为什么tf中的optimizer.get_gradients(...)不起作用? [英] Why Does optimizer.get_gradients(...) in tf not Work?

问题描述

我想使用

I want to use optimizer.get_gradients() from tf2.0 in lieu of GradientTape() a simple model as follows

import numpy as np
import tensorflow as tf

class C3BR(tf.keras.Model):
    def __init__(self, filterNum, kSize, strSize, padMode, dFormat='channels_first'):
        super(C3BR, self).__init__()
        if dFormat == 'channels_first':
            self.conAx = 1
        else:
            self.conAx = -1
        self.kSize = (kSize, kSize, kSize)
        self.conv = layers.Conv3D(filters=filterNum, kernel_size=self.kSize, strides=strSize, padding=padMode, data_format=dFormat)
        self.BN = layers.BatchNormalization(axis=self.conAx)
        self.Relu = layers.ReLU()

    def call(self, inputs, ifTrain=False):
        x = self.conv(inputs)
        x= self.BN(x, training=ifTrain)
        outputs = self.Relu(x)
        return outputs

model = C3BR(32, 3, 1, 'valid')
# model.build_model(input_shape=(2, 4, 64, 64, 64))
# model.summary()
curOpt = tf.keras.optimizers.Adam(learning_rate=1e-4)
x = tf.ones((2, 4, 64, 64, 64), dtype=tf.float32)
yTrue =  tf.ones((2, 32, 62, 62, 62), dtype=tf.float32)
yPred = model(x,ifTrain=True)
loss = tf.reduce_mean(yPred-yTrue)
## Why does not it work?
gradients = curOpt.get_gradients(loss, model.trainable_variables)
curOpt.apply_gradients(zip(gradients, model.trainable_variables)) 
gradNorm = tf.linalg.global_norm(gradients)

但是当我运行代码时，出现了一个错误，说C3BR具有无渐变".

But when I ran the code, an error was raised saying C3BR has None gradient.

我不正确地使用了optimizer.get_gradients(...)吗?

Am I using optimizer.get_gradients(...) incorrectly?

推荐答案

在TF2中处理渐变的推荐方法是使用 tf.GradientTape().我无法告诉您上述原因为何不起作用.但是以下工作.

The recommended way to deal with gradients in TF2 is to use tf.GradientTape(). I can't tell you why the above doesn't work. But the following works.

model = C3BR(32, 3, 1, 'valid')
# model.build_model(input_shape=(2, 4, 64, 64, 64))
# model.summary()
curOpt = tf.keras.optimizers.Adam(learning_rate=1e-4)
x = tf.ones((2, 4, 64, 64, 64), dtype=tf.float32)
yTrue =  tf.ones((2, 32, 62, 62, 62), dtype=tf.float32)
## Why does not it work?
with tf.GradientTape() as tape:
  yPred = model(x,ifTrain=True)
  #print(yPred)
  loss = tf.reduce_mean(yPred-yTrue)

gradients = tape.gradient(loss, model.trainable_variables)
print(gradients)

这篇关于为什么tf中的optimizer.get_gradients(...)不起作用?的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！