Is there any solution for failing to load model in tensorflow2.3?

Question

I try to use tf.keras.models.load_model to load saved model in tensorflow 2.3. However, I got the same error in https://github.com/tensorflow/tensorflow/issues/41535

It seems an important function. But this issue is still not solved. Does anyone know if there is any alternative method to implement the same result?

Answer 1

I found an alternative method to load custom model in tensorflow 2.3. You need to do some following changes. I will explain by some code snapshots

• for __init__() of custom model. Before,

  def __init__(self, mask_ratio=0.1, hyperparam=0.1, **kwargs):
      layers = []
      layer_configs = {}
      if 'layers' in kwargs.keys():
          layer_configs = kwargs['layers']
      for config in layer_configs:
          layer = tf.keras.layers.deserialize(config)
          layers.append(layer)
      super(custom_model, self).__init__(layers)  # custom_model is your custom model class
      self.mask_ratio = mask_ratio
      self.hyperparam = hyperparam
      ...
  

  After,

  def __init__(self, mask_ratio=0.1, hyperparam=0.1, **kwargs):
      super(custom_model, self).__init__()  # custom_model is your custom model class
      self.mask_ratio = mask_ratio
      self.hyperparam = hyperparam
      ...
  

• define two functions in your custom model class

  def get_config(self):
      config = {
          'mask_ratio': self.mask_ratio,
          'hyperparam': self.hyperparam
      }
      base_config = super(custom_model, self).get_config()
      return dict(list(config.items()) + list(base_config.items()))
  @classmethod
  def from_config(cls, config):
      #config = cls().get_config()
      return cls(**config)
  

• After finishing training, save model using 'h5' format

  model.save(file_path, save_format='h5')
  

• Finally, load model as following codes,

  model = tf.keras.models.load_model(model_path, compile=False, custom_objects={'custom_model': custom_model})