your ciou customer op according to the ciou paper seems to have the same mistake the paper have.

[fjzhangcr]

System information

• OS Platform and Distribution (e.g., Linux Ubuntu 16.04):
• TensorFlow version and how it was installed (source or binary):
• TensorFlow-Addons version and how it was installed (source or binary):
• Python version:
• Is GPU used? (yes/no): no Describe the bug your ciou customer op according to the ciou paper seems to have the same mistake the paper have.

A clear and concise description of what the bug is. in the ciou paper <<Distance-IoU Loss: Faster and Better Learning for Bounding Box Regression>> https://arxiv.org/abs/1911.08287 , the deviation of v w.r.t w and h was descriped below: but after my manual calculation the deviation of v w.r.t w was below: that means the dv/dw should have a negative sign, and the dv/dh should not.

your ciou customer op is in https://github.com/tensorflow/addons/pull/914/files the key codes about ciou op are below:

def _get_v(
    b1_height: CompatibleFloatTensorLike,
    b1_width: CompatibleFloatTensorLike,
    b2_height: CompatibleFloatTensorLike,
    b2_width: CompatibleFloatTensorLike,
) -> tf.Tensor:
    @tf.custom_gradient
    def _get_grad_v(height, width):
        arctan = tf.atan(tf.math.divide_no_nan(b1_width, b1_height)) - tf.atan(
            tf.math.divide_no_nan(width, height)
        )
        v = 4 * ((arctan / math.pi) ** 2)

        def _grad_v(dv):
            gdw = dv * 8 * arctan * height / (math.pi ** 2)
            gdh = -dv * 8 * arctan * width / (math.pi ** 2)
            return [gdh, gdw]

        return v, _grad_v

    return _get_grad_v(b2_height, b2_width)

gdw = dv * 8 * arctan * height / (math.pi ** 2) miss a negative sign gdh =-dv * 8 * arctan * width / (math.pi ** 2) should remove the negative sign

the correct code should be :

def _get_v(
    b1_height: CompatibleFloatTensorLike,
    b1_width: CompatibleFloatTensorLike,
    b2_height: CompatibleFloatTensorLike,
    b2_width: CompatibleFloatTensorLike,
) -> tf.Tensor:
    @tf.custom_gradient
    def _get_grad_v(height, width):
        arctan = tf.atan(tf.math.divide_no_nan(b1_width, b1_height)) - tf.atan(
            tf.math.divide_no_nan(width, height)
        )
        v = 4 * ((arctan / math.pi) ** 2)

        def _grad_v(dv):
            gdw = -dv * 8 * arctan * height / (math.pi ** 2)
            gdh =   dv * 8 * arctan * width / (math.pi ** 2)
            return [gdh, gdw]

        return v, _grad_v

    return _get_grad_v(b2_height, b2_width)

Code to reproduce the issue here is my manual deviation test:

import tensorflow as tf
PI=tf.math.asin(1.)*2


with tf.GradientTape() as tape:
    w=tf.Variable(0.5,dtype=tf.float32)
    h=tf.Variable(0.1,dtype=tf.float32)
    v = 4/PI/PI*  tf.square(tf.math.atan(tf.math.divide_no_nan(1., 1.))-tf.math.atan(tf.math.divide_no_nan(w, h)))
theory= 4/PI/PI* 2*(tf.math.atan(1./1.)-tf.math.atan(tf.math.divide_no_nan(w, h)))*(-1)*h/(h**2+w**2)
grad_dv_dw=tape.gradient(v, w)
print(theory.numpy(),grad_dv_dw.numpy())


z=math.pi/4
PI=tf.math.asin(1.)*2
with tf.GradientTape() as tape:
    w=tf.Variable(0.5,dtype=tf.float32)
    h=tf.Variable(0.1,dtype=tf.float32)
    v = 4/PI/PI*tf.square(tf.math.atan(tf.math.divide_no_nan(1., 1.))
          -tf.math.atan(tf.math.divide_no_nan(w, h)))
grad_dv_dwdh=tape.gradient(v, [w,h])
print('scale:',(tf.square(w) + tf.square(h)).numpy())
theorical_dv_dw=(-1.)*h*8/tf.pow(PI,2)*(tf.math.atan(tf.math.divide_no_nan(1., 1.))
                                  -
                                  tf.math.atan(tf.math.divide_no_nan(w, h)))/(tf.square(w) + tf.square(h))
theorical_dv_dh=w*8/tf.pow(PI,2)*(tf.math.atan(tf.math.divide_no_nan(1., 1.))
                                  -
                                  tf.math.atan(tf.math.divide_no_nan(w, h)))/(tf.square(w) + tf.square(h))
print("tf dv/dw dv/dh",[_.numpy() for _ in grad_dv_dwdh])
print('theorical dv/dw dv/dh:',theorical_dv_dw.numpy(),theorical_dv_dh.numpy())

tensorflow console shows:

0.18331422 0.18331422
scale: 0.26
tf dv/dw dv/dh [0.18331422, -0.9165711]
theorical dv/dw dv/dh: 0.18331422 -0.9165711

but when i switch yolov3 ciou algorithm from tensorflow.addon (https://github.com/tensorflow/addons/pull/914/files#diff-f980d3eb2d746478b7664e652682ab3c62b24b7644559b9d5d4849087ec25d91) to mine own one, the model converge in the first epoch, but get nan in the second epoch~~something wents wrong! your code runs well but is not consistent with manual caculation，hope you guy check that whether it is a bug，or i do something wrong. thanks

btw, your customer op works well, while other implementation using tf.stop_gradient function to eliminate the scale factor (w2+h2), suffer in NaN. Provide a reproducible test case that is the bare minimum necessary to generate the problem.

Other info / logs

Include any logs or source code that would be helpful to diagnose the problem. If including tracebacks, please include the full traceback. Large logs and files should be attached.

[bhack]

/cc @fsx950223

[fsx950223]

Maybe you should tell the paper's author about it. I used a custom gradient in order to align with paper.

[fjzhangcr]

hi, thank your for you reply, i have post it to the author's github the same time i post here~~~ https://github.com/Zzh-tju/DIoU/issues/19 i post here, because your customer gradient implementation of ciou works better than other implementations using tf.stop_gradient, i think maybe you know the relationship of the deviation equation and the code of def _grad_v(dv) inside _get_v function, you do the coding on purpose~~ would you please check the deviation above?~~~is the deviation of v w.r.t w wrong? can you see the weird different between the deviation of v w.r.t w and the code inside def _grad_v(dv)?

[seanpmorgan]

TensorFlow Addons is transitioning to a minimal maintenance and release mode. New features will not be added to this repository. For more information, please see our public messaging on this decision: TensorFlow Addons Wind Down

Please consider sending feature requests / contributions to other repositories in the TF community with a similar charters to TFA: Keras Keras-CV Keras-NLP