()(ytr,sample_weight = None)
loss
<tf.Tensor: shape=(),dtype=float32,numpy=6.711515>

计算梯度：

with tf.GradientTape() as g:
    yhat = model(xtr)
    loss = tf.keras.losses.BinaryCrossentropy()(ytr,sample_weight = None)
    grad = g.gradient(loss,model.trainable_variables)
grad
(None,<tf.Tensor: shape=(1,1),numpy=array([[[0.51390326]]],dtype=float32)>)

问题来了。 Variable:0 没有渐变。是什么赋予了？当我去应用渐变时，我收到警告、问题和NaNs

model.opt.apply_gradients(zip(grad,model.trainable_variables))
WARNING:tensorflow:Gradients do not exist for variables ['Variable:0'] when minimizing the loss.

如果我运行训练循环超过几个步骤，则卷积层上的梯度开始为 NaN。

(<tf.Variable 'Variable:0' shape=() dtype=float32,numpy=-6.2>,<tf.Variable 'conv1d_12/kernel:0' shape=(1,numpy=array([[[nan]]],dtype=float32)>)

是什么？

---

编辑：我通过从 makemodel 方法中取出一堆东西并将其放入 __call__ 中设法解决了这个问题。我不知道这为什么有效。也许它会帮助一些未来的谷歌员工？我很乐意给任何能解释为什么这有效的人打 15 分，因为我没有任何线索。

class phmodel(tf.Module):
    def __init__(self,**kwargs):
        super().__init__(**kwargs)
        self.beta = tf.Variable(-10.2) # initialize to -3.2 = log(.04)
        self.timesteps = timesteps
        self.timerange = tf.range(0,dtype='float32')
        self.lr = lr
        self.makemodel()
        self.opt = tf.keras.optimizers.SGD(learning_rate = lr)

       
    def makemodel(self):
        inp = Input((self.timesteps,use_bias = False)(inp) # this is equivalent to XB -- it applies the same weight to each of the 100 timesteps
        self.model = Model(inp,lay)


    def __call__(self,x):
        mout = self.model(x)
        elay = tf.math.exp(mout)
        beta = tf.math.exp(self.beta) #re-exponentiate:  it's trained in logs.  This line is to make the code less confusing
        baseline_hazard = tf.expand_dims(1-tf.math.exp(tf.multiply(-beta,self.timerange)),-1)
        cumulative_hazard = tf.math.cumsum(tf.multiply(baseline_hazard,elay),axis = 1)
        out = 1-tf.math.exp(-cumulative_hazard)
        return out
       
    def train_step(self,self.model.trainable_variables))
        print(loss) if verbose is True else None