search

microsoft / EdgeML

This repository provides code for machine learning algorithms for edge devices developed at Microsoft Research India.
Other
1.59k stars 370 forks source link

Seedot #249

Open mehreenjabeen opened 2 years ago

mehreenjabeen commented 2 years ago

Hi,

I want to generate inference code corresponding to an onnx model. Although readme file in overview section says "SeeDot takes as input trained floating-point models", I cannot find any argument in the SeeDot-dev.py file which takes trained model as input.

Please help to resolve this.

ShikharJ commented 2 years ago

@mehreenjabeen Hi, could you mention which model do you wish to generate the corresponding code for?

mehreenjabeen commented 2 years ago

Hi @ShikharJ,

I have two onnx models for keyword spotting task based on dnn and gru architecture

ShikharJ commented 2 years ago

@mehreenjabeen SeeDot/Shiftry might not have the full support for these models, but you can always try. See this for an example.

mehreenjabeen commented 2 years ago

@ShikharJ: I am following the same link you have shared, but my query is that there is no argument for feeding the trained onnx model to SeeDot-dev.py file. usage: SeeDot-dev.py [-h] [-a] [-e] [-d] [-m] [-n] [-dt] [-t] [-s] [-sf] [-l] [-lsf] [-tdr] [-o]

ShikharJ commented 2 years ago

Ah, sorry, please follow the instructions from here. Please note that the generated seedot_ast.txt will be the input.sd file that you need. As for dumping the model weights, you can simply do an online search on how to dump weights from a trained ONNX model.

mehreenjabeen commented 2 years ago

@ShikharJ Upon running compile.sh from here, following error appears: python: can't open file 'create_input.py': [Errno 2] No such file or directory I ran the below command: ./compile.sh ./models/dnn.onnx

ShikharJ commented 2 years ago

@mehreenjabeen Sorry for the delayed response, it seems that the SeeDot converter is currently broken, and might need a substantial amount of work. We're planning a new release in February, but till then, I don't see this issue getting resolved.