Skip to content

Fix complex vjps for several unary ops #3766

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
obchain wants to merge 1 commit into
base: main
Choose a base branch
from
+50 −7
Open

Fix complex vjps for several unary ops #3766

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
35 changes: 28 additions & 7 deletions mlx/primitives.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1701,7 +1701,10 @@ std::vector<array> Cosh::vjp(
const std::vector<array>& cotangents,
const std::vector<int>& argnums,
const std::vector<array>&) {
return jvp(primals, cotangents, argnums);
// The vjp conjugates the jvp's multiplier (a no-op for real inputs).
return {conjugate(
jvp(primals, {conjugate(cotangents[0], stream())}, argnums)[0],
stream())};
}

std::vector<array> Cosh::jvp(
Expand Down Expand Up @@ -2775,7 +2778,10 @@ std::vector<array> Log1p::vjp(
const std::vector<array>& cotangents,
const std::vector<int>& argnums,
const std::vector<array>&) {
return jvp(primals, cotangents, argnums);
// The vjp conjugates the jvp's multiplier (a no-op for real inputs).
return {conjugate(
jvp(primals, {conjugate(cotangents[0], stream())}, argnums)[0],
stream())};
}

std::vector<array> Log1p::jvp(
Expand Down Expand Up @@ -4867,7 +4873,10 @@ std::vector<array> Sin::vjp(
const std::vector<array>& cotangents,
const std::vector<int>& argnums,
const std::vector<array>&) {
return jvp(primals, cotangents, argnums);
// The vjp conjugates the jvp's multiplier (a no-op for real inputs).
return {conjugate(
jvp(primals, {conjugate(cotangents[0], stream())}, argnums)[0],
stream())};
}

std::vector<array> Sin::jvp(
Expand All @@ -4892,7 +4901,10 @@ std::vector<array> Sinh::vjp(
const std::vector<array>& cotangents,
const std::vector<int>& argnums,
const std::vector<array>&) {
return jvp(primals, cotangents, argnums);
// The vjp conjugates the jvp's multiplier (a no-op for real inputs).
return {conjugate(
jvp(primals, {conjugate(cotangents[0], stream())}, argnums)[0],
stream())};
}

std::vector<array> Sinh::jvp(
Expand Down Expand Up @@ -5435,7 +5447,10 @@ std::vector<array> Square::vjp(
const std::vector<array>& cotangents,
const std::vector<int>& argnums,
const std::vector<array>&) {
return jvp(primals, cotangents, argnums);
// The vjp conjugates the jvp's multiplier (a no-op for real inputs).
return {conjugate(
jvp(primals, {conjugate(cotangents[0], stream())}, argnums)[0],
stream())};
}

std::vector<array> Square::jvp(
Expand Down Expand Up @@ -5607,7 +5622,10 @@ std::vector<array> Tan::vjp(
const std::vector<array>& cotangents,
const std::vector<int>& argnums,
const std::vector<array>&) {
return jvp(primals, cotangents, argnums);
// The vjp conjugates the jvp's multiplier (a no-op for real inputs).
return {conjugate(
jvp(primals, {conjugate(cotangents[0], stream())}, argnums)[0],
stream())};
}

std::vector<array> Tan::jvp(
Expand All @@ -5633,7 +5651,10 @@ std::vector<array> Tanh::vjp(
const std::vector<array>& cotangents,
const std::vector<int>& argnums,
const std::vector<array>&) {
return jvp(primals, cotangents, argnums);
// The vjp conjugates the jvp's multiplier (a no-op for real inputs).
return {conjugate(
jvp(primals, {conjugate(cotangents[0], stream())}, argnums)[0],
stream())};
}

std::vector<array> Tanh::jvp(
Expand Down
22 changes: 22 additions & 0 deletions python/tests/test_autograd.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1339,6 +1339,28 @@ def test_complex_log_vjp(self):
# Check against hand-computed vjps
self.assertTrue(mx.allclose(vjps[0], expected))

def test_complex_unary_vjps(self):
# For a holomorphic f the vjp is cotangent * conj(f'(z)); these ops used
# to delegate to their jvp and drop the conjugate for complex inputs.
mx.random.seed(0)
z = mx.random.normal((3, 4, 5), dtype=mx.complex64)
cotangent = mx.random.normal((3, 4, 5), dtype=mx.complex64)
z = mx.where(abs(z) < 1e-3, 1e-3 + 0j, z)

ops = {
mx.square: lambda x: 2 * x,
mx.sin: mx.cos,
mx.sinh: mx.cosh,
mx.cosh: mx.sinh,
mx.tan: lambda x: 1 / mx.cos(x) ** 2,
mx.tanh: lambda x: 1 - mx.tanh(x) ** 2,
mx.log1p: lambda x: 1 / (1 + x),
}
for fn, deriv in ops.items():
_, (vjp,) = mx.vjp(fn, [z], [cotangent])
expected = cotangent * mx.conj(deriv(z))
self.assertTrue(mx.allclose(vjp, expected, atol=1e-5), msg=str(fn))

def test_complex_abs_grad(self):
mx.random.seed(0)
primal = mx.random.normal((3, 4, 5), dtype=mx.complex64)
Expand Down