Convert Multi-band MelGAN pretrained model to TFLite

Hi,

I’m looking to convert the pretrained universal vocoder for Multi-band MelGAN into a TensorFlow Lite .tflite file for use on Android.

I’m trying to convert multiband_melgan.v1_24k (universal vocoder) to be a drop-in replacement for the mbmelgan.tflite (not a universal vocoder) file in the Android example, found here.

As this is a .h5 file, my understanding is I need to go .h5 -> .pb -> .tflite

The problem is, once it reaches load_model, I get the error ValueError: No model found in config file.

Some research says this is because the model was exported with model.save_weights, this way the architecture information is not saved.

Is there any way I can load the architecture information so I can make this conversion? I believe this model was added by @ZDisket

[OscarVanL]

Hi, apologies, I worked this out on my own.

I needed to use the MultiBangMelGANGeneratorConfig to load the .yaml config that comes with the model.

My code then becomes:

with open(args.model_config) as conf:
    config = yaml.load(conf, Loader=yaml.Loader)
config = MultiBandMelGANGeneratorConfig(**config['multiband_melgan_generator_params'])
mblemgan = TFMBMelGANGenerator(config=config, name='MBMelGAN')
mblemgan._build()
# Load weights from file
mblemgan.load_weights(args.model_path)
print("Saving model as .pb")
mblemgan.save(os.path.join(args.outdir, filename))

Then you can load this into TFLiteConverter.from_saved_model, run convert() and then write this to a .tflite file 😃

[OscarVanL]

Nice, glad you got it to work 😃 did you use TF 2.3.1 or TF nightly?