test.py

def load_initial_weights(self, session):
    """
    As the weights from http://www.cs.toronto.edu/~guerzhoy/tf_alexnet/ come 
    as a dict of lists (e.g. weights['conv1'] is a list) and not as dict of 
    dicts (e.g. weights['conv1'] is a dict with keys 'weights' & 'biases') we
    need a special load function
    """
    
    # Load the weights into memory
    weights_dict = np.load(self.WEIGHTS_PATH, encoding = 'bytes').item()
    
    # Loop over all layer names stored in the weights dict
    for op_name in weights_dict:
        
      # Check if the layer is one of the layers that should be reinitialized
      train_bool = True
      if op_name not in self.SKIP_LAYER:
        train_bool = False
        
      with tf.variable_scope(op_name, reuse = True):
            
        # Loop over list of weights/biases and assign them to their corresponding tf variable
        for data in weights_dict[op_name]:
            
        # Biases
        if len(data.shape) == 1:
              
          var = tf.get_variable('biases', trainable = train_bool)
          session.run(var.assign(data))
              
            # Weights
        else:
              
          var = tf.get_variable('weights', trainable = train_bool)
          session.run(var.assign(data))

I think there is an indentation error starting from

        # Biases
        if len(data.shape) == 1:
              
          var = tf.get_variable('biases', trainable = train_bool)
          session.run(var.assign(data))
              
            # Weights
        else:
              
          var = tf.get_variable('weights', trainable = train_bool)
          session.run(var.assign(data))

this whole block needs to be indented.
For the rest, I think it looks good to me.

Yeah. My bad.