• State
  • Locked Locked
  • Replies 1 reply
  • Subscribers 98 subscribers
  • Views 297 views
  • Users 0 members are here
Support feedback
Options
Options
Related

This thread has been locked.

If you have a related question, please click the "Ask a related question" button in the top right corner. The newly created question will be automatically linked to this question.

AM62A3: How to use multiple binary files as the validation set during model import？

tian liang
Prodigy 100 points
Part Number: AM62A3

Tool/software:

Dear TI expert,

When performing model import on the AM62A, we want to try using a sufficient number of images for calibration under the calibrationOption = 7 setting. However, since our input images are in bayer format RAW images, we set inFileFormat=1. As image input only supports png, jpg, and bmp formats, and not RAW images, we use a single RAW image binary file as the validation set during import. If we want to use multiple RAW images as the validation set, how should we proceed?

Thank you.

  • 0
    TI__Genius 15486 points

    Hello,

    Are you using the tidl_model_import.out? That file has a help dialog that will print with -h

    From this dialog (and some source files within firmware-builder), I see that you should set 'inFileFormat' to be 1 to use raw data and 'rawDataInElementType' based on your data type: 

    "Raw data Format of each input feature. The following types are currently supported:"
    " 0 : unsigned char"
    " 1 : signed char"
    " 2 : unsigned short"
    " 3 : signed short"
    " 6 : float"

    However, this may restrict the number of frames; I see some reference in the code that suggests it may force the number of frames to 1. Using raw input is generally a rarely used and lesser supported option. The documentation is thin here.

    It may be worth considering an alternative method. From the data-layout / format perspective, is there any special about these 'raw' images aside from the fact that they are not fully chromatic, RGB (or otherwise) images? Would it suffice to use a single-channel (e.g. grayscale) input in a format like BMP or PNG?

    If the images are bit-packed in some way, that would impact this, but I assume that would also have to be unpacked before input to a network anyway.

    BR,
    Reese