kshitij12345 commented on a change in pull request #15515: [MXNET-978] Higher
Order Gradient Support `arcsin`, `arccos`.
URL: https://github.com/apache/incubator-mxnet/pull/15515#discussion_r333091929
##########
File path: src/operator/tensor/elemwise_unary_op_trig.cc
##########
@@ -188,7 +188,31 @@ The storage type of ``arcsin`` output depends upon the
input storage type:
.set_attr<nnvm::FGradient>("FGradient", ElemwiseGradUseIn{ "_backward_arcsin"
});
MXNET_OPERATOR_REGISTER_BINARY_WITH_SPARSE_CPU_DR(_backward_arcsin,
-
unary_bwd<mshadow_op::arcsin_grad>);
+
unary_bwd<mshadow_op::arcsin_grad>)
+.set_attr<nnvm::FGradient>("FGradient",
+ [](const nnvm::NodePtr& n, const std::vector<nnvm::NodeEntry>& ograds) {
+ // ograds[0]: head_grad_grads (dL/dxgrad)
+ // inputs[0]: dL/dy
+ // inputs[1]: x (ElemwiseGradUseIn)
+ // f(x) = arcsin(x)
+ // n: f'(x) = 1/(1-x^2)^1/2
+ // f''(x) = f'(x) * x/(1-x^2)
+ // Note: x/(1-x^2) = x * f'(x)^2
+ auto dydx = n->inputs[0];
+ auto x = n->inputs[1];
+ auto dydx_mul_grad_x = nnvm::NodeEntry{n};
+ auto op = mxnet::util::NodeOpGen{n};
+
+ auto grad_x = op.div(dydx_mul_grad_x, dydx);
+ auto grad_x_square = op.square(grad_x);
+ auto grad_x_square_mul_x = op.mul(grad_x_square, x);
+ auto grad_grad_x = op.mul(dydx_mul_grad_x, grad_x_square_mul_x);
+
+ std::vector<nnvm::NodeEntry> ret;
+ ret.emplace_back(op.mul(ograds[0], grad_x));
Review comment:
Sorry for the late reply.
> if the first input is y_grad or dL/dy, this gradient should be dL/(dy*dx) ?
I am not sure of `dL` part because we don't really use it in computing the
loss function.
> didn't we have the convention of x_grad instead of grad_x?
Oops. Thanks. Was a old PR. Will update the names.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
With regards,
Apache Git Services
