tf.keras与 TensorFlow混用，trainable=False设置无效

最新推荐文章于 2026-06-07 10:42:45 发布

原创最新推荐文章于 2026-06-07 10:42:45 发布 · 890 阅读

0 ·

本内容遵循CC 4.0 BY-SA版权协议

本文探讨了在 Keras 中设置 Layer 的 trainable=False 似乎不起作用的问题，并提供了通过 variable_scope 管理训练变量的解决方案，强调了在导入模型时明确指定训练变量的必要性。

一个简单的例子

另外一个例子：来自KerasLayer trainable=false seems to have no effect

解决的办法

一个简单的例子

import tensorflow.compat.v1 as tf
from tensorflow.keras.layers import Conv2D

input = tf.ones([5, 5, 5, 5])

with tf.variable_scope('z'):
    z = tf.Variable(tf.zeros(shape=[3,1],dtype=tf.float32),name='z',trainable=False)

with tf.variable_scope('conv'):
    output1 = Conv2D(filters=3,kernel_size=3,trainable=False)(input)
    # output = tf.keras.layers.BatchNormalization(trainable=False)(input,training=True)

print(tf.get_collection(key=tf.GraphKeys.TRAINABLE_VARIABLES))
print(tf.trainable_variables())

打印：

[<tf.Variable 'conv/conv2d/kernel:0' shape=(3, 3, 5, 3) dtype=float32>, <tf.Variable 'conv/conv2d/bias:0' shape=(3,) dtype=float32>]
[<tf.Variable 'conv/conv2d/kernel:0' shape=(3, 3, 5, 3) dtype=float32>, <tf.Variable 'conv/conv2d/bias:0' shape=(3,) dtype=float32>]

另外一个例子：来自KerasLayer trainable=false seems to have no effect

Setting trainable=False seems to have no effect.

Minimal example:

layer = tf.keras.layers.Dense(
      units=1,
      kernel_initializer=tf.keras.initializers.Constant([[1.0]]),
      bias_initializer=tf.keras.initializers.Constant([1.0]),
      trainable=False)
y = layer(tf.constant([[1.0]]))
with tf.Session() as session:
  session.run(tf.global_variables_initializer())
  for var in tf.global_variables():
    print("var: " + str(var) + " trainable: " + str(var.trainable))

results in:

var: <tf.Variable 'dense/kernel:0' shape=(1, 1) dtype=float32> trainable: True 
var: <tf.Variable 'dense/bias:0' shape=(1,) dtype=float32> trainable: True