Deep_Object_Pose icon indicating copy to clipboard operation
Deep_Object_Pose copied to clipboard

Training old epoch with modified dataset

Open CorbinFerrie opened this issue 3 years ago • 1 comments

Hi, I've realized recently that my dataset for my training set did not include enough lowlight and closeup shots of my object. Is it possible to retrain my model using a somewhat working epoch? Like including more images in my dataset to account for the parts of my model that aren't working? My model is also not really accurate so I had to change some of the sensitivity parameters in the config in order to get any pose information from it. In a bit of a time crunch so I don't have time to retrain for another 5 days like last time. Below are my attached images of what it looks like atm. Thanks, BeliefMapTest2 CloseUp FarAway

CorbinFerrie avatar May 10 '22 21:05 CorbinFerrie

If you create more data, you resume the training with the more data by passing --net and your set of weight. There will be a moment where the loss will be high but it will kick down back.

TontonTremblay avatar May 11 '22 03:05 TontonTremblay