Make input a dictionary for multi-modal object detection #95
Conversation
| @@ -2,6 +2,7 @@ | |||
|
|
|||
| defaults: | |||
| - COCO_TorchvisionFasterRCNN | |||
| - override /model/[email protected]: preprocessor_multi_modal | |||
There was a problem hiding this comment.
Why is this the default now?
There was a problem hiding this comment.
A preprocessor is required and I made the single-modal normalizer the default one.
There was a problem hiding this comment.
But that specifies the multimodal one?
| _target_: torchvision.transforms.Compose | ||
| transforms: | ||
| - _target_: mart.transforms.GetItems | ||
| keys: ${datamodule.test_dataset.modalities} |
There was a problem hiding this comment.
I don't think this is correct. These keys need to be specified independently because I could load the images in [depth, rgb] order but require they be order as [rgb, depth] for the model.
There was a problem hiding this comment.
Would also be nice if this could use the yaml file below (preprocessor_single_modal.yaml).
There was a problem hiding this comment.
Now it use the other yaml file.
I think it's convenient to use interpolation by default. We can always change the keys in experiment.yaml or in the command line if we encounter that rare situation.
There was a problem hiding this comment.
But isn't there a silent bug if I switch the datamodule modality from [rgb, depth] to [depth, rgb]?
| @@ -0,0 +1,12 @@ | |||
| # @package model.modules.preprocessor | |||
There was a problem hiding this comment.
Why not create a modules directory? Why does this live under detection when it has nothing to do with detection?
There was a problem hiding this comment.
model/detection -> model/modules
| @@ -0,0 +1,6 @@ | |||
| # @package model.modules.preprocessor | |||
There was a problem hiding this comment.
Why not create a modules directory? Why does this live under detection when it has nothing to do with detection?
I also think this should just be preprocessor.yaml or something like that. Perhaps 8bit_preprocessor.yaml or something indicating that this is doing 0-255 normalization?
There was a problem hiding this comment.
I named it tuple_normalizer.yaml
dxoigmn
changed the title
What does this PR do?
Using dictionary input like
{"rgb": tensor1, "depth": tensor2}should make it easier to compose multi-modal adversaries.The dictionary is later converted back to tensor so that models understand the input.
This is backward compatible with single-modal object detection, because single-modal is a special case of multi-modal.
Type of change
Please check all relevant options.
Testing
Please describe the tests that you ran to verify your changes. Consider listing any relevant details of your test configuration.
Before submitting
pre-commit run -acommand without errorsDid you have fun?
Make sure you had fun coding 🙃