Edgeai TIDL Tools and Examples - This repository contains Tools and example developed for Deep learning runtime (DLRT) offering provided by TI’s edge AI solutions.
Other
121
stars
27
forks
source link
Questions about python examples running on a target board #29
My question is, when running onnxrt_ep.py on the target board TDA4x,
I feel like the python script's input is still onnx file, not .bin files, where I think .bin files are the product of compiling.
For example, when I check the codes below in the onnxrt_ep.py, config['model_path'] is ../../../models/public/resnet18_opset9.onnx.
# the execution part
EP_list = ['TIDLExecutionProvider', 'CPUExecutionProvider']
sess = rt.InferenceSession(config['model_path'],
providers=EP_list,
provider_options=[delegate_options, {}],
sess_options=so)
Shouldn't the target board inference be loaded from .bin files in model-artifacts? (ex. 191_tidl_io_1.bin, 191_tidl_net.bin)
Why did I have to compile on PC, if the target board can infer models by directly loading from onnx files?
Is there a python script example of creating and inferring a model from .bin files?
The delegate options has the model artifacts folder, name this shall be set to ./model-artifacts that is copied the TDA4x Target file system which has all the bin files for target execution
Hi. Thanks for this great library!
I'm following the Python Examples.: https://github.com/TexasInstruments/edgeai-tidl-tools/blob/master/examples/osrt_python/README.md I was able to compile models on PC and copied
./models
and./model-artifacts
to my TDA4x (SDK8.2)My question is, when running
onnxrt_ep.py
on the target board TDA4x, I feel like the python script's input is stillonnx
file, not.bin
files, where I think.bin
files are the product of compiling.For example, when I check the codes below in the
onnxrt_ep.py
,config['model_path']
is../../../models/public/resnet18_opset9.onnx
..bin
files inmodel-artifacts
? (ex.191_tidl_io_1.bin
,191_tidl_net.bin
)onnx
files?.bin
files?Can you please advise? What am I missing?
FYI, I'm setting like this on the board