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

[AMP Prim OP]support dropout prim ops bfloat16 dtype #54175

Merged
merged 5 commits into from
Jun 1, 2023
Merged

[AMP Prim OP]support dropout prim ops bfloat16 dtype #54175

Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
f4442aa
fix dropout api and support bf16 for prim
Charles-hit May 29, 2023
b1cfb2d
fix code style
Charles-hit May 29, 2023
63364a6
fix dropout test
Charles-hit May 30, 2023
189be83
fix dropout p = 0 test
Charles-hit May 31, 2023
185a15a
Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…
Charles-hit Jun 1, 2023
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
117 changes: 91 additions & 26 deletions test/legacy_test/test_dropout_op.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -311,10 +311,6 @@ def setUp(self):
'is_test': True,
}
self.outputs = {'Out': out}
# Because prim op compare res with dygraph
# when p = 0 dropout api return x,in dygraph mode x_grad = out_grad,
# but in static mode x_grad = []
self.enable_check_static_comp = False

def init_test_case(self):
self.input_size = [32, 64]
Expand Down Expand Up @@ -362,22 +358,10 @@ def setUp(self):
}

def test_check_output(self):
self.check_output()
self.check_output(check_prim=True)

def test_check_grad_normal(self):
self.check_grad(['X'], 'Out')

def test_check_output_for_prim(self):
# greater_equal does't support bfloat16 in cpu
if core.is_compiled_with_cuda():
self.check_output_with_place(core.CUDAPlace(0))

def test_check_grad_for_prim(self):
# greater_equal does't support bfloat16 in cpu
if core.is_compiled_with_cuda():
self.check_grad_with_place(
core.CUDAPlace(0), ['X'], 'Out', only_check_prim=True
)
self.check_grad(['X'], 'Out', check_prim=True)


class TestDropoutOpWithSeedOnCPUPlace(unittest.TestCase):
Expand Down Expand Up @@ -1451,8 +1435,9 @@ def forward(
training=True,
mode="upscale_in_train",
):
y = paddle.assign(x)
out = paddle.nn.functional.dropout(
x=x, p=p, axis=axis, training=training, mode=mode
x=y, p=p, axis=axis, training=training, mode=mode
)
return out

Expand All @@ -1476,6 +1461,16 @@ def apply_to_static(net, use_cinn):
'float32',
places,
),
(
'bfp16',
np.random.rand(100000),
0.3,
False,
'upscale_in_train',
1002,
'bfloat16',
places,
),
(
'fp64',
np.random.rand(100000),
Expand Down Expand Up @@ -1506,6 +1501,16 @@ def apply_to_static(net, use_cinn):
'float32',
places,
),
(
'p=1.0,dtype=bfp16',
np.random.rand(100000),
1.0,
True,
'upscale_in_train',
1002,
'bfloat16',
places,
),
(
'p=1.0,test=False',
np.random.rand(100000),
Expand All @@ -1517,15 +1522,35 @@ def apply_to_static(net, use_cinn):
places,
),
(
'p=0.0',
'p=1.0,test=False,dtype=bfp16',
np.random.rand(100000),
1.0,
False,
'upscale_in_train',
1002,
'bfloat16',
places,
),
(
'p=0.0',
np.random.rand(100000),
0,
True,
'upscale_in_train',
1002,
'float32',
places,
),
(
'p=0.0,dtype=bfp16',
np.random.rand(100000),
0,
True,
'upscale_in_train',
1002,
'bfloat16',
places,
),
(
'downgrade_train',
np.random.rand(100000),
Expand All @@ -1536,6 +1561,16 @@ def apply_to_static(net, use_cinn):
'float32',
places,
),
(
'downgrade_train,dtype=bfp16',
np.random.rand(100000),
0.5,
False,
'downscale_in_infer',
1002,
'bfloat16',
places,
),
(
'fp32_cpu',
np.random.rand(100000),
Expand Down Expand Up @@ -1571,7 +1606,11 @@ def apply_to_static(net, use_cinn):
class TestCompositeDropout(unittest.TestCase):
@classmethod
def setUpClass(cls):
cls.x = cls.x.astype(cls.dtype)
cls.x = (
cls.x.astype(cls.dtype)
if cls.dtype != "bfloat16"
else cls.x.astype("float32")
)
core._set_prim_all_enabled(True)

@classmethod
Expand All @@ -1596,12 +1635,18 @@ def get_eager_desire(self, place):
paddle.set_device("gpu")
core.set_prim_eager_enabled(False)
input_ = paddle.to_tensor(
data=self.x, dtype=self.dtype, place=place, stop_gradient=False
data=self.x,
dtype=self.dtype if self.dtype != "bfloat16" else "float32",
place=place,
stop_gradient=False,
)
output = paddle.nn.functional.dropout(
input_, self.p, training=(not self.is_test), mode=self.mode
)
grad = paddle.grad(output, input_)
if self.dtype == "bfloat16":
output = paddle.cast(output, "float32")
grad[0] = paddle.cast(grad[0], "float32")
return output, grad[0]

def test_static_comp(self):
Expand All @@ -1614,18 +1659,26 @@ def test_static_comp(self):
mp, sp = paddle.static.Program(), paddle.static.Program()
with paddle.static.program_guard(mp, sp):
input_ = paddle.static.data(
'x', shape=self.x.shape, dtype=self.x.dtype
'x',
shape=self.x.shape,
dtype=self.x.dtype
if self.dtype != "bfloat16"
else "float32",
)
input_.stop_gradient = False
y = paddle.assign(input_)
output = paddle.nn.functional.dropout(
input_,
y,
self.p,
training=(not self.is_test),
mode=self.mode,
)
if core._is_fwd_prim_enabled():
primapi.to_prim(mp.blocks)
grad = paddle.static.gradients(output, input_)[0]
if self.dtype == "bfloat16":
output = paddle.cast(output, "float32")
grad = paddle.cast(grad, "float32")
exe = paddle.static.Executor(place)
exe.run(sp)
fwd, rev = exe.run(
Expand Down Expand Up @@ -1662,14 +1715,20 @@ def test_jit_comp(self):
paddle.set_device("gpu")
paddle.seed(self.seed)
input_ = paddle.to_tensor(
data=self.x, dtype=self.dtype, place=place, stop_gradient=False
data=self.x,
dtype=self.dtype if self.dtype != "bfloat16" else "float32",
place=place,
stop_gradient=False,
)
net = PrimNet()
net = apply_to_static(net, False)
output = net(
input_, self.p, training=(not self.is_test), mode=self.mode
)
grad = paddle.grad(output, input_)
if self.dtype == "bfloat16":
output = paddle.cast(output, "float32")
grad[0] = paddle.cast(grad[0], "float32")
fwd_actual.append(output.numpy())
rev_actual.append(grad[0].numpy())
for i in range(len(self.places)):
Expand All @@ -1696,14 +1755,20 @@ def test_jit_comp_with_cinn(self):
paddle.set_device("gpu")
paddle.seed(self.seed)
input_ = paddle.to_tensor(
data=self.x, dtype=self.dtype, place=place, stop_gradient=False
data=self.x,
dtype=self.dtype if self.dtype != "bfloat16" else "float32",
place=place,
stop_gradient=False,
)
net = PrimNet()
net = apply_to_static(net, True)
output = net(
input_, self.p, training=(not self.is_test), mode=self.mode
)
grad = paddle.grad(output, input_)
if self.dtype == "bfloat16":
output = paddle.cast(output, "float32")
grad[0] = paddle.cast(grad[0], "float32")
fwd_actual.append(output.numpy())
rev_actual.append(grad[0].numpy())
i = 0
Expand Down