Skip to content

Commit

Permalink
[Fix] Create Tensor with new_* method to support amp (#2389)
Browse files Browse the repository at this point in the history
  • Loading branch information
@zhouzaida
q.yao authored and zhouzaida committed Nov 23, 2022
1 parent ac47088 commit 3bb0611
Show file tree
Hide file tree
Showing 5 changed files with 116 additions and 184 deletions.
4 changes: 2 additions & 2 deletions mmcv/ops/diff_iou_rotated.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -235,9 +235,9 @@ def box2corners(box: Tensor) -> Tensor:
"""
B = box.size()[0]
x, y, w, h, alpha = box.split([1, 1, 1, 1, 1], dim=-1)
x4 = torch.FloatTensor([0.5, -0.5, -0.5, 0.5]).to(box.device)
x4 = box.new_tensor([0.5, -0.5, -0.5, 0.5]).to(box.device)
x4 = x4 * w # (B, N, 4)
y4 = torch.FloatTensor([0.5, 0.5, -0.5, -0.5]).to(box.device)
y4 = box.new_tensor([0.5, 0.5, -0.5, -0.5]).to(box.device)
y4 = y4 * h # (B, N, 4)
corners = torch.stack([x4, y4], dim=-1) # (B, N, 4, 2)
sin = torch.sin(alpha)
Expand Down
6 changes: 3 additions & 3 deletions mmcv/ops/group_points.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -233,7 +233,7 @@ def forward(
else:
B, nfeatures, nsample = indices.size()
_, C, N = features.size()
output = torch.cuda.FloatTensor(B, C, nfeatures, nsample)
output = features.new_zeros(B, C, nfeatures, nsample)

ext_module.group_points_forward(
features,
Expand Down Expand Up @@ -262,7 +262,7 @@ def backward(ctx, grad_out: torch.Tensor) -> Tuple:
idx, N = ctx.for_backwards

B, C, npoint, nsample = grad_out.size()
grad_features = torch.cuda.FloatTensor(B, C, N).zero_()
grad_features = grad_out.new_zeros(B, C, N)

grad_out_data = grad_out.data.contiguous()
ext_module.group_points_backward(
Expand All @@ -279,7 +279,7 @@ def backward(ctx, grad_out: torch.Tensor) -> Tuple:
B, N, idx, features_batch_cnt, idx_batch_cnt = ctx.for_backwards

M, C, nsample = grad_out.size()
grad_features = torch.cuda.FloatTensor(N, C).zero_()
grad_features = grad_out.new_zeros(N, C)

grad_out_data = grad_out.data.contiguous()
ext_module.stack_group_points_backward(
Expand Down
4 changes: 2 additions & 2 deletions mmcv/ops/three_interpolate.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -38,7 +38,7 @@ def forward(ctx: Any, features: torch.Tensor, indices: torch.Tensor,
B, c, m = features.size()
n = indices.size(1)
ctx.three_interpolate_for_backward = (indices, weight, m)
output = torch.cuda.FloatTensor(B, c, n)
output = features.new_empty(B, c, n)

ext_module.three_interpolate_forward(
features, indices, weight, output, b=B, c=c, m=m, n=n)
Expand All @@ -58,7 +58,7 @@ def backward(
idx, weight, m = ctx.three_interpolate_for_backward
B, c, n = grad_out.size()

grad_features = torch.cuda.FloatTensor(B, c, m).zero_()
grad_features = grad_out.new_zeros(B, c, m)
grad_out_data = grad_out.data.contiguous()

ext_module.three_interpolate_backward(
Expand Down
251 changes: 90 additions & 161 deletions tests/test_ops/test_group_points.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -7,7 +7,8 @@

@pytest.mark.skipif(
not torch.cuda.is_available(), reason='requires CUDA support')
def test_grouping_points():
@pytest.mark.parametrize('dtype', [torch.half, torch.float, torch.double])
def test_grouping_points(dtype):
idx = torch.tensor([[[0, 0, 0], [3, 3, 3], [8, 8, 8], [0, 0, 0], [0, 0, 0],
[0, 0, 0]],
[[0, 0, 0], [6, 6, 6], [9, 9, 9], [0, 0, 0], [0, 0, 0],
Expand Down Expand Up @@ -35,51 +36,37 @@ def test_grouping_points():
[
-0.6646, -0.6870, -0.1125, -0.2224, -0.3445,
-1.4049, 0.4990, -0.7037, -0.9924, 0.0386
]]]).cuda()
]]],
dtype=dtype).cuda()

output = grouping_operation(features, idx)
expected_output = torch.tensor([[[[0.5798, 0.5798, 0.5798],
[-1.3311, -1.3311, -1.3311],
[0.9268, 0.9268, 0.9268],
[0.5798, 0.5798, 0.5798],
[0.5798, 0.5798, 0.5798],
[0.5798, 0.5798, 0.5798]],
[[5.4247, 5.4247, 5.4247],
[1.4740, 1.4740, 1.4740],
[2.1581, 2.1581, 2.1581],
[5.4247, 5.4247, 5.4247],
[5.4247, 5.4247, 5.4247],
[5.4247, 5.4247, 5.4247]],
[[-1.6266, -1.6266, -1.6266],
[-1.6931, -1.6931, -1.6931],
[-1.6786, -1.6786, -1.6786],
[-1.6266, -1.6266, -1.6266],
[-1.6266, -1.6266, -1.6266],
[-1.6266, -1.6266, -1.6266]]],
[[[-0.0380, -0.0380, -0.0380],
[-0.3693, -0.3693, -0.3693],
[-1.8527, -1.8527, -1.8527],
[-0.0380, -0.0380, -0.0380],
[-0.0380, -0.0380, -0.0380],
[-0.0380, -0.0380, -0.0380]],
[[1.1773, 1.1773, 1.1773],
[6.0865, 6.0865, 6.0865],
[2.8229, 2.8229, 2.8229],
[1.1773, 1.1773, 1.1773],
[1.1773, 1.1773, 1.1773],
[1.1773, 1.1773, 1.1773]],
[[-0.6646, -0.6646, -0.6646],
[0.4990, 0.4990, 0.4990],
[0.0386, 0.0386, 0.0386],
[-0.6646, -0.6646, -0.6646],
[-0.6646, -0.6646, -0.6646],
[-0.6646, -0.6646, -0.6646]]]]).cuda()
expected_output = torch.tensor(
[[[[0.5798, 0.5798, 0.5798], [-1.3311, -1.3311, -1.3311],
[0.9268, 0.9268, 0.9268], [0.5798, 0.5798, 0.5798],
[0.5798, 0.5798, 0.5798], [0.5798, 0.5798, 0.5798]],
[[5.4247, 5.4247, 5.4247], [1.4740, 1.4740, 1.4740],
[2.1581, 2.1581, 2.1581], [5.4247, 5.4247, 5.4247],
[5.4247, 5.4247, 5.4247], [5.4247, 5.4247, 5.4247]],
[[-1.6266, -1.6266, -1.6266], [-1.6931, -1.6931, -1.6931],
[-1.6786, -1.6786, -1.6786], [-1.6266, -1.6266, -1.6266],
[-1.6266, -1.6266, -1.6266], [-1.6266, -1.6266, -1.6266]]],
[[[-0.0380, -0.0380, -0.0380], [-0.3693, -0.3693, -0.3693],
[-1.8527, -1.8527, -1.8527], [-0.0380, -0.0380, -0.0380],
[-0.0380, -0.0380, -0.0380], [-0.0380, -0.0380, -0.0380]],
[[1.1773, 1.1773, 1.1773], [6.0865, 6.0865, 6.0865],
[2.8229, 2.8229, 2.8229], [1.1773, 1.1773, 1.1773],
[1.1773, 1.1773, 1.1773], [1.1773, 1.1773, 1.1773]],
[[-0.6646, -0.6646, -0.6646], [0.4990, 0.4990, 0.4990],
[0.0386, 0.0386, 0.0386], [-0.6646, -0.6646, -0.6646],
[-0.6646, -0.6646, -0.6646], [-0.6646, -0.6646, -0.6646]]]],
dtype=dtype).cuda()
assert torch.allclose(output, expected_output)


@pytest.mark.skipif(
not torch.cuda.is_available(), reason='requires CUDA support')
def test_stack_grouping_points():
@pytest.mark.parametrize('dtype', [torch.half, torch.float, torch.double])
def test_stack_grouping_points(dtype):
idx = torch.tensor([[0, 0, 0], [3, 3, 3], [8, 8, 8], [1, 1, 1], [0, 0, 0],
[2, 2, 2], [0, 0, 0], [6, 6, 6], [9, 9, 9], [0, 0, 0],
[1, 1, 1], [0, 0, 0]]).int().cuda()
Expand All @@ -106,130 +93,72 @@ def test_stack_grouping_points():
[
-0.6646, -0.6870, -0.1125, -0.2224, -0.3445,
-1.4049, 0.4990, -0.7037, -0.9924, 0.0386
]]).float().cuda()
]],
dtype=dtype).cuda()
features_batch_cnt = torch.tensor([3, 3]).int().cuda()
indices_batch_cnt = torch.tensor([6, 6]).int().cuda()
output = grouping_operation(features, idx, features_batch_cnt,
indices_batch_cnt)
expected_output = torch.Tensor([[[0.5798, 0.5798, 0.5798],
[-0.7981, -0.7981, -0.7981],
[-0.9280, -0.9280, -0.9280],
[-1.3311, -1.3311, -1.3311],
[1.3687, 1.3687, 1.3687],
[0.9277, 0.9277, 0.9277],
[-0.4164, -0.4164, -0.4164],
[-1.8274, -1.8274, -1.8274],
[0.9268, 0.9268, 0.9268],
[0.8414, 0.8414, 0.8414]],
[[0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000]],
[[0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000]],
[[5.4247, 5.4247, 5.4247],
[1.5113, 1.5113, 1.5113],
[2.3944, 2.3944, 2.3944],
[1.4740, 1.4740, 1.4740],
[5.0300, 5.0300, 5.0300],
[5.1030, 5.1030, 5.1030],
[1.9360, 1.9360, 1.9360],
[2.1939, 2.1939, 2.1939],
[2.1581, 2.1581, 2.1581],
[3.4666, 3.4666, 3.4666]],
[[0.5798, 0.5798, 0.5798],
[-0.7981, -0.7981, -0.7981],
[-0.9280, -0.9280, -0.9280],
[-1.3311, -1.3311, -1.3311],
[1.3687, 1.3687, 1.3687],
[0.9277, 0.9277, 0.9277],
[-0.4164, -0.4164, -0.4164],
[-1.8274, -1.8274, -1.8274],
[0.9268, 0.9268, 0.9268],
[0.8414, 0.8414, 0.8414]],
[[-1.6266, -1.6266, -1.6266],
[-1.0281, -1.0281, -1.0281],
[-1.0393, -1.0393, -1.0393],
[-1.6931, -1.6931, -1.6931],
[-1.3982, -1.3982, -1.3982],
[-0.5732, -0.5732, -0.5732],
[-1.0830, -1.0830, -1.0830],
[-1.7561, -1.7561, -1.7561],
[-1.6786, -1.6786, -1.6786],
[-1.6967, -1.6967, -1.6967]],
[[-0.0380, -0.0380, -0.0380],
[-0.1880, -0.1880, -0.1880],
[-1.5724, -1.5724, -1.5724],
[0.6905, 0.6905, 0.6905],
[-0.3190, -0.3190, -0.3190],
[0.7798, 0.7798, 0.7798],
[-0.3693, -0.3693, -0.3693],
[-0.9457, -0.9457, -0.9457],
[-0.2942, -0.2942, -0.2942],
[-1.8527, -1.8527, -1.8527]],
[[0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000]],
[[0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000]],
[[-0.0380, -0.0380, -0.0380],
[-0.1880, -0.1880, -0.1880],
[-1.5724, -1.5724, -1.5724],
[0.6905, 0.6905, 0.6905],
[-0.3190, -0.3190, -0.3190],
[0.7798, 0.7798, 0.7798],
[-0.3693, -0.3693, -0.3693],
[-0.9457, -0.9457, -0.9457],
[-0.2942, -0.2942, -0.2942],
[-1.8527, -1.8527, -1.8527]],
[[1.1773, 1.1773, 1.1773],
[1.5009, 1.5009, 1.5009],
[2.6399, 2.6399, 2.6399],
[5.9242, 5.9242, 5.9242],
[1.0962, 1.0962, 1.0962],
[2.7346, 2.7346, 2.7346],
[6.0865, 6.0865, 6.0865],
[1.5555, 1.5555, 1.5555],
[4.3303, 4.3303, 4.3303],
[2.8229, 2.8229, 2.8229]],
[[-0.0380, -0.0380, -0.0380],
[-0.1880, -0.1880, -0.1880],
[-1.5724, -1.5724, -1.5724],
[0.6905, 0.6905, 0.6905],
[-0.3190, -0.3190, -0.3190],
[0.7798, 0.7798, 0.7798],
[-0.3693, -0.3693, -0.3693],
[-0.9457, -0.9457, -0.9457],
[-0.2942, -0.2942, -0.2942],
[-1.8527, -1.8527,
-1.8527]]]).cuda().float()
expected_output = torch.tensor(
[[[0.5798, 0.5798, 0.5798], [-0.7981, -0.7981, -0.7981],
[-0.9280, -0.9280, -0.9280], [-1.3311, -1.3311, -1.3311],
[1.3687, 1.3687, 1.3687], [0.9277, 0.9277, 0.9277],
[-0.4164, -0.4164, -0.4164], [-1.8274, -1.8274, -1.8274],
[0.9268, 0.9268, 0.9268], [0.8414, 0.8414, 0.8414]],
[[0.0000, 0.0000, 0.0000], [0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000], [0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000], [0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000], [0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000], [0.0000, 0.0000, 0.0000]],
[[0.0000, 0.0000, 0.0000], [0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000], [0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000], [0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000], [0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000], [0.0000, 0.0000, 0.0000]],
[[5.4247, 5.4247, 5.4247], [1.5113, 1.5113, 1.5113],
[2.3944, 2.3944, 2.3944], [1.4740, 1.4740, 1.4740],
[5.0300, 5.0300, 5.0300], [5.1030, 5.1030, 5.1030],
[1.9360, 1.9360, 1.9360], [2.1939, 2.1939, 2.1939],
[2.1581, 2.1581, 2.1581], [3.4666, 3.4666, 3.4666]],
[[0.5798, 0.5798, 0.5798], [-0.7981, -0.7981, -0.7981],
[-0.9280, -0.9280, -0.9280], [-1.3311, -1.3311, -1.3311],
[1.3687, 1.3687, 1.3687], [0.9277, 0.9277, 0.9277],
[-0.4164, -0.4164, -0.4164], [-1.8274, -1.8274, -1.8274],
[0.9268, 0.9268, 0.9268], [0.8414, 0.8414, 0.8414]],
[[-1.6266, -1.6266, -1.6266], [-1.0281, -1.0281, -1.0281],
[-1.0393, -1.0393, -1.0393], [-1.6931, -1.6931, -1.6931],
[-1.3982, -1.3982, -1.3982], [-0.5732, -0.5732, -0.5732],
[-1.0830, -1.0830, -1.0830], [-1.7561, -1.7561, -1.7561],
[-1.6786, -1.6786, -1.6786], [-1.6967, -1.6967, -1.6967]],
[[-0.0380, -0.0380, -0.0380], [-0.1880, -0.1880, -0.1880],
[-1.5724, -1.5724, -1.5724], [0.6905, 0.6905, 0.6905],
[-0.3190, -0.3190, -0.3190], [0.7798, 0.7798, 0.7798],
[-0.3693, -0.3693, -0.3693], [-0.9457, -0.9457, -0.9457],
[-0.2942, -0.2942, -0.2942], [-1.8527, -1.8527, -1.8527]],
[[0.0000, 0.0000, 0.0000], [0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000], [0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000], [0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000], [0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000], [0.0000, 0.0000, 0.0000]],
[[0.0000, 0.0000, 0.0000], [0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000], [0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000], [0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000], [0.0000, 0.0000, 0.0000],
[0.0000, 0.0000, 0.0000], [0.0000, 0.0000, 0.0000]],
[[-0.0380, -0.0380, -0.0380], [-0.1880, -0.1880, -0.1880],
[-1.5724, -1.5724, -1.5724], [0.6905, 0.6905, 0.6905],
[-0.3190, -0.3190, -0.3190], [0.7798, 0.7798, 0.7798],
[-0.3693, -0.3693, -0.3693], [-0.9457, -0.9457, -0.9457],
[-0.2942, -0.2942, -0.2942], [-1.8527, -1.8527, -1.8527]],
[[1.1773, 1.1773, 1.1773], [1.5009, 1.5009, 1.5009],
[2.6399, 2.6399, 2.6399], [5.9242, 5.9242, 5.9242],
[1.0962, 1.0962, 1.0962], [2.7346, 2.7346, 2.7346],
[6.0865, 6.0865, 6.0865], [1.5555, 1.5555, 1.5555],
[4.3303, 4.3303, 4.3303], [2.8229, 2.8229, 2.8229]],
[[-0.0380, -0.0380, -0.0380], [-0.1880, -0.1880, -0.1880],
[-1.5724, -1.5724, -1.5724], [0.6905, 0.6905, 0.6905],
[-0.3190, -0.3190, -0.3190], [0.7798, 0.7798, 0.7798],
[-0.3693, -0.3693, -0.3693], [-0.9457, -0.9457, -0.9457],
[-0.2942, -0.2942, -0.2942], [-1.8527, -1.8527, -1.8527]]],
dtype=dtype).cuda()
assert torch.allclose(output, expected_output)
Loading

0 comments on commit 3bb0611

Please sign in to comment.