.. include:: /keyword.rst

===========================
Example: Pruning the Model
===========================

This page provides a example on model pruning if the model contains operators or data types not supported when doing the model compilation.
For more discussions about Genio AI, refer to the `Genio Community <https://genio-community.mediatek.com/c/neuropilot/23>`_.

.. _Unsupported Operations in DLA Conversion (Optional):

Unsupported Operations in DLA Conversion 
========================================

If ``ncc-tflite`` reports *unsupported operation* errors while converting a TFLite model to DLA format, the model may need pruning or structural adjustments.

In many cases, the solution is to:

* Identify unsupported operations, including their parameters and tensor shapes.
* Modify or prune those operations.
* Regenerate the TFLite model and re-run the conversion.

Example Error Message
---------------------

The following example shows a typical error when converting a TFLite model to DLA format:

.. prompt:: bash # auto

   # ncc-tflite --arch=mdla3.0 yolox_s_quant.tflite

.. code-block:: bash

   OP[2]: STRIDED_SLICE
    ├ MDLA: Last stride size should be equal to 1
    ├ EDMA: unsupported operation
    ├ EDMA: unsupported operation
   OP[3]: STRIDED_SLICE
    ├ MDLA: Last stride size should be equal to 1
    ├ EDMA: unsupported operation
    ├ EDMA: unsupported operation
   OP[4]: STRIDED_SLICE
    ├ MDLA: Last stride size should be equal to 1
    ├ EDMA: unsupported operation
    ├ EDMA: unsupported operation
   OP[5]: STRIDED_SLICE
    ├ MDLA: Last stride size should be equal to 1
    ├ EDMA: unsupported operation
    ├ EDMA: unsupported operation
   ERROR: Cannot find an execution plan because of unsupported operations
   ERROR: Fail to compile yolox_s_quant.tflite

.. note::

   In this example:

   * The ``STRIDED_SLICE`` operation appears multiple times (OP[2], OP[3], OP[4], OP[5]), and each instance fails for the same reason.
   * On MDLA, the last stride does not meet the requirement **“Last stride size must be equal to 1”**, so the operation is unsupported.
   * On EDMA, ``STRIDED_SLICE`` is not supported at all.
   * Because the converter cannot find a valid execution plan, the compilation fails.

   To resolve this, modify or prune the unsupported ``STRIDED_SLICE`` operations, or adjust stride settings so that they satisfy the MDLA requirements, and then retry DLA conversion.

1. Inspect the Model
--------------------

* Use a visualization tool such as `Netron <https://netron.app/>`_ to inspect:

  * Input and output tensors.
  * Internal layers, including operations that ``ncc-tflite`` reports as unsupported.
  * Tensor dimensions and data types.

* Focus on operations listed in the ``ncc-tflite`` error log (in the example above, ``STRIDED_SLICE``).

.. figure:: /_asset/yoloX-s_unsupported_operation.png
   :alt: yoloX-s unsupported operation in Netron
   :align: center
   :width: 50%

.. raw:: html

   <br><br>

2. Prune or Rewrite the Model
------------------------------

After locating unsupported operations, prune or reshape the model so it becomes compatible with DLA.

One option is to use the ``TFLiteEditor`` class from ``mtk_converter`` to export a subgraph that excludes unsupported operations or simplifies the network structure.

Example script:

.. code-block:: python

   import mtk_converter

   editor = mtk_converter.TFLiteEditor("yolox_s_quant.tflite")
   output_file = "yolox_s_quant_7-303.tflite"
   input_names = ["input.32"]
   output_names = ["1360"]

   _ = editor.export(
       output_file=output_file,
       input_names=input_names,
       output_names=output_names,
   )

This example exports a subgraph whose input tensor is ``input.32`` and output tensor is ``1360``.

.. note::

   The values of ``input_names`` and ``output_names`` are specific to this model.
   Always determine the correct tensor names for your own model by checking:

   * The TFLite graph in Netron, and
   * The ``ncc-tflite --show-io-info`` output, where applicable.

.. figure:: /_asset/yoloX-s_input_name.png
   :alt: Example of input tensor name in Netron
   :align: center
   :width: 80%

.. raw:: html

   <br><br>

.. figure:: /_asset/yoloX-s_output_name.png
   :alt: Example of output tensor name in Netron
   :align: center
   :width: 80%

.. raw:: html

   <br><br>

3. Validate the Pruned Model
----------------------------

After pruning or rewriting the model:

* Run inference with the pruned TFLite model.
* Compare results against the original model on a small set of representative inputs.
* Confirm that accuracy degradation is acceptable for your use case.

Only after this validation step is complete should the pruned model be used to generate the DLA file with ``ncc-tflite``.


Supported Model Formats for ``ncc-tflite``
==========================================

The ``ncc-tflite`` tool accepts **only** TensorFlow Lite models as input.

Supported format
----------------

* ``.tflite`` (TensorFlow Lite FlatBuffer format)

If the original model is created in another framework, convert it to TFLite before using ``ncc-tflite``.
Common conversion flows include:

* PyTorch → ONNX → TensorFlow → TFLite
* TensorFlow SavedModel → TFLite
* ONNX → TensorFlow → TFLite

For official guidance and examples, refer to the TensorFlow and framework‑specific documentation.

Verifying Supported Operators
-----------------------------

To check whether the operators in your TFLite model are supported by ``ncc-tflite``, use:

.. prompt:: bash # auto

   # ./ncc-tflite --show-builtin-ops
   # ./ncc-tflite --show-mtkext-ops

The output lists:

* Built‑in TFLite operations supported by ``ncc-tflite``.
* MediaTek extension operations (MTKEXT) supported by ``ncc-tflite``.

Compare these lists with the operators used in your model (as shown by tools such as Netron or ``ncc-tflite --show-tflite``) to identify potential incompatibilities before starting a full conversion.