tensorflow - creating a variable within tf.variable_scope(name), initialized from another variable's initialized_value -

hey tensorflow community,

i experiencing unexpected naming conventions when using variable_scope in following setup:

with tf.variable_scope("my_scope"):     var = tf.variable(initial_value=other_var.initialized_value()) 

in above, holds

other_var.name = 'outer_scope/my_scope/other_var_name:0' 

i therefore "reusing" same scope @ point in code. intuitively not see issue this, following happens:

var.name = 'outer_scope/my_scope_1/var_name:0' 

so apparently, tf isn't happy "my_scope" , needs append "_1". "outer_scope" remains same, though.

if not initialize "other_var", behaviour not come up.

an explanation appreciated! thx

mat

you might want use tf.get_variable() instead of 'tf.variable`.

with tf.variable_scope('var_scope', reuse=false) var_scope:     var = tf.get_variable('var', [1])     var2 = tf.variable([1], name='var2')     print var.name # var_scope/var:0      print var2.name # var_scope/var2:0  tf.variable_scope('var_scope', reuse=true) var_scope:     var = tf.get_variable('var', [1])     var2 = tf.variable([1], name='var2')     print var.name # var_scope/var:0      print var2.name # var_scope_1/var2:0 

the reason behind think in example, although have "re-entered" variable_scope want, affects variable name scope named name_scope intead of variable_scope might guess. official document here can see that:

when tf.variable_scope("name"), implicitly opens tf.name_scope("name").

name_scope used managing operation names(such add, matmul), because tf.variable operation , operation name "inherited" variables created it, name of name_scope rather variable_scope used prefix.

but if want use tf.variable, can directly use name_scope in with statement:

with tf.name_scope('n_scope') n_scope:     var = tf.variable([1], name='var')     print var.name #n_scope/var_1:0  tf.name_scope(n_scope) n_scope:     var = tf.variable([1], name='var')     print var.name #n_scope/var_1:0 

one thing pay attention should pass argument scope varible captured with statement when want "re-enter" name scope, rather using str scope name:

  tf.name_scope('n_scope') n_scope:       var = tf.variable([1], name='var')       print var.name #n_scope/var_1:0    tf.name_scope('n_scope') n_scope:       var = tf.variable([1], name='var')       print var.name #n_scope_1/var_1:0 

pay attention argument passed tf.name_scope. behavior again described in doc string of name_scope:

the name argument interpreted follows:

1. a string (not ending ‘/’) create new name scope, in name appended prefix of operations created in context. if name has been used before, made unique calling self.unique_name(name).

2. a scope captured g.name_scope(...) scope: statement treated “absolute” name scope, makes possible re-enter existing scopes.

3. a value of none or empty string reset current name scope top-level (empty) name scope.