with tf.variable_scope(scope):
      # get size of features from action_spec and observation_spec
      nonspatial_size = 0
      spatial_features = ['feature_minimap', 'feature_screen']
      initially_zero_features = {'cargo': 500, 'multi_select': 500, 'build_queue': 10, 'single_select': 1}
      for feature_name, feature_dim in observation_spec:
        if feature_name not in spatial_features:
          if feature_name == 'available_actions':
            feature_size = len(action_spec.functions)
          elif feature_name in initially_zero_features:
            feature_size = initially_zero_features[feature_name] * feature_dim[1]
          else:
            feature_size = 1
            for dim in feature_dim:
              feature_size *= dim
          nonspatial_size += feature_size
      screen_channels = observation_spec['screen'][0]
      minimap_channels = observation_spec['minimap'][0]


Old data type for observation_spec was dict-based but I don't know the exact arrangement without reading through the old changelogs.

New data type for observation_spec is a tuple like this:

<class 'tuple'>: ({'action_result': (0,), 'alerts': (0,), 'available_actions': (0,), 'build_queue': (0, 7), 'cargo': (0, 7), 'cargo_slots_available': (1,), 'control_groups': (10, 2), 'game_loop': (1,), 'last_actions': (0,), 'multi_select': (0, 7), 'player': (11,), 'score_cumulative': (13,), 'single_select': (0, 7), 'feature_screen': (17, 84, 84), 'feature_minimap': (7, 64, 64), 'feature_units': (0, 26)},)

Data in the tuple observation_spec is addressable by [0].action_result[0]