optimistic restore

optimistic restore

def optimistic_restore(session, save_file, variable_scope=''):
    '''
    A Caffe-style restore that loads in variables
    if they exist in both the checkpoint file and the current graph.
    Call this after running the global init op.
    By DanielGordon10 on December 27, 2016
    https://github.com/tensorflow/tensorflow/issues/312
    With RalphMao tweak.

    bpugh, July 21, 2017: Added a variable_scope so that a network can be
    loaded within a tf.variable_scope() and still have weights restored.
    '''
    reader = tf.train.NewCheckpointReader(save_file)
    saved_shapes = reader.get_variable_to_shape_map()
    if variable_scope is '':
        saved_shapes_scoped = saved_shapes
        offset = 0
    else:
        saved_shapes_scoped = [variable_scope + '/' + x for x in saved_shapes]
        offset = len(variable_scope) + 1

    var_names = []
    for var in tf.global_variables():
        search_term = var.name.split(':')[0]
        if search_term in saved_shapes_scoped:
            var_names.append((var.name.split(':')[0], var.name.split(':')[0][offset:]))

    name2var = dict(zip(map(lambda x:x.name.split(':')[0],
            tf.global_variables()), tf.global_variables()))
    restore_variables = []
    with tf.variable_scope('', reuse=True):
        for var_name, saved_var_name in var_names:
            try:
                curr_var = name2var[var_name]
                var_shape = curr_var.get_shape().as_list()
                if var_shape == saved_shapes[saved_var_name]:
                    found_variable = tf.get_variable(var_name)
                    restore_variables.append(found_variable.assign(reader.get_tensor(saved_var_name)))
            except:
                print("{} couldn't be loaded.".format(saved_var_name))
    session.run(restore_variables)