How many sequence files do you recommend to have when training a model to generalize well?
Could you please elaborate more for your question? Thanks.
When I am training, the validation loss is used to help understand if the training is overfitting to the training data. I am not able to see this loss while training using tlt 2.0 using faster rcnn.
By having more of these sequences from different videos it will help the model understand different scenes.
TLT provides “tlt-evalute” tool to check specific epoch’s mAP during training.
You can find it in other networks. For example, below spec means evaluation on the validation set will run every 10 epochs. You can see the validation accuracy, which is more indicative than validation loss.
But currently, faster-rcnn does not support this yet. Other kinds of detection networks suppport it.
Thank you Morgan.