tensorflow-model-converter

Intro

Tensorflow has support to read models from multiple versions but lacks export functionality to save models to a different version. For example, one can not read a tensorflow 2.x model into 1.x due to the introduction of "ragged tensors".

This repo tries to fill that gap. It does so by loading a model and re-saving it into a target tensorflow version.

The program tests the models on a sample batch in an attempt to validate that the conversion was successful.

Known limitations

Testing only works for models with one input branch
Small numerical differences can occur when resaving the model (due to Tensorflow versions i believe). Make sure to benchmark the new model.

Requirements

Python 3.7+
Pip (very lightweight)
Docker (website)

Install

git clone https://github.com/OlofHarrysson/tensorflow-model-converter.git
cd tensorflow-model-converter
pip install -r requirements.txt
python main.py --help

Model compatibility

The project is new so haven't been tested very much. The plan is to have a compatibility matrix here for successfully/failed model conversions. For now, a list will do.

Functional keras models all versions
Nested functional keras model that was created through tensorflow.js

After conversion

Because some models have issues with normal saving/loading we save the model in multiple formats. Once the model is converted you can play around this code to figure out how to load your model.

from tensorflow import keras
import json

# Load via keras. If this works, nothing below is needed.
model_h5_path = 'model.h5'
model = keras.models.load_model(model_h5_path)

# Load model from json
model_json_path = 'model.json'
with open(model_json_path) as infile:
    json_model = json.load(infile)
model = keras.models.model_from_json(json.dumps(json_model))

# Load weights saved with model.save_weights(path/to/model)
h5_weights_path = 'model_weights.h5'
model.load_weights(h5_weights_path)

# Loading weights that are saved in the "tf-format" https://www.tensorflow.org/guide/keras/save_and_serialize#saving_subclassed_models
# For this the model has to be saved with model.save_weights('path_to_my_weights', save_format='tf')
tf_weights_path = 'model_weights_tf'
model.load_weights(tf_weights_path)

model.summary()

Feature requests

Vote for features or add suggestions you'd like to see at https://internetport.hellonext.co/b/tensorflow-model-converter

Pull requests are welcome :)