Easy way to set pre-trained weights in tensorflow?

Asked Sep 19 '19 at 10:06

Active Sep 19 '19 at 12:44

Viewed 33 times

Is there an easy way to set pretrained weights to a tensorflow layer? (w/o keras)

In keras it's easy as model.set_weights(...) or layer.set_weights(...) - but I'm breaking my head around this for more than a day in tf, and can't seem to find anything that works.

edited Sep 19 '19 at 12:44

stephen_mugisha

asked Sep 19 '19 at 10:06

Maverick Meerkat

5,737
3
47
66

You can use [`tf.assign`](https://www.tensorflow.org/api_docs/python/tf/assign) so assign value for individual variables. – carobnodrvo Sep 19 '19 at 11:42
If I use `tf.layers.dense` how do I know the name of the weight variables? – Maverick Meerkat Sep 19 '19 at 12:10
@DavidRefaeli Where are you copying the data from? – carobnodrvo Sep 19 '19 at 12:35
1

@cookiemonster you are right, the solution there works. Thanks, sorry I didn't find it by my own but I searched for more than a few hours. – Maverick Meerkat Sep 19 '19 at 12:47
@carobnodrvo from a matlab matrix file – Maverick Meerkat Sep 19 '19 at 12:48

Easy way to set pre-trained weights in tensorflow?

0 Answers0