@ops.RegisterGradient("Mod")
def _mod_grad(op, grad):
x, y = op.inputs
gz = grad
x_grad = gz
y_grad = tf.reduce_mean(-(x // y) * gz, reduction_indices=[0], keep_dims=True)
return x_grad, y_grad

I would like to use tf.mod in my graph, as well, in order to define a loss function that minimizes the error in a variable that wraps (e.g. an angle with a range of [-pi,pi)). For example, if
a1 = pi - eps and a2 = -pi + eps,
then I would like
a1 - a2 = -2 * eps,
as opposed to
a1 - a2 = -2 * eps + 2 * pi.

In order to test the above solution, must I rebuild Tensorflow from source? I'm currently using 0.12.0rc1.
Thanks!

Edit - Instructions for registering a gradient function can be found here: https://www.tensorflow.org/how_tos/adding_an_op/#implement_the_gradient_in_python

@kleinma For Python-only changes such as above, you need not rebuild TensorFlow from source

I'm not sure if mod in models is well-defined enough to know what the gradient should be in all cases. The angle case makes sense, but lots of uses of mod (binning, etc) are inherently discontinuous, certainly with integers. @girving, do you think we should provide a default mod gradient, or just let people override them appropriately for their application when they see this error?

The gradient of mod is zero a.e., so it's TensorFlow gradient is zero everywhere. We should definitely define a zero gradient (returning None).

@aselle Unassigning myself since I'm on leave.

@sherrym would you take a look since @girving is out? It sounds as if the gradient should be 0 for mod as defined, but there would be some customers for an angular_mod with a different gradient.

Any news guys? @girving

Wait a sec guys! The gradient of tf.mod(x, y) w.r.t x is 1, except for x / y = integer where it is not defined... but since these points are a set of measure zero, the gradient should always be one (?). Gradient of int division (//) or floor is of course zero...

One plot says it all:

xvals = np.linspace(0, 10, 100)
yvals = xvals % 3.14
plt.plot(xvals, yvals)

Anyway, looking forward for the gradients!

Ah, thought we were talking about integer mod, which has no gradient. Yes, float mod has gradient 1 w.r.t. x a.e., and also an a.e. defined gradient w.r.t. y.

OK actually I just assumed tf.mod = % but didn't think about it further...

@harpone check this out: http://math.stackexchange.com/questions/1849280/derivative-of-remainder-function-wrt-denominator

@yuefengz Can you take a look?

That stackexchange link is very helpful. I will implement the gradient function.

@yuefengz thanks!

@yuefengz any updates on your part?

@yuntatsai is working on a change, we will check it soon. Thanks for your patience.