TensorFlow Lite

dpinthinker · May 10, 2020, 4:25am

tensorflow/tensorflow/blob/master/tensorflow/lite/python/lite.py#L604


      
          Args:
            optimizers: List of strings that represents the list of optimizers.
          
          
Returns:
            tf.ConfigProto.
          """
          if not optimizers:
            optimizers = []
          # MLIR converter will take care of constant folding instead of grappler.
          if not self.experimental_new_converter:
            optimizers.append("constfold")
          
          
is_only_flex_enabled = (
              set([OpsSet.SELECT_TF_OPS]) == set(self.target_spec.supported_ops))
          if is_only_flex_enabled:
            # The layout optimizer turns NHCW to NCHW. This provides performance
            # optimizations when Flex mode is enabled. However, this is not compatible
            # with builtin ops.
            optimizers.append("layout")
          return _get_grappler_config(optimizers)

因为转的是 graphdef 格式，需要其他命令。

可以在如下网址看下:

https://tensorflow.google.cn/lite/convert/cmdline_examples?hl=zh-cn#convert_a_tensorflow_graphdef_

https://tensorflow.google.cn/lite/convert/cmdline_examples?hl=zh-cn#convert_a_tensorflow_graphdef_for_quantized_inference_

或者，找到 saved_model 模型