How to use custom models to active learning

[CastleDream]

I have tried many many times to use tensorflowjs_converter to change my tensorflow.pd to a model.json
while I still can't figure out the specific format of model.json and classes.json in the example of cocoSSDModel which is the pre-trained model in Vott' active learning.

I want to know which method do you adopt to convert and get the model.json and classes.json
which one of TensorFlow SavedModel, TensorFlow Hub module, Keras HDF5 or tf.keras SavedModel???
and also, I noticed that the format of model.json of mine is different from the given, even I have the
same version with the Vott，i.e. tensorflowjs=1.0.3.

can someone give me a complete and detailed example of a custom model in active learning of Vott

[ghunkins]

+1 on this. Excellent tool, would love custom active learning.

[nansravn]

+1 on this. I'm trying to use for active learning a custom vision model that has been exported as a docker image. I have this model running with docker on the path http://127.0.0.1/image, but I can't use this locally hosted API for active learning. =/

[JacopoMangiavacchi]

@nansravn you can't reuse a custom vision model exported as a docker image. You need to convert the model to the TF.js format needed by VoTT.

Once you have your model saved in the TF.js format put the model.json file, the weights binary files referenced in the model.json file and the custom classes.json file containing your labels on a folder and point to that folder from VoTT Active Learning settings. Please look at the cocoSSDModel on this repo for a sample of all of these files.

#829

[naveenchary121]

i need help in creating vott active learning model plzzz help me

[murali513]

I want to know which method do you adopt to convert and get the model.json and classes.json
which one of TensorFlow SavedModel, TensorFlow Hub module, Keras HDF5 or tf.keras SavedModel???

[RafayAK]

@nansravn you can't reuse a custom vision model exported as a docker image. You need to convert the model to the TF.js format needed by VoTT.

Once you have your model saved in the TF.js format put the model.json file, the weights binary files referenced in the model.json file and the custom classes.json file containing your labels on a folder and point to that folder from VoTT Active Learning settings. Please look at the cocoSSDModel on this repo for a sample of all of these files.

#829

Where do I put the pre and post-processing steps?

[ghunkins]

@RafayAK An option may be to look into using a tf.keras.layers.Lambda layer.

[RafayAK]

@ghunkins I'm not sure I can do the pre and post-processing steps of a Keras Retina-Net-like model in lambda layers and successfully convert to Tensorflow Js. Do you have any resources or references?
I have an open issue on this #963