Open kfengtee opened 5 months ago
The recognition model fundamentally used now is built with pytorch and the export is in onnx.
Therefore I am not sure if both can be merged reliably
Moreover the logmelcalc.tflite has a small bug due to which the graph cannot process more than 1 sample at a time i.e batch size should be 1
Due to the same reasons both the base model and the preprocessing graphs where not merged together
I see, thanks for the explanation! Hmm yeah I think merging the new onnx model with tflite graph will be very tricky.
Do you still keep a copy of the script/weights used to convert the old Tensorflow model (e.g: in SavedModel/Keras/etc format/etc) into .tflite
files (for both logmelcalc.tflite
and baseModel.tflite
)? I'm keen to give it a try and see how to make logmelcalc.tflite
process in batch + merge with the old base model.
p/s: understand that the new pytorch/onnx model performs much better than the old tensorflow model, but the size is a bit too huge + personally inclined to use TFLite as the edge inference engine, hence still interested to play around with the old version first haha
Hi, first of all, thank you for sharing this awesome work!
I noticed that the first model version was divided into two separate graphs
logmelcalc.tflite
andbaseModel.tflite
. Wondering if there is any rationale behind this?Context: I am considering merging these two graphs into a single
.tflite
file for easier management. So wanted to check if this is feasible and if there are any potential issues I should be aware of.