tensorflow 怎么只训练一部分变量

版权声明：本文为博主原创文章，未经博主允许不得转载。 https://blog.csdn.net/zjucor/article/details/85246368

https://stackoverflow.com/questions/37326002/is-it-possible-to-make-a-trainable-variable-not-trainable/37327561

After looking at the documentation and the code, I was not able to find a way to remove a Variable from the TRAINABLE_VARIABLES.

Here is what happens:

• The first time tf.get_variable('weights', trainable=True) is called, the variable is added to the list of TRAINABLE_VARIABLES.
• The second time you call tf.get_variable('weights', trainable=False), you get the same variable but the argument trainable=False has no effect as the variable is already present in the list of TRAINABLE_VARIABLES (and there is no way to remove it from there)

First solution

When calling the minimize method of the optimizer (see doc.), you can pass a var_list=[...] as argument with the variables you want to optimizer.

For instance, if you want to freeze all the layers of VGG except the last two, you can pass the weights of the last two layers in var_list.

Second solution

You can use a tf.train.Saver() to save variables and restore them later (see this tutorial).

• First you train your entire VGG model with all trainable variables. You save them in a checkpoint file by calling saver.save(sess, "/path/to/dir/model.ckpt").
• Then (in another file) you train the second version with non trainable variables. You load the variables previously stored with saver.restore(sess, "/path/to/dir/model.ckpt").

Optionally, you can decide to save only some of the variables in your checkpoint file. See the doc for more info.