TDA4VM: using edgeai-tidl-tools to train custum model

oren sanderovich

Part Number: TDA4VM

Hi,

I am trying to quatize a new custum model using the tlrt_delegate.py i have changed from https://github.com/TexasInstruments/edgeai-tidl-tools/blob/master/README.md#python-examples.

I didn get any error when running python3.6 tflrt_delegate.py -c.

But when i am runnig the python3.6 tflrt_delegate.py the output array returned by infer_image is full of 0.

When running the same scripts with the models zoo, i get valid results, and an output image is being creayted.

I have the following questions:

1. how do i use greysacle images? i see the you have the following line:

imgs.append(Image.open(image_files[i]).convert('RGB').resize((new_width, new_height), PIL.Image.LANCZOS))

2. Do you support png files?

Can you see other erros from the logs?

Thanks

Oren

logs are attached.

models_out_build.txt

over 3 years ago

0 kumar.desappan over 3 years ago

TI__Mastermind 22145 points

1. The input tensor preparation has to be the same as the one used during the training of the model. Please refer to the code used for the training model that you are trying to deploy. The example provided here are for RGB images, Our analytics solution doesn't have any limitation on input tensor type.

2. We do support openCv or PIL for image read in these examples, yes these libs can support PNG file type

0 oren sanderovich over 3 years ago in reply to kumar.desappan

Prodigy 50 points

so i can get just get rid of the RGB sectionB,BGR,? there us no configuration if image type (RGB,BGR,greyscale?)? inWidth = 32

if i use tidl_model_import.out i have below configuration:
inHeight = 64
inNumChannels = 3
inDataNorm = 0
inFileFormat = 1

should i configure this also on the pythion example, if so, where?

0 Fadil Muhammed over 3 years ago in reply to oren sanderovich

TI__Prodigy 565 points

In python , inHeight ,inNumChannels values are taken from model itself(github.com/.../tflrt_delegate.py, so you don't need to configure separately.

Processors

Processors forum

TDA4VM: using edgeai-tidl-tools to train custum model