AM5728: TIDL import model : Floating Point exception

rohit yadav

Intellectual 400 points

Part Number: AM5728

Hi,

We are trying to run our tensorflow lite model on DSP.

But when we use the import tool —tidl_model_import.out to import TensorFlow lite model to what TIDL surpports,

we got error : Floating point exception (core dumped)

Config file

----------------

randParams = 0

modelType = 3

quantizationStyle = 1

quantRoundAdd = 50

numParamBits = 8

inputNetFile = "/home/tidl/facenet.tflite"

inputParamsFile = "NA"

outputNetFile = "/home/tidl/model_Net.bin"

outputParamsFile = "/home/tidl/model_params.bin"

sampleInData = "/home/tidl/sample.jpg"

tidlStatsTool = "/opt/ti-processor-sdk-linux-am57xx-evm-06.03.00.106/linux-devkit/sysroots/x86_64-arago-linux/usr/bin/eve_test_dl_algo_ref.out"

Conversion logs :

Please find logs attached.

8054.error.txt

over 5 years ago

0 Praveen Eppa1 over 5 years ago

TI__Genius 17580 points

Hi Rohit,

Is it possible for you share the tflite model to replicate the issue?

Thanks,

Praveen

0 rohit yadav over 5 years ago in reply to Praveen Eppa1

Intellectual 400 points

Hi Praveen,

Following is the link to tflite model file :

drive.google.com/.../view

0 Yuan Zhao over 5 years ago in reply to rohit yadav

TI__Expert 4095 points

Hi Rohit,

From your log, you can see that TIDL network didn't have the correct tensor dimensions. You need to specify the input dimensions and provide valid images/inputs for calibration. e.g.

inWidth  = 160
inHeight = 160
inNumChannels = 3
inElementType = 1
rawSampleInData = 0
preProcType  = 2
sampleInData = "./a_reasonable_network_input_iamge.png/jpg/etc"

Once you fix the input in the import config file, you will run into another problem. TIDL has a constraint on the InnerProductLayer. If you can modify the tflite network, please insert a "Flatten" layer before the "FullyConnected"/"InnerProduce" layer.

- Yuan

Processors

Processors forum

AM5728: TIDL import model : Floating Point exception