Home Forums Model Development Kit (MDK) Model.json documentation

Viewing 1 post (of 1 total)
  • Author
    Posts
  • #16286
    Andrei Stoian
    Participant

    Hello,

    I am creating a network, inspired by the GTI VGG version, to perform classification of sounds, represented as 112×112 value mel-spectrograms. These values are now in floating point.

    I am training the network with pytorch and then I apply the quantization steps that I converted from the train.py script. I get good accuracy with the fully quantized model with 2 bit precision for all convolution layers (like GTI VGG).

    However, I am a bit stuck when exporting to the chip. In all _model.jsons that are included in the MDK the “input format” is CS_BGR24_PLANAR and operation is IMAGEREADER.

    1. Are there other options for “input format” and “operation”? I would like to plugin a FLOAT 2D array as input. Is it possible to have an input with only 1 channel? In float format?
    2. Do I need to write the _dat.json myself or do you have a tool to generate it from the model definition?
    3. Do you have documentation for the other fields in the model.json (the user’s guide seems to detail only the dat.json in the appendix).

    Regards,
    Andrei Stoian
    THALES

Viewing 1 post (of 1 total)
  • You must be logged in to reply to this topic.