Ineffective Python TFLite ImportOption 'save_temp_tfl_input'

[riestmo-nxp]

What happened?

The Python API of the compiler for tflite exposes an import option save_temp_tfl_input for debugging purposes. However, it is unused in the upstream TFLite to TOSA API and consequently the import option is ineffective, e.g. no TFLite representation of the model is saved. I would expect to either remove the import option or to (re-)add the creation of a TFLite representation if the option is set.

Steps to reproduce your issue

1. Follow the TensorFlow Lite integration guide
2. Run the iree_tflite_compile.compile_file method with passing the save_temp_tfl_input option
3. See that no TFLite representation of the model is created

What component(s) does this issue relate to?

Python

Version information

v3.0.0

Additional context

No response

[ScottTodd]

(well this turned into a wall of text, sorry about that 😅)

This is some old code that could use some help with maintenance, patches are welcome.

History

Looking through the git history with git log -S save_temp_tfl_input, that option was part of an early API implementation which was replaced with a newer implementation that included the option but did not use it. See where #7201 deleted bindings/python/iree/compiler/tflite.py.

Relevant source files from history:

• https://github.com/iree-org/iree/blob/6f61506384bfc6176d0143b512229662645a54c0/bindings/python/iree/compiler/tflite.py
• https://github.com/iree-org/iree/blob/6f61506384bfc6176d0143b512229662645a54c0/integrations/tensorflow/iree_tf_compiler/iree-import-tflite-main.cpp

Back then, we had a C++ dependency on TensorFlow and TensorFlow Lite. We eventually broke that hard dependency and replaced it with a Python dependency instead. The current files are here:

• compiler/bindings/python/iree/compiler/tools/tflite.py
• integrations/tensorflow/python_projects/iree_tflite/iree/tools/tflite/scripts/iree_import_tflite/__main__.py

Looks like I updated some TensorFlow code in #13287 to remove unused options like save_temp_tf_input and save_temp_mid_level_input but missed the TensorFlow Lite code.

What to do now

I would expect to either remove the import option or to (re-)add the creation of a TFLite representation if the option is set.

I think at this point the save_temp_tfl_input and save_temp_iree_input options should be merged, e.g. by deleting save_temp_tfl_input

      save_temp_tfl_input: Optionally save the IR that results from importing
        the flatbuffer (prior to any further transformations).
      save_temp_iree_input: Optionally save the IR that is the result of the
        import (ready to be passed to IREE).

All that iree-import-tflite does in iree_import_tflite/__main__.py now is call the upstream tflite_to_tosa_bytecode function from tensorflow.mlir.experimental, which is "ready to be passed to IREE".

Other details

I think the temp file should also be saved as .mlirbc since it is bytecode (not text) here:

iree/compiler/bindings/python/iree/compiler/tools/tflite.py

Lines 96 to 98 in 07f81f0

	# Saving the file, use tfs.
	tfl_iree_input = tfs.alloc_optional(
	"tfl-iree-input.mlir", export_as=options.save_temp_iree_input

If you want to get MLIR text (in the TOSA dialect), you can run the bytecode through iree-opt or iree-ir-tool like so

iree-ir-tool cp tfl-iree-input.mlirbc -o tfl-iree-input.mlir