TDA4VM: MNIST model import error TIDL

Padmasree N

Part Number: TDA4VM

Hello Team,

I am trying to deploy MNIST pre-trained model from github in TIDL. I got the following warning message.

WARNING: Conv Layer Conv_0's coeff cannot be found(or not match) in coef file, Random coeff will be generated! Only for evaluation usage! Results are all random!

Please find the onnx model below

The TIDL_import_config file is as shown below.

modelType          = 2
numParamBits       = 8
inputNetFile      = "../../test/testvecs/models/public/onnx/mnist_8_opset.onnx"
outputNetFile      = "../../test/testvecs/config/tidl_models/onnx/mnist.bin"
outputParamsFile   = "../../test/testvecs/config/tidl_models/onnx/tidl_io_mnist_"
inDataNorm = 0
inWidth = 28
inHeight = 28
inNumChannels = 3
inData = ../../test/testvecs/config/detection_list_mnist.txt
inDataNamesList = input.1
postProcType = 1

Also, im getting "Segmentation fault" for all models including sample tutorial models.

------------------ Network Compiler Traces -----------------------------
Segmentation fault (core dumped)

Please let me know the solution for the same.

Also, kindly provide me some detailed explanation about numParamBits and inNumChannels params.

Thanks in advance!

Regards,

Padmasree N.

over 3 years ago

0 Anshu Jain over 3 years ago

TI__Guru 56820 points

Hi Padmasree,

Padmasree N said:
WARNING: Conv Layer Conv_0's coeff cannot be found(or not match) in coef file, Random coeff will be generated! Only for evaluation usage! Results are all random!

This indicates that parameters for a given layer parameters are not found in the network itself. Is this model some dummy models without trained parameters?

Padmasree N said:
Also, kindly provide me some detailed explanation about numParamBits and inNumChannels params.

Please refer : https://software-dl.ti.com/jacinto7/esd/processor-sdk-rtos-jacinto7/08_04_00_02/exports/docs/tidl_j721e_08_04_00_12/ti_dl/docs/user_guide_html/md_tidl_model_import.html#tidl_model_import_basic for these parameters. Let us know if something is not clear from it.

Looking at the network dimension it looks like you have 28 batches/ROI's can you confirm if that's the case?

Regards,
Anshu

0 Padmasree N over 3 years ago in reply to Anshu Jain

Intellectual 470 points

Hello Anshu,

Thanks for your reply!

The model is trained with parameters. And, 28x28 is the input image dimension; 32 is the batch size ,1 input image channel

Since, I am very new to TIDL, I have a question.

1) Does tensor channel mean the input image channel size or any other?

2) After importing the model to TIDL format, the corresponding .bin net file generated is further used with OpenVX API to deploy on target, or is there any other method for inference on target?

Please let me know the same.

Thank you.

Regards,

Padmasree N.

0 Padmasree N over 3 years ago in reply to Padmasree N

Intellectual 470 points

Dear Anshu Jain,

Any updates on my query?

Hoping for your reply

Thanks in advance!

0 Anshu Jain over 3 years ago in reply to Padmasree N

TI__Guru 56820 points

Padmasree N said:
1) Does tensor channel mean the input image channel size or any other?

Are you talking about inNumChannels as mentioned here :

https://software-dl.ti.com/jacinto7/esd/processor-sdk-rtos-jacinto7/08_04_00_02/exports/docs/tidl_j721e_08_04_00_12/ti_dl/docs/user_guide_html/md_tidl_model_import.html#tidl_model_import_basic

If yes then this is number on input channels in the networks input.

Padmasree N said:
2) After importing the model to TIDL format, the corresponding .bin net file generated is further used with OpenVX API to deploy on target, or is there any other method for inference on target?

Refer this : https://software-dl.ti.com/jacinto7/esd/processor-sdk-rtos-jacinto7/08_04_00_02/exports/docs/tidl_j721e_08_04_00_12/ti_dl/docs/user_guide_html/index.html

Regards,
Anshu

0 Padmasree N over 3 years ago in reply to Anshu Jain

Intellectual 470 points

Hello Anshu,

Thanks for your confirmation!

Please let me know how to get rid of the warnings.

WARNING: Conv Layer Conv_0's coeff cannot be found(or not match) in coef file, Random coeff will be generated! Only for evaluation usage! Results are all random!

Regards,

Padmasree N.

0 kumar.desappan over 3 years ago in reply to Padmasree N

TI__Mastermind 22145 points

Hi Padmasree ,

The number of In channels in your model is 1. So it is expected that the third line in the below is set to 1, but you are setting it as 3 which is not matching with model hat is why you are observing this error

inWidth = 28
inHeight = 28
inNumChannels = 3

Please change it to "inNumChannels = 1" and try agagin

Regards,

Kumar.D

Processors

Processors forum

TDA4VM: MNIST model import error TIDL