TensorFlow.js Graph Model Converter

The purpose of this library is to import TFJS graph models into Tensorflow. This allows you to use TensorFlow.js models with Python in case you don't have access to the original formats or the models have been created in TFJS.

Disclaimer

I'm neither a Python developer, nor do I know TensorFlow or TensorFlow.js. I created this package solely because I ran into an issue when trying to convert a pretrained TensorFlow.js model into a different format. I didn't have access to the pretrained original TF model and didn't have the resources to train it myself. I soon learned that I'm not alone with this issue so I sat down and wrote this little library.

If you find any part of the code to be non-idiomatic or know of a simpler way to achieve certain things, feel free to let me know, since I'm a beginner in both Python and especially TensorFlow (used it for the very first time in this very project).

Prerequisites

• tensorflow 2.1+
• tensorflowjs 1.5.2+

Compatibility

The converter has been tested with tensorflowjs v3.13.0, tensorflow v2.8.0 and Python 3.9.10.

Installation

pip install tfjs-graph-converter

Usage

After the installation, you can run the packaged tfjs_graph_converter binary for quick and easy model conversion.

Positional Arguments

Options

Compatibility Modes

Models are converted to optmimised native Tensorflow operators by default. This can cause problems if the converted model is subseuently converted to another format (ONNX, TFLite, older TFJS versions, etc.) The --compat_mode option can be used to avoid incompatible native operations such as fused convolutions. Available options are:

Advanced Options

These options are intended for advanced users who are familiar with the details of TensorFlow and TensorFlow Serving.

Specifying --outputs can be useful for multi-head models to select the default output for the main signature. The CLI only handles the default signature of the model. Multiple signatures can be created using the API.

The method name must be handled with care, since setting the wrong value might prevent the signature from being valid for use with TensorFlow Serving. The option is available, because the converter only generates predict-signatures. In case the model is a regression model or a classifier with the matching outputs, the correct method name can be forced using the --method_name option.

Alternatively, you can create your own converter programs using the module's API. The API is required to accomplish more complicated tasks, like packaging multiple TensorFlow.js models into a single SavedModel.

Example

To convert a TensorFlow.js graph model to a TensorFlow frozen model (i.e. the most common use case?), just specify the directory containing the model.json, followed by the path and file name of the frozen model like so:

tfjs_graph_converter path/to/js/model path/to/frozen/model.pb

Advanced Example

Converting to TF SavedMovel format adds a lot of options for tweaking model signatures. The following example converts a Posenet model, which is a multi-head model.

We want to select only two of the four possible outputs and rename them in the model's signature, as follows:

• Input: input (from sub_2)
• Outputs: offsets and heatmaps (from float_short_offsets and float_heatmaps)

tfjs_graph_converter \
    ~/models/posenet/model-stride16.json \
    ~/models/posenet_savedmodel \
    --output_format tf_saved_model \
    --outputs float_short_offsets,float_heatmaps \
    --rename float_short_offsets:offsets,float_heatmaps:heatmaps,sub_2:input

After the conversion, we can examine the output and verify the new model signature:

saved_model_cli show --dir ~/models/posenet_savedmodel --all

MetaGraphDef with tag-set: 'serve' contains the following SignatureDefs:

signature_def['serving_default']:
  The given SavedModel SignatureDef contains the following input(s):
    inputs['input'] tensor_info:
        dtype: DT_FLOAT
        shape: (1, -1, -1, 3)
        name: sub_2:0
  The given SavedModel SignatureDef contains the following output(s):
    outputs['heatmaps'] tensor_info:
        dtype: DT_FLOAT
        shape: (1, -1, -1, 17)
        name: float_heatmaps:0
    outputs['offsets'] tensor_info:
        dtype: DT_FLOAT
        shape: (1, -1, -1, 34)
        name: float_short_offsets:0
  Method name is: tensorflow/serving/predict

Usage from within Python

The package installs the module tfjs_graph_converter, which contains all the functionality used by the converter script. You can leverage the API to either load TensorFlow.js graph models directly for use with your TensorFlow program (e.g. for inference, fine-tuning, or extending), or use the advanced functionality to combine several TFJS models into a single SavedModel. The latter is only supported using the API (it's just a single function call, though, so don't panic 😉)

Important Note

By default, Python code that includes the library will see CUDA devices disabled (i.e. not visible in the program). This is done because the library uses some low-level APIs that don't allow disabling GPUs from Python code. Unfortunately some GPUs support CUDA but don't have the compute capabilities or VRAM required to convert certain models. For this reason, CUDA devices are disabled by default and the converter and scripts using it use CPUs only.

This behaviour can now be disabled by calling enable_cuda() before any Tensorflow or converter function is called. This will re-enable the use of CUDA-capable devices, but may result in errors during model loading/conversion depending on the installed GPU hardware.

API Documentation