Skip to content
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

Jump to bottom

[Relay] add test for second order ad #2754

Merged
merged 5 commits into from
Mar 30, 2019
Merged

[Relay] add test for second order ad #2754

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
bace027
do second order
MarisaKirisame Mar 9, 2019
33dd6d5
add comment
MarisaKirisame Mar 9, 2019
cbd6a35
better name
MarisaKirisame Mar 10, 2019
aa98c9a
use tvm assert all close
MarisaKirisame Mar 23, 2019
84fa952
refire ci
MarisaKirisame Mar 28, 2019
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
19 changes: 18 additions & 1 deletion python/tvm/relay/op/_tensor_grad.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3,7 +3,7 @@
from __future__ import absolute_import
from ..expr import const
from .op import register_gradient
from .transform import collapse_sum_like, where
from .transform import collapse_sum_like, broadcast_to_like, where
from .tensor import exp, negative, power, less
from .tensor import zeros_like, ones_like

Expand Down Expand Up @@ -77,3 +77,20 @@ def divide_grad(orig, grad):
x, y = orig.args
return [collapse_sum_like(grad / y, x),
collapse_sum_like(- (grad * orig / y), y)]


@register_gradient("zeros_like")
def zeros_like_grad(orig, grad):
"""Returns [0]"""
return [orig]

@register_gradient("ones_like")
def ones_like_grad(orig, grad):
"""Returns [0]"""
return [zeros_like(orig.args[0])]

@register_gradient("collapse_sum_like")
def collapse_sum_like_grad(orig, grad):
"""Returns [broadcast_to_like(grad, x), 0]"""
x, y = orig.args
return [broadcast_to_like(grad, x), zeros_like(y)]
74 changes: 48 additions & 26 deletions tests/python/relay/test_pass_gradient.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -20,8 +20,8 @@ def test_id():
ex = create_executor()
x = rand(dtype, *shape)
forward, (grad,) = ex.evaluate(back_func)(x)
np.testing.assert_allclose(forward.asnumpy(), x.asnumpy())
np.testing.assert_allclose(grad.asnumpy(), np.ones_like(x.asnumpy()))
tvm.testing.assert_allclose(forward.asnumpy(), x.asnumpy())
tvm.testing.assert_allclose(grad.asnumpy(), np.ones_like(x.asnumpy()))


def test_add():
Expand All @@ -35,8 +35,8 @@ def test_add():
ex = create_executor()
x = rand(dtype, *shape)
forward, (grad,) = ex.evaluate(back_func)(x)
np.testing.assert_allclose(forward.asnumpy(), 2 * x.asnumpy())
np.testing.assert_allclose(grad.asnumpy(), 2 * np.ones_like(x.asnumpy()))
tvm.testing.assert_allclose(forward.asnumpy(), 2 * x.asnumpy())
tvm.testing.assert_allclose(grad.asnumpy(), 2 * np.ones_like(x.asnumpy()))


def test_temp_add():
Expand All @@ -51,8 +51,8 @@ def test_temp_add():
ex = create_executor()
x = rand(dtype, *shape)
forward, (grad,) = ex.evaluate(back_func)(x)
np.testing.assert_allclose(forward.asnumpy(), 4 * x.asnumpy())
np.testing.assert_allclose(grad.asnumpy(), 4 * np.ones_like(x.asnumpy()))
tvm.testing.assert_allclose(forward.asnumpy(), 4 * x.asnumpy())
tvm.testing.assert_allclose(grad.asnumpy(), 4 * np.ones_like(x.asnumpy()))


def test_sub():
Expand All @@ -66,8 +66,8 @@ def test_sub():
ex = create_executor()
x = rand(dtype, *shape)
forward, (grad,) = ex.evaluate(back_func)(x)
np.testing.assert_allclose(forward.asnumpy(), np.zeros_like(x.asnumpy()))
np.testing.assert_allclose(grad.asnumpy(), np.zeros_like(x.asnumpy()))
tvm.testing.assert_allclose(forward.asnumpy(), np.zeros_like(x.asnumpy()))
tvm.testing.assert_allclose(grad.asnumpy(), np.zeros_like(x.asnumpy()))


def test_broadcast_add():
Expand All @@ -90,11 +90,11 @@ def test_broadcast_add():
relay.TupleType([t1, t2])]))
ex = create_executor()
forward, (grad_x, grad_y) = ex.evaluate(full_func)(x_nd, y_nd)
np.testing.assert_allclose(forward.asnumpy(), expected_forward)
np.testing.assert_allclose(grad_x.asnumpy(),
np.ones_like(expected_forward).sum(axis=2, keepdims=True))
np.testing.assert_allclose(grad_y.asnumpy(),
np.ones_like(expected_forward).sum(axis=(0, 1), keepdims=True).squeeze(axis=0))
tvm.testing.assert_allclose(forward.asnumpy(), expected_forward)
tvm.testing.assert_allclose(grad_x.asnumpy(),
np.ones_like(expected_forward).sum(axis=2, keepdims=True))
tvm.testing.assert_allclose(grad_y.asnumpy(),
np.ones_like(expected_forward).sum(axis=(0, 1), keepdims=True).squeeze(axis=0))


def test_broadcast_subtract():
Expand All @@ -117,11 +117,11 @@ def test_broadcast_subtract():
relay.TupleType([t1, t2])]))
ex = create_executor()
forward, (grad_x, grad_y) = ex.evaluate(full_func)(x_nd, y_nd)
np.testing.assert_allclose(forward.asnumpy(), expected_forward)
np.testing.assert_allclose(grad_x.asnumpy(),
np.ones_like(expected_forward).sum(axis=2, keepdims=True))
np.testing.assert_allclose(grad_y.asnumpy(),
-np.ones_like(expected_forward).sum(axis=(0, 1), keepdims=True).squeeze(axis=0))
tvm.testing.assert_allclose(forward.asnumpy(), expected_forward)
tvm.testing.assert_allclose(grad_x.asnumpy(),
np.ones_like(expected_forward).sum(axis=2, keepdims=True))
tvm.testing.assert_allclose(grad_y.asnumpy(),
-np.ones_like(expected_forward).sum(axis=(0, 1), keepdims=True).squeeze(axis=0))


def test_tuple():
Expand All @@ -147,10 +147,10 @@ def test_tuple():
expected_forward = x_np + y_np - z_np
ex = create_executor()
forward, (grad_x, grad_y, grad_z) = ex.evaluate(back_func)(x_nd, y_nd, z_nd)
np.testing.assert_allclose(forward.asnumpy(), expected_forward)
np.testing.assert_allclose(grad_x.asnumpy(), np.ones_like(grad_x.asnumpy()))
np.testing.assert_allclose(grad_y.asnumpy(), np.ones_like(grad_y.asnumpy()))
np.testing.assert_allclose(grad_z.asnumpy(), -1 * np.ones_like(grad_z.asnumpy()))
tvm.testing.assert_allclose(forward.asnumpy(), expected_forward)
tvm.testing.assert_allclose(grad_x.asnumpy(), np.ones_like(grad_x.asnumpy()))
tvm.testing.assert_allclose(grad_y.asnumpy(), np.ones_like(grad_y.asnumpy()))
tvm.testing.assert_allclose(grad_z.asnumpy(), -1 * np.ones_like(grad_z.asnumpy()))


def test_pow():
Expand All @@ -168,8 +168,9 @@ def test_pow():
i_nd = rand(dtype, *shape)
ex = create_executor(mod=mod)
forward, (grad_i,) = ex.evaluate(back_func)(i_nd)
np.testing.assert_allclose(forward.asnumpy(), 8 * i_nd.asnumpy())
np.testing.assert_allclose(grad_i.asnumpy(), 8 * np.ones_like(grad_i.asnumpy()))
tvm.testing.assert_allclose(forward.asnumpy(), 8 * i_nd.asnumpy())
tvm.testing.assert_allclose(grad_i.asnumpy(), 8 * np.ones_like(grad_i.asnumpy()))


def test_ref():
shape = (10, 10)
Expand All @@ -187,8 +188,28 @@ def test_ref():
x_nd = rand(dtype, *shape)
ex = create_executor()
forward, (grad_x,) = ex.evaluate(back_func)(x_nd)
np.testing.assert_allclose(forward.asnumpy(), 2 * x_nd.asnumpy())
np.testing.assert_allclose(grad_x.asnumpy(), 2 * np.ones_like(grad_x.asnumpy()))
tvm.testing.assert_allclose(forward.asnumpy(), 2 * x_nd.asnumpy())
tvm.testing.assert_allclose(grad_x.asnumpy(), 2 * np.ones_like(grad_x.asnumpy()))


def test_square_second_order():
shape = (10, 10)
dtype = 'float32'
t = relay.TensorType(shape, dtype)
x = relay.var("x", t)
func = relay.Function([x], x * x)
back_func = relay.ir_pass.infer_type(gradient(func))
y = relay.var("y", t)
back_func_adjusted = relay.Function([y], relay.TupleGetItem(relay.TupleGetItem(back_func(y), 1), 0))
back_func_adjusted = relay.ir_pass.infer_type(back_func_adjusted)
back_back_func = relay.ir_pass.infer_type(gradient(back_func_adjusted))
assert back_func.checked_type == relay.FuncType([t], relay.TupleType([t, relay.TupleType([t])]))
x_nd = rand(dtype, *shape)
ex = create_executor()
forward, (grad_x,) = ex.evaluate(back_back_func)(x_nd)
tvm.testing.assert_allclose(forward.asnumpy(), 2 * x_nd.asnumpy())
tvm.testing.assert_allclose(grad_x.asnumpy(), 2 * np.ones_like(grad_x.asnumpy()))


if __name__ == "__main__":
test_id()
Expand All @@ -200,3 +221,4 @@ def test_ref():
test_tuple()
test_pow()
test_ref()
test_square_second_order()