Loss_rpn_box_reg is nan [torchvision 0.3]

[kl_divergence]

I’m using torchvision 0.3.
Here’s an example from both datasets I’m using:

Here’s my model :

def get_model(num_classes):
    model = torchvision.models.detection.fasterrcnn_resnet50_fpn(pretrained=True)
    in_features = model.roi_heads.box_predictor.cls_score.in_features
    model.roi_heads.box_predictor = FastRCNNPredictor(in_features, num_classes) # replace the pre-trained head with a new one
    return model.cuda()

I’ve made both the dataset classes identical as per torchvision docs. I’ve seen the same behaviour when I train on high res images for target_dataset also after some epochs. What can done to deal with this ? What can be the reason for getting ‘nan’ for the first dataset in first run?
I’ve tried lowering lr, normalizing images but still the same issue persists.

[Magolor (Magolor Chen)]

fasterrcnn_resnet50_fpn takes bbox [x1,y1,x2,y2] as input but the data set is using [x1,y1,width,height] it seems.