Object detection preprocessing layer raise ValueError with length of boxes

[jaytwoCaribou]

Hi,

The closest issue I found about the problem I mention in this issue is the following. However I use a tf.data pipeline, as recommended in the provided answer.
I have several questions, regarding the description of the issue :

• May I miss something in my minimal provided code example ?
• May I miss something in the versions of keras-cv and tf and their cross compatibilities ?
• Should I upgrade to keras 3 (keras-hub), but I read that ragged tensor are not supported ?

Current Behavior:

I try to follow the tutorial in keras 2 to train a Yolov8 here, but it seems that a ValueError is encountered during the preprocessing layer when I add augmentation. Without augmentation the data pipeline seems to run well. It does raise the error with other preprocessing layer keras_cv.layers.RandomShear or keras_cv.layers.JitteredResize.

ValueError: Expected len(boxes.shape)=2, or len(boxes.shape)=3, got len(boxes.shape)=4
which comes from the following lines
keras_cv/src/bounding_box/utils.py", line 158, in _format_inputs * raise ValueError
It seems that the keras_cv preprocessing layer is adding a dimension in the ragged_tensor of bboxes.

Expected Behavior:

• pipeline running without error

Steps To Reproduce:

Here is a minimal exemple to reproduce the error I get

import tensorflow as tf
from tensorflow import keras

import keras_cv
import numpy as np

def load_dataset(images, classes, bbox):
    image = images
    bounding_boxes = {
        "classes": tf.cast(classes, dtype=tf.float32),
        "boxes": bbox,
    }
    return {"images": tf.cast(image, tf.float32), "bounding_boxes": bounding_boxes}

classes = [
    [8, 8, 8, 8, 8],         # 5 classes
    [1],                  # 1 class
    [12, 14, 14],     # 3 classes
    ]

bbox = [
    [[199.0, 19.0, 390.0, 401.0],
    [217.0, 15.0, 270.0, 157.0],
    [393.0, 18.0, 432.0, 162.0],
    [1.0, 15.0, 226.0, 276.0],
    [19.0, 95.0, 458.0, 443.0]],     #image 1 has 4 objects
    [[52.0, 117.0, 109.0, 177.0]],   #image 2 has 1 object
    [[88.0, 87.0, 235.0, 322.0],
    [113.0, 117.0, 218.0, 471.0], 
    [113.0, 117.0, 218.0, 471.0]],   #image 3 has 2 objects
    ]

images = np.random.randint(0, 255, (3, 340, 340))

bbox = tf.ragged.constant(bbox)
classes = tf.ragged.constant(classes)
images = tf.ragged.constant(images)

augmenter = keras.Sequential(
    layers=[
        keras_cv.layers.RandomFlip(mode="horizontal", bounding_box_format="rel_xyxy"),
    ]
)


data = tf.data.Dataset.from_tensor_slices((images, classes, bbox))
ds = data.map(load_dataset)
ds = ds.ragged_batch(2, drop_remainder=True)
ds = ds.map(augmenter)

Version:

Here is the package I add in the pyproject.toml:

• python = ">=3.10,<3.13"
• tensorflow = "2.13"
• keras-cv = "0.6.4"

Additional information : I tried many combination with more recent packages (keras-cv 0.9, tf 2.17, tf2.14, ...), this is one of them, I never managed to solve the error.

Anything else:

I find it difficult to find documentation about keras-cv layers as the documentation always redirect to keras-hub, but it may be normal as the migration is ongoing ? Is there a specific documentation page that is maintained during migration ?

Anyway, thank you for this amazing work and futur contributions,

[sachinprasadhs]

Thanks for reporting the issue! We have consolidated the development of KerasCV into the new KerasHub package, which supports image, text, and multi-modal models. Please read the announcement. KerasHub will support all the core functionality of KerasCV.

KerasHub can be installed with !pip install -U keras-hub. Documentation and guides are available at keras.io/keras_hub.

With our focus shifted to KerasHub, we are not planning any further development or releases in KerasCV. If you encounter a KerasCV feature that is missing from KerasHub, or would like to propose an addition to the library, please file an issue with KerasHub.