[RetinaNet] Image Converter and ObjectDetector #1906
Conversation
…neeli/keras-hub into sineeli/add-retinanet-phase-2
There was a problem hiding this comment.
Thanks for the PR @sineeli!! Looks generally good!! I have left a few comments.
Also, I want to make sure the -The code usage is updated to reflect the correct implementation - https://docs.google.com/document/d/15FUEP_vNehwLWJLragXhPFkYcbmmpbo0NYh5vL6q1xA/edit?tab=t.0
|
Kept some layers |
| "max_level": 7, | ||
| "num_scales": 3, | ||
| "aspect_ratios": [0.5, 1.0, 2.0], | ||
| "anchor_size": 8, | ||
| }, | ||
| input_data={ |
There was a problem hiding this comment.
define input_data also in setup()
There was a problem hiding this comment.
input_data varies in each test_case so kept it separately also how we assert changes
- Correct test cases.
…et format from preprocessor
… can effect the bounding boxes and the ops i backend framework dependent
- Add required docstrings - Use `center_xywh` encoding for retinanet as per torch weights
… arg for prediction head configuration
…e extraction from image encoder
There was a problem hiding this comment.
Thanks for the updates @sineeli. Left a few comments! This is looking good!
| Args: | ||
| anchors: `Tensors`. Anchor boxes with shape of `(N, 4)` where N is the | ||
| number of anchors. | ||
| boxes: `Tensors` Bounding boxes to encode. Boxes can be of be shape |
There was a problem hiding this comment.
with shape (B, N, 4) for batched boxes or (N, 4) for a single set of boxes. N should match the number of anchors.
|
|
||
| @keras_hub_export("keras_hub.layers.RetinaNetImageConverter") | ||
| class RetinaNetImageConverter(ImageConverter): | ||
| backbone_cls = RetinaNetBackbone |
There was a problem hiding this comment.
The converter has no resizing option, this is something we need to support for vision models. The backbone is taking image_shape as an input. This means we have to support image resizing in preprocessor.
| positive_threshold=0.5, | ||
| negative_threshold=0.4, | ||
| box_variance=[0.1, 0.1, 0.2, 0.2], | ||
| box_variance=[1.0, 1.0, 1.0, 1.0], |
There was a problem hiding this comment.
can you please explain why this update?
| anchors=anchor_boxes, | ||
| boxes=matched_gt_boxes, | ||
| anchor_format=self.bounding_box_format, | ||
| box_format=self.bounding_box_format, | ||
| encoding_format=self.encoding_format, | ||
| variance=self.box_variance, | ||
| image_shape=image_shape, | ||
| ) |
There was a problem hiding this comment.
for the return statement of this method. assign the value to box_targets and class_targets and then return them for clarity
keras_hub/src/models/retinanet/retinanet_object_detector_test.py
Outdated
Show resolved
Hide resolved
|
Weights Transfer Check:
|
|
Trained on |
…kend cause some gpu misplacement
This PR covers preprocessor for RetinaNet object detector and RetinaNet model itself. #1756
ImageObjectDetectorImageObjectDetectorPreprocessorRetinaNetObjectDetectorRetinaNetObjectDetectorPreprocessorRetinaNetImageConverter