Option to deactivate online augmentation

Accelerated Computing Intelligent Video Analytics TAO Toolkit
1

Continuing the discussion from Does TLT flip images during training a classifier?:

Is there an approximate schedule when the online augmentation deactivation will be implemented in TLT 3.0 and released with an update?

Due to a graphics card change to the RTX3070 I really need this feature as I can’t use TLT 2.0 anymore.

Thanks.

2

The change for "enable or disable horizontal_flip " will be available in next release.

3

Thanks @Morganh for the very quick reply, can you give a rough release date yet?

Unfortunately I have to get out my gtx1060 again to use TLT 2.0, because I have to disable the augmentation methods zoom-in, zoom-out and color jittering as well.

4

Hi Tim_P,
Maybe you misunderstand Does TLT flip images during training a classifier? . That topic is only talking about “enable or disable horizontal_flip”.
I’m afraid your question is different from that topic.
In TLT 2.0 or TLT 3.0, you can disable the augmentation methods such as zoom-in, zoom-out and color.

5

Oh OK, I assumed there was no function at all in TLT 3.0 to disable the various augmentations. Thanks for the information, if in the next release then still the flip function can be switched off I am happy and can leave my rtx3070 installed :-)

6

For reference, in object detection network, if you want to disable augmentation methods such as zoom-in, zoom-out , hflip, vflip ,translation and color, please set as below.

spatial_augmentation {
hflip_probability: 0.0
vflip_probability: 0.0
zoom_min: 1.0
zoom_max: 1.0
translate_max_x: 0.0
translate_max_y: 0.0
}
color_augmentation {
hue_rotation_max: 0.0
saturation_shift_max: 1.0
contrast_scale_max: 0.0
contrast_center: 0.5
}

7

Thank you @Morganh, great support, you have helped me a lot.

1 Like
8

Sorry to resurrect the post, maybe I was too unclear, because I use the Retinanet object detector. With the TLT 2.0 it worked exactly as you posted. But when I try this configuration in TLT 3.0, depending on how the configurations were inserted in the specification file, it gives the below listed error messages.

Could it be that the configuration and disabling of the individual online augmentations is not yet possible in all object detectors, for DetectNet_v2 I also found this option in the docs https://docs.nvidia.com/metropolis/TLT/tlt-user-guide/text/object_detection/detectnet_v2.html#augmentation-module, but not for RetinaNet

augmentation_config {
  output_width: 672
  output_height: 384
  output_channel: 3

    color_augmentation {
       hue_rotation_max: 0.0
       saturation_shift_max: 1.0
       contrast_scale_max: 0.0
       contrast_center: 0.5
     }

    spatial_augmentation {
       vflip_probability: 0.0
       zoom_min: 1.0
       zoom_max: 1.0
       translate_max_x: 0.0
       translate_max_y: 0.0
    }
}

Error:
google.protobuf.text_format.ParseError: 59:5 : Message type "AugmentationConfig" has no field named "color_augmentation".

or

augmentation_config {
  output_width: 672
  output_height: 384
  output_channel: 3
}

color_augmentation {
  hue_rotation_max: 0.0
  saturation_shift_max: 1.0
  contrast_scale_max: 0.0
  contrast_center: 0.5
  }

spatial_augmentation {
  vflip_probability: 0.0
  zoom_min: 1.0
  zoom_max: 1.0
  translate_max_x: 0.0
  translate_max_y: 0.0
}

Error:
google.protobuf.text_format.ParseError: 60:1 : Message type "Experiment" has no field named "color_augmentation".

9

@Tim_P
For TLT 3.0_dp version, Retinanet, there is no color or spatial augmentation configuration. It will be configurable in next release.

10

@Morganh
Good to know, thanks again.