Python 如何在Tensorflow中禁用回溯

我有一个用Hessian的Newton-Rapson方法优化函数的代码

import tensorflow as tf

var = tf.Variable([3., -1., 0., 1.], dtype=tf.float32, name="var")

for i in range(10):

    with tf.GradientTape() as tape1:
        with tf.GradientTape() as tape2:
            func1 = (var[0] + 10*var[1])**2 + (5 *(var[2] - var[3])**2) + (var[1] - (2 * var[2]))**4 + (10*(var[0] - var[3])**4)
            g =  tape1.gradient(func1, var)

        gt = tf.reshape(g, [1, 4])
        h = tape2.jacobian(g, var)
        h_inv = tf.linalg.inv(h)
        dot = tf.matmul( gt, h_inv )
        var_next = var - dot

    var = tf.Variable([var_next[0, 0], var_next[0, 1], var_next[0, 2], var_next[0, 3]])
    print('jacobian at ', i , 'step: ', gt)

当我把它放到循环中时，我得到：

警告：tensorflow:最近6次调用中有6次调用触发了tf.function Retracting

---

我知道TF函数被回溯时会发生这种情况，因为它的参数在形状或数据类型（对于张量）甚至值（Python或np对象或变量）上都会发生变化。在我的例子中，参数更改是正常的。

使用以下解决方法关闭警告

import logging, os

logging.disable(logging.WARNING)
os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"

import tensorflow as tf

var = tf.Variable([3., -1., 0., 1.], dtype=tf.float32, name="var")

for i in range(10):

    with tf.GradientTape() as tape1:
        with tf.GradientTape() as tape2:
            func1 = (var[0] + 10*var[1])**2 + (5 *(var[2] - var[3])**2) + (var[1] - (2 * var[2]))**4 + (10*(var[0] - var[3])**4)
            g =  tape1.gradient(func1, var)

        gt = tf.reshape(g, [1, 4])
        h = tape2.jacobian(g, var)
        h_inv = tf.linalg.inv(h)
        dot = tf.matmul( gt, h_inv )
        var_next = var - dot

    var = tf.Variable([var_next[0, 0], var_next[0, 1], var_next[0, 2], var_next[0, 3]])
    print('jacobian at ', i , 'step: ', gt)

输出

jacobian at  0 step:  tf.Tensor([[ 306.      -144.00002   -2.      -310.     ]], shape=(1, 4), dtype=float32)
jacobian at  1 step:  tf.Tensor([[ 94.81535    -1.1852257   2.3704371 -94.815445 ]], shape=(1, 4), dtype=float32)
jacobian at  2 step:  tf.Tensor([[ 28.09341     -0.35116196   0.7023215  -28.093409  ]], shape=(1, 4), dtype=float32)
jacobian at  3 step:  tf.Tensor([[ 8.324021   -0.1040398   0.20810343 -8.324027  ]], shape=(1, 4), dtype=float32)
jacobian at  4 step:  tf.Tensor([[ 2.4663827  -0.03082165  0.0616576  -2.466382  ]], shape=(1, 4), dtype=float32)
jacobian at  5 step:  tf.Tensor([[ 0.73078156 -0.00913089  0.01827012 -0.7307824 ]], shape=(1, 4), dtype=float32)
jacobian at  6 step:  tf.Tensor([[ 0.21652769 -0.00270531  0.0054124  -0.21652697]], shape=(1, 4), dtype=float32)
jacobian at  7 step:  tf.Tensor([[ 0.06415595 -0.00080431  0.00160385 -0.06415619]], shape=(1, 4), dtype=float32)
jacobian at  8 step:  tf.Tensor([[ 0.01900928 -0.00023716  0.00047522 -0.01900923]], shape=(1, 4), dtype=float32)
jacobian at  9 step:  tf.Tensor([[ 5.6324019e-03 -7.0029724e-05  1.4065549e-04 -5.6322156e-03]], shape=(1, 4), dtype=float32) 

---

我在Tensorflow 2.4.1中没有看到任何警告。您能试用最新的Tensorflow版本并确认吗？谢谢！