Implement ObjectDetection support

[benjamintli]

Summary

This PR is an attempt to implement support for ObjectDetection models in optimum ExecuTorch. This PR adds ExecuTorchModelForObjectDetection, ObjectDetectionExportableModule, a task for object-detection, and a test for DETR type models.

Notes

• ObjectDetectionExportableModule traces an object detection model, and stores num_channels, image_size, get_label_ids and get_label_names
  • num_channels is not consistently defined in configs for the existing object detection models, so the ObjectDetectionExportableModule will try a few different config options, and if it can't resolve it it'll default to 3 (RGB), which I feel like is a sensible default. the priority goes kwargs defined num_channels -> config defined num_channels -> 3
  • image_size is also not defined in configurations typically, as some models have dynamic sizes; there's not really a sensible default, since users would pick a size to use at inference/deployment time. So for this, the image size is passed in via the CLI and will only be used for object-detection models
  • get_label_ids and get_label_names are two flat lists that are used to construct id2label; for some reason executorch doesn't seem to support storing dicts as values in constant_methods (it gets flattened)
• ExecuTorchModelForObjectDetection has 3 attributes: num_channels, image_size, and id2label, which is a dict of class ids to labels.
• added timm as a dev dependency for DETR

Testing

• ran unit tests locally
• tried running this script: https://gist.github.com/benjamintli/c188ee848fd945ec3ef78558a5803b6c. Output:

Loaded ExecuTorchModelForObjectDetection
=== Running inference ===
Using a slow image processor as `use_fast` is unset and a slow processor was saved with this model. `use_fast=True` will be the default behavior in v4.52, even if the model was saved with a slow processor. This will result in minor differences in outputs. You'll still be able to use a slow processor with `use_fast=False`.
[cpuinfo_utils.cpp:71] Reading file /sys/devices/soc0/image_version
[cpuinfo_utils.cpp:87] Failed to open midr file /sys/devices/soc0/image_version
logits shape: torch.Size([1, 100, 92])
pred_boxes shape: torch.Size([1, 100, 4])

=== Detections (confidence > 0.7) ===
  remote: 1.00 @ box [38.75445556640625, 70.63130187988281, 177.2999267578125, 117.51690673828125]
  remote: 0.99 @ box [334.3410949707031, 73.98548126220703, 369.325439453125, 188.1864776611328]
  couch: 1.00 @ box [-0.009555816650390625, 1.426863670349121, 639.6910400390625, 474.4981994628906]
  cat: 1.00 @ box [11.712303161621094, 51.7900390625, 314.5272216796875, 469.29779052734375]
  cat: 1.00 @ box [345.2006530761719, 23.41107940673828, 640.120849609375, 370.40679931640625]

=== Drawing bounding boxes ===
Saved result to detections_output.jpg

Done!

This is my first PR in this repo, so I'm open to any feedback!

[benjamintli]

@guangy10 @larryliu0820 sorry for the tag, but would either of you be able to take a look at my PR? Not sure if there's a process I should have been following here 😅

[JacobSzwejbka]

Optimum CI is in a rough way right now so Ill mostly just look at your new test

[benjamintli]

Hey @larryliu0820 @JacobSzwejbka would you be able to take a look at my revision? Thank you!