Closed kulkarnikeerti closed 1 year ago
@kulkarnikeerti
I don't think it is a problem. In this project, gt_tensor
contains a total of three predicted labels: objectness
, category label
, and bounding box label
. Among them, the objectness
label is very important, which determines the positive and negative samples in the training process. Usually, each target will be assigned with at least one positive sample, that is, gt_tensot[x, y, 0]=1.0
. If a certain gt_tensot[x, y, 0]=0.0
, it indicates that the corresponding prediction is a negative sample, and only the objectness loss is calculated, and the classification loss and regression loss are not calculated.
I am trying to use mapillary with yolo-nano model. And I get the target tensor after applying this
gt_creator
with 0's in it. Inif
condition which is present in thegt_creator
function (wheregt_tensor
is assigned with values) is not assigning the values togt_tensor
, hence it just returns the tensor wit 0's which is initialized in the beginning.https://github.com/yjh0410/PyTorch_YOLO-Family/blob/24b92e3989a7506ba7e047c284fef69d7250a3b9/train.py#L302
Okay, so I tried to run the original script instead with
mapillary
data, even withcoco
dataset its the same. I get a tensor with 0's. This isn't the expected behavior right? We should have values in thegt_tensor
?Thanks in advance, Keerti