轻松学习 TFLM Day 5：找出当前模型需要注册哪些 op摘要：本文解决 TFLM 开发中的一个实际问题——拿到.tflite模型后，如何在MicroOpResolver里注册所需的 op。文章先给出常用模型的 op 注册结论，再讲解.tflite中 op 信息的存储结构（operator_codes与operators[].opcode_index），并提供了两种解析脚本（基于flatbuffers官方库和纯标准库手写解析）。最后总结了 builtin/custom op 的区别、MicroMutableOpResolverN中 N 的选取规则，以及排查Missing registration的完整流程，帮助读者快速定位并解决 op 注册问题。Day 2 我们知道 TFLM 执行模型时，会通过MicroOpResolver找到每个算子的 kernel。Day 3 和 Day 4 又讲了怎么写和组织自定义优化 kernel。今天解决一个非常实际的问题：我手里有一个 .tflite 模型，到底要在 resolver 里 Add 哪些 op？这件事很重要。注册少了，运行时报错；注册多了，固件变大。TFLM 的风格是“模型用什么，我就注册什么”。1. 结论先行：常用模型需要注册哪些 op我对当前仓库里的主要.tflite模型做了本地解析。下面是常用模型结果。1.1 hello_world模型：tensorflow/lite/micro/examples/hello_world/models/hello_world_float.tflite tensorflow/lite/micro/examples/hello_world/models/hello_world_int8.tflite python/tflite_micro/sine_float.tflite需要的 op：FULLY_CONNECTED最小 resolver：tflite::MicroMutableOpResolver1resolver;resolver.AddFullyConnected();说明：虽然模型里有 3 个 operator，但都是同一种 op，所以只注册一次FULLY_CONNECTED。1.2 simple_add_model模型：tensorflow/lite/micro/examples/memory_footprint/models/simple_add_model.tflite tensorflow/lite/micro/python/tflite_size/tests/simple_add_model.tflite需要的 op：ADD最小 resolver：tflite::MicroMutableOpResolver1resolver;resolver.AddAdd();1.3 micro_speech_quantized模型：tensorflow/lite/micro/examples/micro_speech/models/micro_speech_quantized.tflite需要的 op：DEPTHWISE_CONV_2D FULLY_CONNECTED RESHAPE SOFTMAX最小 resolver：tflite::MicroMutableOpResolver4resolver;resolver.AddDepthwiseConv2D();resolver.AddFullyConnected();resolver.AddReshape();resolver.AddSoftmax();1.4 micro_speech audio_preprocessor_float模型：tensorflow/lite/micro/examples/micro_speech/models/audio_preprocessor_float.tflite需要的 op：CAST CONCATENATION CUSTOM:SignalEnergy CUSTOM:SignalFftAutoScale CUSTOM:SignalFilterBank CUSTOM:SignalFilterBankLog CUSTOM:SignalFilterBankSpectralSubtraction CUSTOM:SignalFilterBankSquareRoot CUSTOM:SignalPCAN CUSTOM:SignalRfft CUSTOM:SignalWindow MUL RESHAPE STRIDED_SLICE最小 resolver：tflite::MicroMutableOpResolver14resolver;resolver.AddCast();resolver.AddConcatenation();resolver.AddEnergy();resolver.AddFftAutoScale();resolver.AddFilterBank();resolver.AddFilterBankLog();resolver.AddFilterBankSpectralSubtraction();resolver.AddFilterBankSquareRoot();resolver.AddPCAN();resolver.AddRfft();resolver.AddWindow();resolver.AddMul();resolver.AddReshape();resolver.AddStridedSlice();这些Signal*都是 TFLM signal custom op，注册函数在：tensorflow/lite/micro/micro_mutable_op_resolver.h signal/micro/kernels/1.5 micro_speech audio_preprocessor_int8模型：tensorflow/lite/micro/examples/micro_speech/models/audio_preprocessor_int8.tflite需要的 op：ADD CAST CONCATENATION CUSTOM:SignalEnergy CUSTOM:SignalFftAutoScale CUSTOM:SignalFilterBank CUSTOM:SignalFilterBankLog CUSTOM:SignalFilterBankSpectralSubtraction CUSTOM:SignalFilterBankSquareRoot CUSTOM:SignalPCAN CUSTOM:SignalRfft CUSTOM:SignalWindow DIV MAXIMUM MINIMUM MUL RESHAPE STRIDED_SLICE最小 resolver：tflite::MicroMutableOpResolver18resolver;resolver.AddAdd();resolver.AddCast();resolver.AddConcatenation();resolver.AddEnergy();resolver.AddFftAutoScale();resolver.AddFilterBank();resolver.AddFilterBankLog();resolver.AddFilterBankSpectralSubtraction();resolver.AddFilterBankSquareRoot();resolver.AddPCAN();resolver.AddRfft();resolver.AddWindow();resolver.AddDiv();resolver.AddMaximum();resolver.AddMinimum();resolver.AddMul();resolver.AddReshape();resolver.AddStridedSlice();1.6 keyword_scrambled模型：tensorflow/lite/micro/models/keyword_scrambled.tflite tensorflow/lite/micro/models/keyword_scrambled_8bit.tflite需要的 op：FULLY_CONNECTED QUANTIZE SOFTMAX SVDF最小 resolver：tflite::MicroMutableOpResolver4resolver;resolver.AddFullyConnected();resolver.AddQuantize();resolver.AddSoftmax();resolver.AddSvdf();1.7 person_detect模型：tensorflow/lite/micro/models/person_detect.tflite需要的 op：AVERAGE_POOL_2D CONV_2D DEPTHWISE_CONV_2D RESHAPE SOFTMAX最小 resolver：tflite::MicroMutableOpResolver5resolver;resolver.AddAveragePool2D();resolver.AddConv2D();resolver.AddDepthwiseConv2D();resolver.AddReshape();resolver.AddSoftmax();这个模型很适合 AI 芯片后端验证，因为CONV_2D和DEPTHWISE_CONV_2D是常见加速重点。1.8 person_detect_vela模型：tensorflow/lite/micro/models/person_detect_vela.tflite需要的 op：CUSTOM:ethos-u最小 resolver：tflite::MicroMutableOpResolver1resolver;resolver.AddEthosU();说明：这个模型已经被 Vela 处理成 Ethos-U custom op，TFLM 看到的是一个 custom op，而不是普通的CONV_2D、DEPTHWISE_CONV_2D列表。1.9 mnist_lstm模型：tensorflow/lite/micro/examples/mnist_lstm/trained_lstm.tflite tensorflow/lite/micro/examples/mnist_lstm/trained_lstm_int8.tflite需要的 op：FULLY_CONNECTED RESHAPE SOFTMAX UNIDIRECTIONAL_SEQUENCE_LSTM最小 resolver：tflite::MicroMutableOpResolver4resolver;resolver.AddFullyConnected();resolver.AddReshape();resolver.AddSoftmax();resolver.AddUnidirectionalSequenceLSTM();1.10 dtln_noise_suppression模型：tensorflow/lite/micro/examples/dtln/dtln_noise_suppression.tflite需要的 op：FULLY_CONNECTED LOGISTIC UNIDIRECTIONAL_SEQUENCE_LSTM最小 resolver：tflite::MicroMutableOpResolver3res
