The tools in this directory make it easier to distribute TFLite models to multiple devices over networks with the sole aim of benchmarking latency performance. The intended workflow is as follows:

The stripping tool empties eligible constants from a TFLite flatbuffer to reduce its size.
This lean model can be easily transported to devices over a network.
The reconstitution tool on the device takes in a flatbuffer in memory, and fills in the appropriate buffers with random data.

As an example, see the before/after sizes for MobileNetV1:

Float: 16.9MB -> 12KB
Quantized: 4.3MB -> 17.6 KB

NOTE: This tool only supports single subgraphs for now.

There are two tools in this directory:

1. Stripping buffers out of TFLite flatbuffers

This tool takes in an input flatbuffer, and strips out (or 'empties') the buffers (constant data) for tensors that follow the following guidelines:

Are either of: Float32, Int32, UInt8, Int8
If Int32, the tensor should have a min of 10 elements

The second rule above protects us from invalidating constant data that cannot be randomised (for example, Reshape 'shape' input).

To run the associated script:

bazel run -c opt tensorflow/lite/tools/strip_buffers:strip_buffers_from_fb -- --input_flatbuffer=/input/path.tflite --output_flatbuffer=/output/path.tflite

2. Stripping buffers out of TFLite flatbuffers

The idea here is to reconstitute the lean flatbuffer Model generared in the above step, by filling in random data whereever necessary.

The prototype script can be called as:

bazel run -c opt tensorflow/lite/tools/strip_buffers:reconstitute_buffers_into_fb -- --input_flatbuffer=/input/path.tflite --output_flatbuffer=/output/path.tflite

C++ Library

Both the above tools are present as stripping_lib in this directory, which mutate the flatbuffer(s) in-memory. This ensures we can do the above two steps without touching the filesystem again.