I want to use TLT for my own custom dataset. and I used TLT-v2-py3, My dataset has 7.5K images and I resized all of images/annotations into 1040x784, I used both detectnet2-res18/res50 for fine-truning.
I setted occlued = 0 for all objects, but I have some objects that overlaped with others.
Q1- For any input size of multiplying 16, It used pre-trained weights or maybe train from scratch?
Q2- For both netwrok, I get about same result after 60 epoch, why? this is related to dataset or network. even I get same reslut with yolo3-darknet19.
- End user can train with ngc pretrained model or train from scratch. It is not related to input size.
- It is not strange. I am afraid it is related to your dataset.
TLT has capability of writing training logs in txt files or showing the results during in the training like tensorboard?
Reference: How to read confidence in classification?