ximilar-com / tf-image Goto Github PK

Hi,

Thanks for the library!
I am new in TF library.
I would like to know if tf-image can be implemented on TPU and TF Records? If yes, can you pls provide and example.

Many thanks!

Yesterday I would like to try my code with Python 3.11 and TF 2.12.1. I'm using Python 3.8 and TF 2.10.1 as default, and there is everything OK.

But after upgrading I got this error:

File "/usr/local/lib/python3.11/dist-packages/tf_image/core/convert_type_decorator.py", line 15, in wrap  *
        bboxes_type = bboxes.dtype

AttributeError: 'float' object has no attribute 'dtype'

Does the repository work with Python 3.11 and TF 2.12.1? Do you know that?

Hi there.
In many of the tf_image.core.bboxes functions the required format is [ymin, xmin, ymin, xmax].
Is this a typo?
Should the format be [ymin, xmin, ymax, xmax]?

thank you!

Thank you for you contribution.

I am training Mask-RCNN model and I have a CPU bottleneck. I had to move from passing data via generators to tf.data. Unfortunately, now I do not have a possibility to do image augmentation (implementation of model which uses generators also uses imgaug library).

I found your library, and I would like to add a data augmentation also for segmentation masks (beside bounding boxes).

1. Do I understand correctly that I need mainly to make changes in tf_image.core.bboxes.rotate.random_rotate?
2. Are all operations made by your library fully processed on GPU, so there will not be increase in CPU utilization?

@tf.function
def _bboxes_to_relative(image, bboxes):
    image_height, image_width = tf.shape(image)[0], tf.shape(image)[1]
    bboxes_update = tf.cast(tf.stack([image_height, image_width, image_height, image_width]), dtype=tf.float32)
    return bboxes / bboxes_update


@tf.function
def _bboxes_to_absolute(image, bboxes):
    image_height, image_width = tf.shape(image)[0], tf.shape(image)[1]
    bboxes_update = tf.cast(tf.stack([image_height, image_width, image_height, image_width]), dtype=tf.float32)
    return bboxes * bboxes_update

in the above codes in here,
original bboxes is update with bboxes_update.

I think that bboxes_update might be tf.cast(tf.stack([image_width, image_height, image_width, image_height]), dtype=tf.float32), because bboxes value is [xmin, ymin, xmax, ymax]!

I am confused why bboxes is divided by [image_height, image_width, image_height, image_width], not [image_width, image_height, image_width, image_height]