Remove bounding box utils and refactor retinanet #2039
Conversation
|
Not sure what best to do about this test failure. @sineeli do you know the minimal version of Keras we would need for all of this? |
It would require latest release |
| "`bounding_box.to_dense(bounding_boxes)` before passing your boxes " | ||
| "to `bounding_box.mask_invalid_detections()`." | ||
| ) | ||
| validation.validate_bounding_boxes(bounding_boxes) |
There was a problem hiding this comment.
Should we consider making this public API in Keras? (Though we can keep this usage for now)
There was a problem hiding this comment.
Or could we trigger this with public API? E.g. keras.utils.bounding_boxes.convert_format to self.bounding_box_format like we do in call?
There was a problem hiding this comment.
I think need to either remove this or duplicate it here. import keras.src.layers.preprocessing.image_preprocessing.bounding_boxes will means KerasHub will not even import for older than the latest version of keras (including the keras currently preinstalled in colab images).
We don't want that.
| image_size=None, | ||
| scale=None, | ||
| offset=None, | ||
| bounding_box_format, |
There was a problem hiding this comment.
should we have a default for this?
There was a problem hiding this comment.
I mean we can have same as default format used for weights, but it can vary as per user dataset.
There was a problem hiding this comment.
Yeah, but I think a default is still probably helpful? I see a lot of bounding_box_format="xyxy", defaults in Keras.
There was a problem hiding this comment.
Our default weights are in yxyx. But yes we can have xyxy(torch) follows or yxyx(tensorflow) follows this. All the tensorflow datasets will have yxyx, I would say yxyx would be good option.
| @@ -40,12 +55,14 @@ def call(self, inputs): | |||
| if self.norm_std: | |||
| x = x / self._expand_non_channel_dims(self.norm_std, x) | |||
|
|
|||
| return x | |||
| return x, y | |||
There was a problem hiding this comment.
probably this should be keras.utils.pack_x_y_sample_weight(x, y, sample_weight)?
There was a problem hiding this comment.
This still needs doing I think (otherwise you toss sample_weight).
There was a problem hiding this comment.
Still a few comments to address. The big one, we need to get rid of the private import of keras for that validation function. Let's remove or duplicate.
We might also want to set things up so the library still works with older keras versions, and throws an error when using an OD model that 3.8 is required. But I can help with that part.
| image_size=None, | ||
| scale=None, | ||
| offset=None, | ||
| bounding_box_format, |
There was a problem hiding this comment.
Yeah, but I think a default is still probably helpful? I see a lot of bounding_box_format="xyxy", defaults in Keras.
| @@ -40,12 +55,14 @@ def call(self, inputs): | |||
| if self.norm_std: | |||
| x = x / self._expand_non_channel_dims(self.norm_std, x) | |||
|
|
|||
| return x | |||
| return x, y | |||
There was a problem hiding this comment.
This still needs doing I think (otherwise you toss sample_weight).
| "`bounding_box.to_dense(bounding_boxes)` before passing your boxes " | ||
| "to `bounding_box.mask_invalid_detections()`." | ||
| ) | ||
| validation.validate_bounding_boxes(bounding_boxes) |
There was a problem hiding this comment.
I think need to either remove this or duplicate it here. import keras.src.layers.preprocessing.image_preprocessing.bounding_boxes will means KerasHub will not even import for older than the latest version of keras (including the keras currently preinstalled in colab images).
We don't want that.
|
Oh yeah make sense, we can only import when we have Keras>=3.8.0 version and then throw error when this doesn't satisfy? I hope this works. |
| class NonMaxSuppression(keras.layers.Layer): | ||
| """A Keras layer that decodes predictions of an object detection model. | ||
| Args: | ||
| bounding_box_format: The format of bounding boxes of input dataset. | ||
| Refer | ||
| TODO: link keras core bounding box docs | ||
| Refer: | ||
| for more details on supported bounding box formats. |
There was a problem hiding this comment.
add link to keras API
non_max_suppression,anchor_generator, andbox_matcherinto the modeling layers for better integration.