Restore vgg16 network in tensorflow - r

Question

This one has been giving me a headache for quite some time now, even though it seems to be very basic.
I have the vgg16 network downloaded as a .cpkt
(from https://github.com/tensorflow/models/blob/master/slim/README.md#Pretrained)
Now what I want to do is loading for example the tensor of the first convolution layer of this network as an array in R.
I tried
restorer = tf$train$Saver()
sess = tf$Session()
restorer$restore(sess, "/home/beheerder/R/vgg_16.ckpt")
But then I do not see any variables apearing in my enviroment.
I'm working in R, but an awnser in Python is OK as well, as I can probably translate it to R.

Answer 1

Saver takes the variables to restore in constructor. In other words, you have to create the variables before you can restore them. Here is the example from Saver's doc:
v1 = tf.Variable(..., name='v1')
v2 = tf.Variable(..., name='v2')
# Pass the variables as a dict:
saver = tf.train.Saver({'v1': v1, 'v2': v2})
# Or pass them as a list.
saver = tf.train.Saver([v1, v2])
If you were to run the first line of your code in python you would get:
In [1]: import tensorflow as tf
In [2]: saver = tf.train.Saver()
---------------------------------------------------------------------------
ValueError Traceback (most recent call last)
<ipython-input-2-18da33d742f9> in <module>()
----> 1 saver = tf.train.Saver()
/usr/local/lib/python2.7/dist-packages/tensorflow/python/training/saver.pyc in __init__(self, var_list, reshape, sharded, max_to_keep, keep_checkpoint_every_n_hours, name, restore_sequentially, saver_def, builder, defer_build, allow_empty, write_version, pad_step_number)
1054 self._pad_step_number = pad_step_number
1055 if not defer_build:
-> 1056 self.build()
1057 if self.saver_def:
1058 self._check_saver_def()
/usr/local/lib/python2.7/dist-packages/tensorflow/python/training/saver.pyc in build(self)
1075 return
1076 else:
-> 1077 raise ValueError("No variables to save")
1078 self._is_empty = False
1079 self.saver_def = self._builder.build(
ValueError: No variables to save
You can see how model variables are created before being restored in the 20 lines starting from https://github.com/tensorflow/models/blob/master/slim/train_image_classifier.py#L338
This code gets executed if you make a call to train_image_classifier.py similar to the flower example in https://github.com/tensorflow/models/blob/master/slim/README.md#fine-tuning-a-model-from-an-existing-checkpoint