在 Windows 上从 .pb 转换为 .tflite 时出错

Question

I have a frozen tensorflow object detection model frozen_inference_graph.pb and I need to convert it in .tflite format to use it in an android app.

I am using tflite_convert as suggested here, https://codelabs.developers.google.com/codelabs/tensorflow-for-poets-2-tflite/#2

tflite_convert 
--graph_def_file=frozen_inference_graph.pb 
--output_file=new_graph.tflite 
--input_format=TENSORFLOW_GRAPHDEF 
--output_format=TFLITE 
--input_shape=1,224,224,3 
--input_array=image_tensor 
--output_array=detection_boxes,detection_scores,detection_classes,num_detections 
--inference_type=FLOAT 
--input_data_type=FLOAT

This is the error that I'm getting:

Check failed: array.data_type == array.final_data_type Array "image_tensor" 
has mis-matching actual and final data types (data_type=uint8, final_data_type=float).
Fatal Python error: Aborted

Edit: I read the tflite docs and it was mentioned that, only models frozen using freeze.py can be converted using tflite_convert . But I used export_inference_graph.py to get frozen_inference_graph.pb . Is there any other way to convert an object detection model to tflite . I am using the model ssd_mobilenet_v1_coco_11_06_2017 .

Answer 1

The convertor can change input and/or output types, but only one side: if the model has float input tensor, you can create a tflite model with quantized (uint8) input , but not vice versa.

Remove --input_data_type=FLOAT from your command, and it should work.