Convert Subclassed Speech Recognition Model to Tensorflow.js

user_7411958 0
Viewed 176

I have a subclassed Speech Recognition model (link) with which I'd like to make inferences on my node.js server. I am trying to convert it using tfjs but because its a subclassed model I'm getting the following error:

NotImplementedError: Saving the model to HDF5 format requires the model to be a Functional model or a Sequential model. It does not work for subclassed models, because such models are defined via the body of a Python method, which isn't safely serializable. Consider saving to the Tensorflow SavedModel format (by setting save_format="tf") or using `save_weights`.

I am following the official tutorial, which doesn't count this scenario in. And surprisingly I couldn't find much info on the web apart from a closed issue.

Any ideas on how to convert a Subclassed Model to tensorflowjs?

tensorflowjs-converter tensorflow.js keras tensorflow speech-recognition
2 Answers

Ok, so I was specifically trying to convert a speech recognition model (link above) and it seems that most such models aren't supported at the moment by tfjs (including mozilla's deepspeech).

It will specifically throw this error:

ValueError: Unsupported Ops in the model before optimization
AudioSpectrogram

The command used being in this case:

tensorflowjs_converter path/to/qnet15/ path/to/qnet15/converted/ --input_format=tf_saved_model --output_format=tfjs_graph_model

This error can be silenced, however, by adding the --skip_op_check flag to the above command. It will generated the expected model.json with its corresponding weight binaries after a bunch of warnings.

But, if you try inference @node, the same error occurs:

Promise {
  <rejected> TypeError: Unknown op 'AudioSpectrogram'. File an issue at https://github.com/tensorflow/tfjs/issues so we can add it, or register a custom execution with tf.registerOp()

The model is basically useless. There is an open issue for this feature since some years now.

user_74119580

Instead of using model.save_weights() as in the tutorial you should use the other option which is model.save("my_model_dir") and you can check here for confirmation.

After saving the model directory you would want to convert it using the Tensorflowjs Converter

$ tensorflowjs_converter \
    --input_format=tf_saved_model \
    my_model_dir \ # input_path 
    converted_model # output_path
user_133376350
Related