Detectron2: How to output prediction in pred_masks_rle instead of pred_masks

[RomRoc]

I'm running instance segmentation inference on a model trained with custom dataset:

predictor = DefaultPredictor(cfg)
outputs = predictor(im)

I need to save "outputs" dict to filesystem, but it has very big size, because prediction is saved in outputs['instances'].pred_masks
How can I save predictions in in outputs['instances'].pred_masks_rle instead of outputs['instances'].pred_masks, that is more memory efficient?
Thanks

[ppwwyyxx]

The "RLE" format is defined by COCO api and you can use COCO api to convert masks to RLE. That's unrelated to detectron2.

[RomRoc]

Ok sorry, anyway if you can suggest any possible way to save efficiently instance prediction to filesystem, I will appreciate.

[ppwwyyxx]

Using COCO api to convert masks to RLE is an efficient way.

[RomRoc]

Ok I solved using pycocotools, I put the code below in case it's useful to someone.

Now I use instances_to_json to convert instances to json, then I save it to disk.
Is there an opposite function converting json to instances?
Thanks

import pycocotools.mask as mask_util

instances = outputs['instances']

instances.pred_masks_rle = [mask_util.encode(np.asfortranarray(mask)) for mask in instances.pred_masks]
for rle in instances.pred_masks_rle:
    rle['counts'] = rle['counts'].decode('utf-8')

instances.remove('pred_masks')

# TO TEST INVERT CONVERSION
instances.pred_masks = np.stack([mask_util.decode(rle) for rle in instances.pred_masks_rle])