[TF Lite] How to Convert Tensorflow Model to TF Lite Model

Description

This is the basic guide for converting tensorflow model to tensorflow lite model through TOCO.

1. Choose Tensorflow Model

We can get pre-trained tensorflow model at the offcial document.
In this guide, ssd_mobilenet_v2_coco will be used.

2. Download Tensorflow Model

curl http://download.tensorflow.org/models/object_detection/ssd_mobilenet_v2_coco_2018_03_29.tar.gz | tar xzv -C ${PATH_TO_TF_MODEL}

3. Clone Tensorflow git repository

$ git clone https://github.com/tensorflow/tensorflow.git

4. Convert with TOCO

../tensorflow$ bazel run --config=opt //tensorflow/contrib/lite/toco:toco -- --input_file=${PATH_TO_TF_MODEL}/frozen_inference_graph.pb --output_file=${PATH_TO_TFLITE_MODEL}/ssd_mobilenet_v2_coco.tflite --inference_type=FLOAT --input_shape=1,300,300,3 --input_array=Preprocessor/sub --output_arrays=concat,concat_1

Description of the option

• --input_shape: The dimension of input image tensor
• --input_array: Set the input tensor from tensorflow graph
• --output_arrays: Set the output tensor from tensorflow graph

To set above options in person, you should know the graph structure of the tensorflow model.
If we want to use the whole models, with original in/out tensor, we have to implement the custom operator.
In this guide, we will set input/output tensor in person for not using custom operators.

When bazel build is failed

1. SSL exception
If you confront with below log, it could be related with proxy.

...

ERROR: /home/tensorflow/tensorflow/tools/pip_package/BUILD:100:1: no such package 
'@eigen_archive//': Error downloading [https://bitbucket.org/eigen/eigen/get/default.tar.gz]
to /home/.cache/bazel/_bazel_david/9a6a3e878a69de23dea27378a84b120f/external/eigen_archive/default.tar.gz:
sun.security.validator.ValidatorException: PKIX path building failed:
sun.security.provider.certpath.SunCertPathBuilderException:
unable to find valid certification path to requested target and referenced by
'//tensorflow/tools/pip_package:licenses'.
ERROR: Analysis of target '//tensorflow/tools/pip_package:build_pip_package' failed; build aborted.

...

The simplest way to solve this problem is creating a simple file server with python.

1. get the files in person

$ wget https://bitbucket.org/eigen/eigen/get/default.tar.gz -P ${PATH_TO_STORE}

2. run simple python ftp server at localhost

$ python -m SimpleHTTPServer 8000

3. edit ../tensorflow/tensorflow/workspace.bzl append http://localhost:8000/default.tar.gz at the archive address

...

   tf_http_archive(
       name = "eigen_archive",
       urls = [
           "http://localhost:8000/default.tar.gz",
           ...
       ],
       sha256 = "d956415d784fa4e42b6a2a45c32556d6aec9d0a3d8ef48baee2522ab762556a9",
       strip_prefix = "eigen-eigen-fd6845384b86",
       build_file = clean_dep("//third_party:eigen.BUILD"),
   )

...