Creating your own custom pretraining weights .pkl for Aerial Imagery with Detectron2

Hello all , As per available pretrained nets which were trained on msra dataset or coco dataset and we have various versions of resnet50 ,101 and 152 . I was planning go ahead with resnet101 as we train it or fine tune it using existing .pkl , i want to create my own .pkl for further training to simplify like create my massive aerial imagery dataset and train it on top of coco dataset so it would have learned features of earlier and now new aerial features than use it for further training of which we create .pth file .
I tried convert .pth file to .pkl using script convert-torchvision-to-d2.py available in d2 tools folder however it could not be converted so i changed line `newmodel[k] = obj.pop(old_k).detach().numpy()` to `newmodel[k] = obj.pop(old_k)` it got converted however gave error when used for training . Please help me out thank you !

Start with tools/convert-torchvision-to-d2.py and reproduce the reported .pth-to-.pkl conversion using both versions of the changed line. Capture the training error and compare the generated weights with Detectron2's expected format; done means the custom aerial-image weights load successfully for further training.

Written by the indexing model from the issue text.