No variables in saved checkpoint

General Discussion
#1

Write checkpoint:

  class Mymodule (tf.Module):
    def __init__ (self):
      super().__init__()
      self.m = tf.Variable (5.0)
      self.b = tf.Variable (5.0)
    def __call__ (self, x):
      return self.m * x + self.b

  m = Mymodule()
  cp = tf.train.Checkpoint (model=m)
  print ("cp:", cp)
  cp.write ("mycheckpoint")
  print ("list_variables:", tf.train.list_variables ("mycheckpoint"))

Output:

cp: <tensorflow.python.training.tracking.util.Checkpoint object at 0x7f95fc917940>
list_variables: [('_CHECKPOINTABLE_OBJECT_GRAPH', []), ('model/b/.ATTRIBUTES/VARIABLE_VALUE', []), ('model/m/.ATTRIBUTES/VARIABLE_VALUE', [])]
#3

Your tf.Module has variables. You can verify with m.variables before saving the checkpoint.
But I don’t think It is available in that train checkpoint call without you have trained the model.

1 Like