Как сохранить модель keras?

from tensorflow.keras.models import load_model
        model = load_model("model.h5")

class KKK():
    
    def model(self, input_shape):
        seq = Sequential()
        seq.add(Conv2D(6, 3, 3, activation="relu", input_shape=input_shape, padding='valid', data_format="channels_last"))
        seq.add(MaxPooling2D(pool_size=(2, 2)))
        seq.add(Dropout(0.25))
        seq.add(Conv2D(12, 3, 3, activation="relu", padding='valid', data_format="channels_last"))
        seq.add(MaxPooling2D(pool_size=(2, 2), data_format='channels_last'))
        seq.add(Dropout(0.25))
        seq.add(Flatten())
        seq.add(Dense(128, activation='relu'))
        seq.add(Dropout(0.1))
        seq.add(Dense(50, activation='relu'))
        return seq
 
    def euclidean_distance(self, vects):
        x, y = vects
        return K.sqrt(K.sum(K.square(x - y), axis=1, keepdims=True))
 
    def eucl_dist_output_shape(self, shapes):
        shape1, shape2 = shapes
        return (shape1[0], 1)
 
    def contrastive_loss(self, y_true, y_pred):
        margin = 1
        return K.mean(y_true * K.square(y_pred) + (1 - y_true) * K.square(K.maximum(margin - y_pred, 0)))
 
    def train(self):
        X, Y = self.get_data()
        X.shape = (20000, 2, 56, 46, 1)
        x_train, x_test, y_train, y_test = train_test_split(X, Y, test_size=.25)
        input_dim = x_train.shape[2:]
        img_a = Input(shape=input_dim)
        img_b = Input(shape=input_dim)
        base_network = self.model(input_dim)
        feat_vecs_a = base_network(img_a)
        feat_vecs_b = base_network(img_b)
        distance = Lambda(self.euclidean_distance, output_shape=self.eucl_dist_output_shape)([feat_vecs_a, feat_vecs_b])
        rms = RMSprop()
        model = Model(inputs=[img_a, img_b], outputs=distance)
        model.compile(loss=self.contrastive_loss, optimizer=rms)
        model.save("model.h5")
 
from tensorflow.keras.models import load_model
model = load_model("model.h5")

module_objects = module_objects or {}

cls = module_objects.get(class_name)
if cls is None:
    raise ValueError('Unknown ' + printable_module_name + ': ' + class_name)

elif isinstance(identifier, six.string_types):
    function_name = identifier
    if custom_objects and function_name in custom_objects:
      fn = custom_objects.get(function_name)
    elif function_name in _GLOBAL_CUSTOM_OBJECTS:
      fn = _GLOBAL_CUSTOM_OBJECTS[function_name]
    else:
      fn = module_objects.get(function_name)
      if fn is None:
        raise ValueError('Unknown ' + printable_module_name + ':' +
                         function_name)

elif isinstance(identifier, six.string_types):
    object_name = identifier
    if custom_objects and object_name in custom_objects:
      obj = custom_objects.get(object_name)
    elif object_name in _GLOBAL_CUSTOM_OBJECTS:
      obj = _GLOBAL_CUSTOM_OBJECTS[object_name]
    else:
      obj = module_objects.get(object_name)
      if obj is None:
        raise ValueError('Unknown ' + printable_module_name + ':' + object_name)
    # Classes passed by name are instantiated with no args, functions are
    # returned as-is.
    if tf_inspect.isclass(obj):
      return obj()
    return obj
  else:
    raise ValueError('Could not interpret serialized ' + printable_module_name +
                     ': ' + identifier)

else:
    module_objects = module_objects or {}
    cls = module_objects.get(class_name)
    if cls is None:
      raise ValueError('Unknown ' + printable_module_name + ': ' + class_name)
  return (cls, config['config'])

else:
      module_objects = module_objects or {}
      obj = module_objects.get(object_name)
      if obj is None:
        raise ValueError('Unknown ' + printable_module_name + ':' + object_name)

>>> module_objects = None or {}
>>> module_objects.get
<built-in method get of dict object at 0x0000000003025040>
>>>

distance = Lambda(self.euclidean_distance, output_shape=self.eucl_dist_output_shape)([feat_vecs_a, feat_vecs_b])

def eucl_dist_output_shape(self, shapes):
        shape1, shape2 = shapes
        return (shape1[0], 1)

>>> tensorflow.__version__
'2.1.0'
>>>