Does TLT flip images during training a classifier?

Accelerated Computing Intelligent Video Analytics TAO Toolkit
1

Hello,

I am training a TLT Resnet18 classifier on TLT 2.0.

I have a dataset which contains some shapes pointing in specific direction and the classname is the direction itself. (Ex: An object facing right should be labelled as Right, etc)

Does the training process automatically do some flipping augmentation? If so, is there any way to turn it off?

Thanks

3

Do you mean you were training with TLT classification network, right?
Can you share your training spec?

4

Hey,
Thanks for your reply.
Yes I am training a TLT Classification network

model_config {
  arch: "resnet",
  n_layers: 10

  use_bias: False
  use_batch_norm: True
  all_projections: False
  freeze_blocks: 0
  freeze_blocks: 1
  freeze_blocks: 2
  freeze_blocks: 3
  # image size should be "3,H,W", where H,W >= 16
  input_image_size: "3,224,224"

}

train_config {
  train_dataset_path: "/data/train"
  val_dataset_path:   "/data/valid"

  pretrained_model_path: "/workspace/experiment/pretrained/tlt_pretrained_classification_vresnet10/resnet_10.hdf5"
  optimizer: "sgd"
  batch_size_per_gpu: 256
  n_epochs: 60
  n_workers: 16

  # regularizer
  reg_config {
    type: "L2"
    scope: "Conv2D,Dense"
    weight_decay: 0.00005
  }

  # learning_rate
  lr_config {
    scheduler: "step"
    learning_rate: 0.0005
    step_size: 6
    gamma: 0.3
  }
}

So the problem I am facing is that when I separate the objects into classes say ‘Left’ and ‘Right’, my accuracy is always 50% but when I group the classes that are opposite (eg: left and right become 1 class, top and bottom become 1 class, etc), the accuracy goes very high.
Based on this observation, I want to verify if there is any flipping augmentation happening during the training. Because my data is direction specific, flipping would affect the accuracy.

Thanks.

5

Yes, there is horizontal_flip by default during training.

6

I see, thanks.
By any chance is there a way to disable it?

7

Currently not. Internal team will work for it.

8

That would be great.
Thank you very much :)