Hello! I have 2 question about training data for TLT classifier.
How should i preprocess data for people classifier, being used after people detector(peoplenet_v2)? I tried to train two different models, one on pictures with random resolutions and another on images rescaled to 224x224, but none of them works correctly. I got different results on deepstream and tlt validation/inference. I had similar problem, when using classifier as primary model and rescaling training data to 224x224 solved problem. Maybe i should resize pictures to 244, keeping its aspect ratio?
Can i use data augmentation in classifier models simultaneously with training, like with detectors? As i understand documentation, i have to augment data before training with tlt-augment. And i haven’t seen any augmentation fields in example configs nor in docs for spec files.
For TLT classification network, the input images do not need to be manually resized. The input dataloader resizes images as needed.
Please prepare the split dataset for different classes.
For reference, please trigger jupyter notebook and try to run the example.
examples/classification/classification.ipynb
What happens with data in deepstream pipeline after detection, but before classification? Classifier gets on input image and pointer to box? And what happened after with box(e.g, resize)?
I’m using peoplenet_v2 as detector with resnet18 classifier from ngc.nvidia.com, which was trained in tlt on custom dataset. As i said, i had similar problem and solved it with resizing dataset manually. I can provide training configs, if it’s necessary.