Convertion possiblity

[ChabaneAmaury]

Hello,

Would it be possible to add a convertion script, which would take an already trained and saved "regular" model to convert it to the edge's version?

Best,

[FouadSakr]

Hello,

This feature is not supported but your contribution is welcome!

A script for this feature can look as follows (Assuming you saved your model and preprocessing params using pickle):
1- Read your saved model (using pickle),
2- Extract model parameters (e.g., in the case of SVM: support vectors, bias, etc.),
3- Pass these params to the save_params function (e.g., in SVM_OM.py),
4- Read the preprocessing parameters (e.g., PCA, Normalization, etc.) using the pickle framework also.
5- Extract preprocessing parameters (e.g, MinMax Scaler, PCA_components, etc.)
6- Pass preprocessing params to Preprocessing_OM.py.
7- Finally, the output source and header files containing model params and preprocessing params will be saved and ready to be used on the microcontrollers using the Micro-LM module.

Best,

[ChabaneAmaury]

Hello,

The complicated part is that I meant something like convert a Tensorflow model to this KerasClassifier (for example). Unfortunately, there seems to be some bugs when trying to save a tensorflow model using pickle, and same appears to be when saving to a .h5 file. The only way I found right now is:
1- Save the structure to a JSON file
2- Save the weights in a .h5 file
3- Load back the structure
4- Load back the weights
5.1- (Optional, depend on the method) Assume a loss function and optimizer (since they are not savec)
5- Pass all of this in a function called by the KerasClassifier constructor
6- Manually defining the inner model by calling the build_fn method (otherwise we need to call the fit() method and thus assume the loss function and optimizer like if the training was not already done)
7- Then we have our KerasClassifier, that we can pass through the ANN_OM.saveParams, but we lack a lot of infos..
I am not an expert unfortunately, though I think there might be a more elegant solution, but right now it is the only way I found. I would happily request your thoughts on this.

Best,