Unable To Load Model Weights While Predicting (using Pytorch)

I have trained a Mask RCNN network using PyTorch and am trying to use the obtained weights to predict the location of apples in an image.. I am using the dataset from this paper, a

Solution 1:

There is no 'model' parameter in the saved checkpoint. If you look in train_rcnn.py:106:

torch.save(model.state_dict(), os.path.join(args.output_dir, 'model_{}.pth'.format(epoch)))

you see that they save just the model parameters. It should've been something like:

torch.save({
    "model": model.state_dict(),
    "optimizer": optimizer.state_dict(),
    "lr_scheduler": lr_scheduler.state_dict()
}, os.path.join(args.output_dir, 'model_{}.pth'.format(epoch)))

so then after loading you get a dictionary with 'model', and the other parameters they appear to be wanting to keep.

This seems to be a bug in their code.