嗨，我转换了一个“resnet-like”.tflite模型（由int 8量化），发现与cpu相比，它在六边形上有一定的精度损失，然后我打印了输出概率，它们确实不同。

我曾经发现了你们提供的“inference_diff”工具（https://github.com/tensorflow/tensorflow/tree/master/tensorflow/lite/tools/evaluation/tasks/inference_diff），现在用它进行测试，结果如下：

./run_eval --model_file=resnet_quantized.tflite --delegate=hexagon                                                                                             
INFO: Initialized TensorFlow Lite runtime.
loaded libcdsprpc.so
Hexagon delegate created.
INFO: TfLiteHexagonDelegate delegate: 34 nodes delegated out of 36 nodes with 1 partitions.
VERBOSE: Replacing 34 node(s) with delegate (TfLiteHexagonDelegate) node, yielding 3 partitions.
INFO: Created TensorFlow Lite XNNPACK delegate for CPU.
native : lite/tools/evaluation/stages/inference_profiler_stage.cc:78 Test interpreter has been initialized.
native : lite/tools/evaluation/stages/tflite_inference_stage.cc:146 
native : lite/tools/evaluation/stages/inference_profiler_stage.cc:92 Reference interpreter (1 thread on CPU) has been initialized.
Num evaluation runs: 50
Reference run latency: avg=240374(us), std_dev=142248(us)
Test run latency: avg=19779(us), std_dev=5008(us)
OutputDiff[0]: avg_error=0.000274915, std_dev=4.89316e-05

看起来差别不大，但一般来说，量化模型的输出概率大多为0，所以这可能没有多大意义。然后我删除模型末尾的softmax操作，并重试：

./run_eval --model_file=resnet_quantized_without_softmax.tflite --delegate=hexagon                                                                             
INFO: Initialized TensorFlow Lite runtime.
loaded libcdsprpc.so
Hexagon delegate created.
INFO: TfLiteHexagonDelegate delegate: 33 nodes delegated out of 35 nodes with 1 partitions.
VERBOSE: Replacing 33 node(s) with delegate (TfLiteHexagonDelegate) node, yielding 3 partitions.
INFO: Created TensorFlow Lite XNNPACK delegate for CPU.
native : lite/tools/evaluation/stages/inference_profiler_stage.cc:78 Test interpreter has been initialized.
native : lite/tools/evaluation/stages/tflite_inference_stage.cc:146 
native : lite/tools/evaluation/stages/inference_profiler_stage.cc:92 Reference interpreter (1 thread on CPU) has been initialized.
Num evaluation runs: 50
Reference run latency: avg=234793(us), std_dev=146766(us)
Test run latency: avg=19949.9(us), std_dev=5333(us)
OutputDiff[0]: avg_error=2.66955, std_dev=0.73623

现在差别更大了，我想知道这些是否正常，模型的准确度损失大概在什么水平，我上传了这两个模型，分别命名为resnet_quantized. tflite. tar. gz和resnet_quantized_without_softmax. tflite。
谢谢你
resnet_quantized.tflite.tar.gz
resnet_quantized_without_softmax.tflite.tar.gz