Errors: Disable online augmentation RetinaNet

Accelerated Computing Intelligent Video Analytics TAO Toolkit
1

Hi,

in the documentation for the TLT 3.0 the information to disable the online augmentation in the RetinaNet (lower note) is not quite correct and leads to errors. I use the TLT 3.0 version from 06.09.21 (v3.0-py3).

If the min and max parameters random_crop_scale and/or zoom_out_scale are both set to 1.0 or 1 as described in the note, the training aborts with the following error message:

ValueError: It must be `min_scale < max_scale`.

and if you set the random_crop_ar (min and max) parameters to 1 the following error message is generated:

ValueError: It must be `min_aspect_ratio < max_aspect_ratio`.

You can work around the problems with minimally different values (the following values are the smallest possible):

augmentation_config {
    output_width: 672
    output_height: 384
    output_channel: 3
    random_crop_min_scale: 0.9999999
    random_crop_max_scale: 1.0
    random_crop_min_ar: 0.9999999
    random_crop_max_ar: 1.0
    zoom_out_min_scale: 1.0
    zoom_out_max_scale: 1.0000001
    brightness: 0
    contrast: 0
    saturation: 0
    hue: 0
    random_flip: 0
}

Can you reproduce this discrepancy between documentation and program or I am doing something wrong?

P.S. I am very thankful that in v3.0-py3 compared to v3.0-dp-py3 there is now the possibility to turn off the online augmentation in the RetinaNet.

3

Thanks for the finding. I will check internally.