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

[feat] add some GB ops #1361

Open
wants to merge 20 commits into
base: main
Choose a base branch
from
Open

[feat] add some GB ops #1361

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
20 commits
Select commit Hold shift + click to select a range
8fcc8e9
[feat] add isinf、trunc、round、hardsigmoid、elu、threshold
Nov 20, 2024
3ae1f36
[feat] add isinf、trunc、round、hardsigmoid、elu、prelu、selu、softplus
Nov 21, 2024
f61c61c
add isinf、trunc、round、hardsigmoid、elu、prelu、selu、softplus、softsign
Nov 21, 2024
4c1a999
[feat] add relu_backward
Nov 22, 2024
1e98d59
[feat] add erf back
Nov 24, 2024
f09e78e
add batch_norm_GB
Nov 25, 2024
a6dbbb6
add group_norm_GB
Nov 27, 2024
4a1484c
remove unused var
Nov 28, 2024
6e71c12
add dropout backward
Nov 29, 2024
fe30ff7
add group norm backward
Nov 29, 2024
070ea65
add group norm backward test condition
Nov 29, 2024
c471ddd
add batch norm backward
Nov 29, 2024
eff9287
format code style
Dec 2, 2024
f67d287
[remove] fp16 test in dropout test
Dec 2, 2024
0ddf5bb
try fix dropout bakward test
Dec 2, 2024
dcae615
check whether the diopiDropoutBackward function exists
Dec 2, 2024
3f772b3
modify dropoutbackward fp32 -> torch_input.dtype
Dec 2, 2024
8c6a1cc
modify dropout fp32&fp64 atol/rtol to 1e-3
Dec 2, 2024
11f6b97
dropoutbackward call aten impl
Dec 2, 2024
4bbc1ab
modify batch_norm_gb rtol/atol 1e-2
Dec 4, 2024
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
249 changes: 245 additions & 4 deletions diopi_test/python/configs/diopi_configs.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3,6 +3,197 @@


diopi_configs = {
'has_inf': dict(
name=["isinf"],
interface=["torch"],
atol=1e-3,
rtol=1e-4,
tensor_para=dict(
args=[
{
"ins": ['input'],
"shape": ((), (1024,), (2, 4096), (64, 28, 28), (32, 64, 112, 112), (64, 3, 7, 28, 28), (0,), (256, 0), (8, 0, 128)),
"dtype": [np.float16, np.float32, np.float64, np.int16, np.int32, np.int64, np.uint8, np.int8],
},
],
),
),

'trunc': dict(
name=["trunc"],
interface=["torch"],
atol=1e-3,
rtol=1e-4,
tensor_para=dict(
args=[
{
"ins": ['input'],
"shape": ((2, 16, 32, 56, 56), (2, 64, 32, 32), (2, 96, 28), (2, 16)),
"dtype": [np.float32, np.float16, np.float64],
},
],
),
),

'round': dict(
name=["round"],
interface=["torch"],
atol=1e-3,
rtol=1e-4,
tensor_para=dict(
args=[
{
"ins": ['input'],
"shape": ((2, 16, 32, 56, 56), (2, 64, 32, 32), (2, 96, 28), (2, 16)),
"dtype": [np.float32, np.float16, np.float64],
},
],
),
),

'round': dict(
name=["hardsigmoid"],
atol=1e-3,
rtol=1e-4,
tensor_para=dict(
args=[
{
"ins": ['input'],
"shape": ((2, 16, 32, 56, 56), (2, 64, 32, 32), (2, 96, 28), (2, 16)),
"dtype": [np.float32, np.float16, np.float64],
},
],
),
),

'elu': dict(
name=["elu"],
atol=1e-3,
rtol=1e-4,
para=dict(
alpha=[0.234, 4.8, -10, 1.0],
),
tensor_para=dict(
args=[
{
"ins": ["input"],
"shape": ((2, 16, 32, 56, 56), (2, 64, 32, 32), (2, 96, 28), (2, 16)),
"dtype": [np.float32, np.float16, np.float64],
},
],
),
),

'prelu': dict(
name=["prelu"],
atol=1e-3,
rtol=1e-4,
dtype=[np.float32, np.float16, np.float64],
tensor_para=dict(
args=[
{
"ins": ["input"],
"shape": ((2, 16, 32, 56, 56), (2, 64, 32, 32), (2, 96, 28), (2, 16)),
},
{
"ins": ["weight"],
"shape": ((16,), (64,), (96,), (1,)),
},
],
),
),

'selu': dict(
name=["selu"],
dtype=[np.float32, np.float16, np.float64],
atol=1e-3,
rtol=1e-4,
tensor_para=dict(
args=[
{
"ins": ["input"],
"shape": ((2, 16, 32, 56, 56), (2, 64, 32, 32), (2, 96, 28), (2, 16)),
},
],
),
),

'softplus': dict(
name=["softplus"],
atol=1e-3,
rtol=1e-4,
para=dict(
beta=[0.234, 4.8, -10, 1.0],
threshold=[0.234, 4.8, -10, 1.0]
),
tensor_para=dict(
args=[
{
"ins": ["input"],
"shape": ((2, 16, 32, 56, 56), (2, 64, 32, 32), (2, 96, 28), (2, 16)),
"dtype": [np.float32, np.float16, np.float64],
},
],
),
),

'softsign': dict(
name=["softsign"],
atol=1e-3,
rtol=1e-4,
tensor_para=dict(
args=[
{
"ins": ["input"],
"shape": ((2, 16, 32, 56, 56), (2, 64, 32, 32), (2, 96, 28), (2, 16)),
"dtype": [np.float32, np.float16, np.float64],
},
],
),
),

"batch_norm_GB": dict(
name=["batch_norm_GB"],
interface=['CustomizedTest'],
dtype=[np.float32, np.float16, np.float64],
atol=1e-2,
rtol=1e-2,
atol_half=1e-1,
rtol_half=1e-2,
para=dict(
training=[True, True, True],
momentum=[0.01, 0.01, 0.01],
axis=[0, 1, 2],
eps=[1e-4, 1e-4, 1e-4],
),
tensor_para=dict(
args=[
{
"ins": ["input"],
"requires_grad": [True],
"shape": ((2, 64, 32, 32),(2, 64, 32, 32),(2, 64, 32, 32)),
"gen_fn": "Genfunc.randn",
},
{
"ins": ["running_mean"],
"shape": ((2,), (64,), (32,)),
"gen_fn": "Genfunc.zeros",
},
{
"ins": ["running_var"],
"shape": ((2,), (64,), (32,)),
"gen_fn": "Genfunc.ones",
},
{
"ins": ["weight", "bias"],
"requires_grad": [True],
"shape": ((2,), (64,), (32,)),
"gen_fn": "Genfunc.randn",
},
]
),
),

# FIXME batch_norm输入0size的张量报错
'batch_norm': dict(
name=["batch_norm"],
Expand Down Expand Up @@ -507,17 +698,34 @@
args=[
{
"ins": ['input'],
"requires_grad": [True],
"shape": ((), (1024,), (2, 4096), (64, 28, 28),
(32, 64, 112, 112), (64, 3, 7, 28, 28),
(0,), (256, 0), (8, 0, 128)),
"dtype": [np.float16, np.float32, np.float64,
np.int16, np.int32, np.int64,
np.uint8, np.int8],
"dtype": [np.float16, np.float32, np.float64],
"gen_fn": 'Genfunc.randn',
},
],
),
),

'erf': dict(
name=['erf'],
interface=['torch'],
dtype=[np.float16, np.float32, np.float64],
tensor_para=dict(
gen_fn='Genfunc.randn',
args=[
{
"ins": ['input'],
"requires_grad": [True],
"shape": ((), (1, ), (1024,), (364800, 4), (2, 128, 3072),
(256, 128, 3, 3),
(2, 31, 512, 6, 40), (0,), (16, 0)),
},
],
),
),

'relu_no_contiguous': dict(
name=["relu"],
Expand Down Expand Up @@ -4902,6 +5110,8 @@
name=["dropout"],
no_output_ref=True,
is_inplace=True,
atol=1e-3,
rtol=1e-3,
para=dict(
p=[0.5, 0, 0.1, 0.4],
training=[True, True, True, False]
Expand All @@ -4911,7 +5121,7 @@
{
"ins": ['input'],
"shape": ((2, 4096), (32, 49, 256), (2, 16, 64, 64), (1, 2304, 1, 1, 1)),
"dtype": [np.float16, np.float32, np.float64],
"dtype": [np.float32, np.float64],
"gen_fn": 'Genfunc.positive',
},
],
Expand Down Expand Up @@ -6996,6 +7206,37 @@
]
),
),

'group_norm_GB': dict(
name=['group_norm_GB'],
interface=['CustomizedTest'],
atol=1e-4,
rtol=1e-5,
para=dict(
num_groups=[32, 4, 5, 1],
eps=[1e-05, 1e-05, 1e-05, 1e-05],
reduced_axes = [[2, 3], [1, 3], [0, 3], [2, 3]],
channel_axis = [1, 2, 1, 0]
),
tensor_para=dict(
args=[
{
"ins": ["input"],
"requires_grad": [True],
"shape": ((2, 256, 7, 10), (2, 256, 12, 12),
(12, 15, 8, 9),(3, 6, 9, 0)),
"dtype": [np.float32, np.float64, np.float16],
},
{
"ins": ["weight", "bias"],
"requires_grad": [True],
"shape": ((256,), (12,),
(15,), (3,)),
"dtype": [np.float32, np.float64, np.float16],
},
]
),
),

'unique': dict(
name=['unique'],
Expand Down
56 changes: 56 additions & 0 deletions diopi_test/python/conformance/customized_test.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -891,3 +891,59 @@ def pool3d(input, kernel_size, stride, padding, dilation, ceil_mode, count_inclu
def layer_normGB(input, weight, bias, eps, normalized_shape):
return torch.nn.functional.layer_norm(input=input, weight=weight, bias=bias, eps=eps, normalized_shape=normalized_shape)

def batch_norm_GB(input, running_mean, running_var, weight, bias, training=False, momentum=0.1, eps=1e-05, axis=1):
dim = input.dim()
dims = list(range(dim))
dims.remove(axis)
dims.insert(1, axis)
permuted_input = input.permute(dims)
out = torch.nn.functional.batch_norm(
permuted_input,
running_mean,
running_var,
weight=weight,
bias=bias,
training=training,
momentum=momentum,
eps=eps,
)
out = out.permute(dims)
return out

def group_norm_GB(input, num_groups, weight=None, bias=None, eps=1e-05, reduced_axes=[2, 3], channel_axis=1):

input_dims = list(input.size())
reduced_axes_set = set(reduced_axes)
dims = []
non_reduced_dims = []

for i, size in enumerate(input_dims):
if i == channel_axis:
continue
elif i in reduced_axes_set:
continue
else:
non_reduced_dims.append(i)
N = 1
for i in non_reduced_dims:
N = N * input.size(i)
HxW = 1
for i in reduced_axes:
HxW = HxW * input.size(i)
C = input.size(channel_axis)
dims = non_reduced_dims + [channel_axis] + reduced_axes
permuted_input = input.permute(dims)
reshaped_input = permuted_input.reshape([N, C, HxW, 1]).contiguous()
out = torch.nn.functional.group_norm(
reshaped_input,
num_groups,
weight=weight,
bias=bias,
eps=eps
)

reversed_order = [0]*len(dims)
for i in range(1, len(dims)):
reversed_order[dims[i]] = i
return out.reshape(permuted_input.shape).permute(reversed_order)

Loading