I am doing a simple scan to see how dataset size affects training. Basically, I took 10% of the coco dataset and trained a yolov3 net (from scratch) to just look for people. Then I took 20% of the coco dataset and did the same thing.... all the way to 100%. What is strange is that all 9 nets are getting similar loss at the end (~7.5). I must be doing something wrong, right? I expected to see an exponential curve where loss started out high and assymptotically approached some value as the dataset increased to 100%. If it didn't approach a value (and still had a noticeable slope at 100%), then that meant more data could help my algorithm.
This is my .data file: classes= 1 train = train-run-less.txt valid = data/coco/5k.txt names = data/humans.names backup = backup
I am trying to train just one class (person) from the coco dataset. Something is not making sense, and in a sanity test, I discovered that the loss drops even if the training folder only contains 1 image (which doesnt even have people in it). I thought the way this worked was that it trained on the "train" images, then it tested the neural net on the "valid" images. How is it getting better at finding people in the "valid" images if it hasnt trained on a single one??
Basically I am trying to answer the question: "how much accuracy can I expect to gain as I increase the data?"