从相同检查点提取的经过重新训练的Tflite / Pb模型给出不同的结果

Question

After I retrained pre-trained ssd mobilenet v1 model with my own image dataset using object_detection\\model_main.py script, I exported both .pb freeze graph (with export_inference_graph.py script)

python models\research\object_detection\export_inference_graph.py 
--input_type image_tensor 
--input_shape=1,300,300,3 
--pipeline_config_path ssd_mobilenet_v1_test.config 
--trained_checkpoint_prefix training/model.ckpt 
--output_directory export\freeze\

and .tflite graph (with export_tflite_ssd_graph.py script and tflite_convert).

python models\research\object_detection\export_tflite_ssd_graph.py 
--input_type image_tensor 
--pipeline_config_path ssd_mobilenet_v1_test.config 
--trained_checkpoint_prefix training/model.ckpt 
--output_directory export\tflite\ 
--max_detections 16 
--add_postprocessing_op=true

tflite_convert 
--output_file=export\tflite\model.tflite 
--graph_def_file=export\tflite\tflite_graph.pb 
--input_shapes=1,300,300,3 
--input_arrays=normalized_input_image_tensor 
--output_arrays=TFLite_Detection_PostProcess,TFLite_Detection_PostProcess:1,TFLite_Detection_PostProcess:2,TFLite_Detection_PostProcess:3 
--inference_type=QUANTIZED_UINT8 
--mean_values=128 
--std_dev_values=128 
--default_ranges_min=0 
--default_ranges_max=6 
--allow_custom_ops

Pb graph seems to work just fine, but the tflite one false detect everything on android, so I get 16 out of 16 possible detections whatever image I pass to it, even image filled with black colour (I test it on android device. It works well with pre-trained model).

Changing convert options like disabling/enabling quantizing, image std/mean didn't change anything. I also compared my tflite graph to example mobilenet graph and they look pretty similar. Any ideas what can cause that problem?

(windows 10/cuda 9.0/cudnn 7.0/tf-nightly-gpu/models-master)

Answer 1

The output tensors from the tflite model appear to return some extreme values (ex: 5e35 or -3e34). Since some of these score values are greater than 1, it counts as a detection.

My solution, replace all values greater than a limit (I did 1e5) with 0. (Python was faster.)

tensor[tensor > 1e5] = 0

It is weird this doesn't happen with the example detector.tflite or an exported frozen inference graph. There must be a proper way for exporting tflite models.