Making `transformation` an optional parameter in FasterRCNN

Hello,

🚀 Feature

I think it would be more generic to have transform(https://github.com/pytorch/vision/blob/3d65fc6723f1e0709916f24d819d6e17a925b394/torchvision/models/detection/faster_rcnn.py#L231) as a function that can be modified by users rather than a default one.

Motivation

I am applying transformations separately as a part of data augmentation, which includes cropping and resizing. Hence I would prefer to not do the twice while retraining FasterRCNN.

Pitch

I would like to have a fixed size input to be fed into the network for variable-sized images. At present, I do this by resizing the images separately as a part of DataLoader and adjust the parameters of GeneralizedRCNNTransform accordingly.

Alternatives

My present way of using FasterRCNN is an alternative. Since my set of transformations are pre-defined, I have to apply hacks such as setting mean to 0., std to 1. and altering min and max sizes to my default value(this would mean that scale=1 and interpolation would return the same image.

Additional context

While the input to the network is fixed size, I apply many other augmentations such as mirror, random cropping etc, inspired by SSD based networks. Hence I would prefer to do all augmentation in a separate place once instead of twice.

Thank you!

Edit : If you think this would be a meaningful change, I will be happy to send a Pull Request.

Start with torchvision/models/detection/faster_rcnn.py at the referenced transform parameter and read GeneralizedRCNNTransform. Trace how FasterRCNN applies transformations during training and inference, then define the behavior when users provide their own preprocessing. Done means users can avoid duplicate resizing and augmentation while variable-sized inputs remain supported.