TDA4VM: TDA4VM: getting object detection tensor and csv outputs

Hi Adam,

Can you please let us know if you tried one of the OpenVx application or CCS based TIDL test bench?

Regards,
Shyam

Hi,

Thanks for answering.

Meanwhile, I progressed in the process. Right now I'm using only TIDL- from model import to emulation and target execution.

1. Regarding Q2 from the original message - I have found that setting writeTraceLevel to 2/3 (in the config file) should provide tracing all the model layers. However, it happens only during import (for feature map comparison and quantization debugging). Why isn't it happening also on simulation or target inference?

2. And one more question - I have read in this thread that to enable single-channel models, I can only use raw binary inputs by setting inFileFormat=1 in the import confile file. In that format, is it possible to provide a batch of images and not only a single image each time?

Thanks,

Adam

Thanks a lot for your answer.

I have currently only one issue left- how to create a raw binary file from a single-channel png image, so the TIDL model will read it properly. As mentioned in the "Steps to Debug" chapter in the TIDL user guide, I'm setting foldPreBnConv2D = 0 in the import config file, and comparing the output of the first batchnorm layer to the original input binary.

First, I have created the input binary in several ways, mostly with the imagemagick's convert command. When I'm reading this file with np.fromfile(<file>, dtype=np.uint8), it looks exactly like the original png file.

Later, I import my model with this config file:

modelType = 2
numParamBits = 16
numFeatureBits = 16
inputNetFile = "../../test/testvecs/models/public/onnx/yolov3ours/model_sim.onnx"
outputNetFile = "../../test/testvecs/config/tidl_models/onnx/tidl_net_yolo3ours.bin"
outputParamsFile = "../../test/testvecs/config/tidl_models/onnx/tidl_io_yolo3ours_"
inFileFormat = 1
inDataNorm = 1
inMean = 0
inScale = 0.00392156862745098
inWidth = 416
inHeight = 416
inNumChannels = 1
inData = "../../test/testvecs/input/image2_magick9.y"
perfSimConfig = ../../test/testvecs/config/import/device_config.cfg
inElementType = 0
outDataNamesList = "output0,output1,output2"
postProcType = 0
debugTraceLevel = 2
writeTraceLevel = 3
foldPreBnConv2D = 0

Lastly, I extract the output of the batchnorm output, read it with np.fromfile(<file>, dtype=np.float32) and reshaping it to the original image size, I get a different result (left- original image, right- processed image).

Can you help me figure this out?

Thanks,

Adam

Looks like the resolution of the trace file that you are using is not the same as what you are expecting.

Can you share the trace file name used in the below command

read it with np.fromfile(<file>, dtype=np.float32

You are probably right, but only 32bit tracing resolution gives me the array size which I accept to receive (length of 416*416).

I just found out that by importing the model with numParamBits and numFeatureBits equal to 8 (and not 16 like before), I do receive the input I expected.

Attached the trace file. Thanks!

4152.input.tar

Do we have any open questions in this thread?

If not can I close this thread?

Yes, I am still waiting for a solution to how to parse correctly a raw binary for a 16-bit quantized model. I also attached before the output of the first batchnorm layer.

I am also attaching now the input binary which was used. It was saved in uint8 format:

input_uint8.tar

Thanks!

gentle reminder

Another gentle reminder

Hi Adam,

    If you are using 16 bit quantization then input raw data should be 16 bit. Is this the information you are looking for?

Regards,

Anshu