How to debug invalid output

The TFLM debugging output tools allow TFLM users to easily debug their models by providing a tool that will compare the intermediate values(output of each OP/Kernel) from a model post invoke between the TFLM and TfLite. As well as a way to compare intermediate values between TFLM x86 implementations and Optimized Implementations.

How to debug TFLM Interpreter output on embedded targets

First you call a C++ binary that takes a TfLite model and returns a file that has random inputs and their corresponding output values for each layer of the model it was provided.

The second is you provide a TfLite model and file outputted by C++ binary above to a python script. The script runs TFLM x86 inference comparison to the expected output.

How to debug TFLM Python Interpreter output

Using a python script mentioned in the section above when only a TfLite model is provided as input, the script generates random input and compares TFLM vs TfLite inference outputs for each layer of the model.

C++ Expected Layer by Layer Output Tool on TFLite Micro

This C++ binary allows you to pass in a TfLite model and returns a flatbuffer file with input and the corresponding output values appended into it that can be passed into a python debugging tool which can compare those golden values vs the x86 TFLM reference kernel implementation.

The C++ Tool/binary will write a debugging file to the path provide in 2nd arg using the tflite_model provided in the 1st arg.

Command bazel/blaze:

 bazel run tensorflow/lite/micro/tools:layer_cc -- \
    </path/to/input_model.tflite>
   </path/to/output.file_name>

How to Build using Makefile :

make -f tensorflow/lite/micro/tools/make/Makefile layer_by_layer_output_tool -j24

Python Layer by Layer Debugging Tool

The Python Tool/Script can first be used to compare TFLM vs Tflite outputs for random inputs by only providing a TfLite file.

TfLite vs TFLM command:

 bazel run tensorflow/lite/micro/tools:layer_by_layer_debugger -- \
    --input_tflite_file=</path/to/my_model.tflite>

The Python Tool/Script can also be used to compare TFLM's python x86 output vs expected output provided by the C++ Tool/binary.

TFLM vs Expected Command:

  bazel run tensorflow/lite/micro/tools:layer_by_layer_debugger -- \
    --input_tflite_file=</path/to/my_model.tflite> \
    --layer_by_layer_data_file=</path/to/my_debug_flatbuffer_file>

Optional Flags:

--print_dump When this flag is set, it will print the TFLM output for each layer that is compared.

--rng Integer random number seed for generating input data for comparisons against TFLite. (Default: 42)