Skip to content

Navigation Menu

Sign in
Appearance settings

Search code, repositories, users, issues, pull requests...

Provide feedback

We read every piece of feedback, and take your input very seriously.

Saved searches

Use saved searches to filter your results more quickly
Sign up
Appearance settings
This repository was archived by the owner on May 29, 2023. It is now read-only.

Commit 29f0829

Browse files
authored
SparseConv from Open3D (#29)
1 parent 18edc45 commit 29f0829

File tree

18 files changed

+1066
-125
lines changed

18 files changed

+1066
-125
lines changed

‎.github/workflows/main.yml‎

Lines changed: 4 additions & 7 deletions
Original file line numberDiff line numberDiff line change
@@ -12,7 +12,7 @@ on:
1212

1313
env:
1414
OPENVINO_VERSION: 2021年4月2日
15-
VERSION: 2021年4月2日.3
15+
VERSION: 2021年4月2日.4
1616
DIST_VERSION: 20214752
1717
DIST_WIN: https://registrationcenter-download.intel.com/akdlm/irc_nas/18320/w_openvino_toolkit_p_20214752.exe
1818
DIST_MAC: https://registrationcenter-download.intel.com/akdlm/irc_nas/18317/m_openvino_toolkit_p_20214752.dmg
@@ -134,9 +134,6 @@ jobs:
134134
test_lnx:
135135
needs: build_lnx
136136
runs-on: ubuntu-18.04
137-
strategy:
138-
matrix:
139-
torch-version: [1.7.0, 1.10.0]
140137

141138
steps:
142139
- uses: actions/checkout@v2
@@ -149,7 +146,7 @@ jobs:
149146
run: |
150147
sudo apt-get install -y python3-setuptools libopencv-dev
151148
python3 -m pip install --upgrade pip
152-
python3 -m pip install torch==${{ matrix.torch-version }} torchvision
149+
python3 -m pip install -r tests/requirements.txt
153150
python3 -m pip install -U protobuf
154151
python3 -m pip install openvino-dev[onnx]==${{env.OPENVINO_VERSION}}
155152
@@ -161,7 +158,7 @@ jobs:
161158
162159
- name: Test
163160
run: |
164-
python3 -m unittest run_tests.py
161+
python3 -m pytest tests/run_tests.py
165162
166163
publish:
167164
if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
@@ -177,4 +174,4 @@ jobs:
177174
run: |
178175
python3 -m pip install --upgrade pip
179176
python3 -m pip install twine
180-
python3 -m twine upload wheel*/*.whl --skip-existing
177+
python3 -m twine upload --repository testpypi wheel*/*.whl --skip-existing

‎README.md‎

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -6,6 +6,7 @@ Repository with guides to enable some layers from PyTorch in Intel OpenVINO:
66
* [torch.fft](examples/fft)
77
* [nn.functional.grid_sample](https://github.com/dkurt/openvino_pytorch_layers/tree/master/examples/grid_sample)
88
* [torchvision.ops.DeformConv2d](examples/deformable_conv)
9+
* [SparseConv](examples/sparse_conv) from [Open3D](https://github.com/isl-org/Open3D)
910

1011

1112
## OpenVINO Model Optimizer extension
Lines changed: 24 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,24 @@
1+
import torch
2+
import torch.nn as nn
3+
import torch.nn.functional as F
4+
5+
class CalculateGrid(torch.autograd.Function):
6+
@staticmethod
7+
def symbolic(g, in_positions):
8+
return g.op("org.open3d::calculate_grid", in_positions)
9+
10+
@staticmethod
11+
def forward(self, in_positions):
12+
filter = torch.Tensor([[-1, -1, -1], [-1, -1, 0], [-1, 0, -1], [-1, 0, 0],
13+
[0, -1, -1], [0, -1, 0], [0, 0, -1],
14+
[0, 0, 0]]).to(in_positions.device)
15+
16+
out_pos = in_positions.long().repeat(1, filter.shape[0]).reshape(-1, 3)
17+
filter = filter.repeat(in_positions.shape[0], 1)
18+
19+
out_pos = out_pos + filter
20+
out_pos = out_pos[out_pos.min(1).values >= 0]
21+
out_pos = out_pos[(~((out_pos.long() % 2).bool()).any(1))]
22+
out_pos = torch.unique(out_pos, dim=0)
23+
24+
return out_pos + 0.5
Lines changed: 40 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,40 @@
1+
import numpy as np
2+
import argparse
3+
import torch
4+
import torch.nn as nn
5+
from torch.autograd import Variable
6+
from .calculate_grid import CalculateGrid
7+
8+
9+
class MyModel(nn.Module):
10+
def __init__(self):
11+
super(MyModel, self).__init__()
12+
self.calculate_grid = CalculateGrid()
13+
14+
def forward(self, x):
15+
return self.calculate_grid.apply(x)
16+
17+
18+
def export(num_points, max_grid_extent):
19+
# Generate a list of unique positions and add a mantissa
20+
np.random.seed(32)
21+
torch.manual_seed(11)
22+
23+
inp_pos = np.random.randint(0, max_grid_extent, [num_points, 3])
24+
inp_pos = torch.tensor(inp_pos) + torch.rand(inp_pos.shape, dtype=torch.float32) # [0, 1)
25+
26+
model = MyModel()
27+
with torch.no_grad():
28+
torch.onnx.export(model, (inp_pos), 'model.onnx',
29+
input_names=['input'],
30+
output_names=['output'],
31+
operator_export_type=torch.onnx.OperatorExportTypes.ONNX_ATEN_FALLBACK)
32+
33+
ref = model(inp_pos).detach().numpy()
34+
35+
# Pad values with espetial end line (-1, 0, 0) and zeros
36+
ref = np.concatenate((ref, [[-1, 0, 0]]))
37+
ref = np.pad(ref, ((0, inp_pos.shape[0] - ref.shape[0]), (0, 0)))
38+
39+
np.save('inp', inp_pos.detach().numpy())
40+
np.save('ref', ref)

‎examples/sparse_conv/export_model.py‎

Lines changed: 69 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,69 @@
1+
import numpy as np
2+
import argparse
3+
import torch
4+
import torch.nn as nn
5+
from torch.autograd import Variable
6+
from .sparse_conv import SparseConvONNX, SparseConvTransposeONNX
7+
8+
9+
def export(num_inp_points, num_out_points, max_grid_extent, in_channels,
10+
filters, kernel_size, normalize, transpose):
11+
np.random.seed(324)
12+
torch.manual_seed(32)
13+
14+
if transpose:
15+
sparse_conv = SparseConvTransposeONNX(in_channels=in_channels,
16+
filters=filters,
17+
kernel_size=kernel_size,
18+
use_bias=False,
19+
normalize=False)
20+
else:
21+
sparse_conv = SparseConvONNX(in_channels=in_channels,
22+
filters=filters,
23+
kernel_size=kernel_size,
24+
use_bias=False,
25+
normalize=False)
26+
27+
# Generate a list of unique positions and add a mantissa
28+
def gen_pos(num_points):
29+
inp_pos = np.random.randint(0, max_grid_extent, [num_points, 3])
30+
inp_pos = np.unique(inp_pos, axis=0).astype(np.float32)
31+
inp_pos = torch.tensor(inp_pos) + torch.rand(inp_pos.shape, dtype=torch.float32) # [0, 1)
32+
return inp_pos
33+
34+
inp_pos = gen_pos(num_inp_points)
35+
out_pos = gen_pos(num_out_points) if num_out_points else inp_pos
36+
37+
features = torch.randn([inp_pos.shape[0], in_channels])
38+
39+
voxel_size = torch.tensor(1.0)
40+
sparse_conv.eval()
41+
42+
new_kernel = torch.randn(sparse_conv.state_dict()["kernel"].shape)
43+
sparse_conv.load_state_dict({"kernel": new_kernel,
44+
"offset": sparse_conv.state_dict()["offset"]})
45+
46+
with torch.no_grad():
47+
torch.onnx.export(sparse_conv, (features, inp_pos, out_pos, voxel_size), 'model.onnx',
48+
input_names=['input', 'input1', 'input2', 'voxel_size'],
49+
output_names=['output'],
50+
operator_export_type=torch.onnx.OperatorExportTypes.ONNX_ATEN_FALLBACK)
51+
52+
ref = sparse_conv(features, inp_pos, out_pos, voxel_size)
53+
np.save('inp', features.detach().numpy())
54+
np.save('inp1', inp_pos.detach().numpy())
55+
np.save('inp2', out_pos.detach().numpy())
56+
np.save('ref', ref.detach().numpy())
57+
58+
59+
if __name__ == "__main__":
60+
parser = argparse.ArgumentParser(description='Generate ONNX model and test data')
61+
parser.add_argument('--num_points', type=int)
62+
parser.add_argument('--max_grid_extent', type=int)
63+
parser.add_argument('--in_channels', type=int)
64+
parser.add_argument('--filters', type=int)
65+
parser.add_argument('--kernel_size', type=int)
66+
args = parser.parse_args()
67+
68+
export(args.num_points, args.max_grid_extent,
69+
args.in_channels, args.filters, args.kernel_size)

‎examples/sparse_conv/sparse_conv.py‎

Lines changed: 55 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,55 @@
1+
import torch
2+
import torch.nn as nn
3+
import torch.nn.functional as F
4+
from open3d.ml.torch.layers import SparseConv, SparseConvTranspose
5+
6+
class SparseConvFunc(torch.autograd.Function):
7+
@staticmethod
8+
def symbolic(g, cls, feat, in_pos, out_pos, voxel_size):
9+
kernel = cls.state_dict()["kernel"]
10+
offset = cls.state_dict()["offset"]
11+
kernel = g.op("Constant", value_t=kernel)
12+
offset = g.op("Constant", value_t=offset)
13+
return g.op("org.open3d::SparseConv", feat, in_pos, out_pos, kernel, offset)
14+
15+
@staticmethod
16+
def forward(self, cls, feat, in_pos, out_pos, voxel_size):
17+
return cls.origin_forward(feat, in_pos, out_pos, voxel_size)
18+
19+
20+
class SparseConvONNX(SparseConv):
21+
"""
22+
This is a support class which helps export network with SparseConv in ONNX format.
23+
"""
24+
def __init__(self, *args, **kwargs):
25+
super().__init__(*args, **kwargs)
26+
self.origin_forward = super().forward
27+
28+
def forward(self, feat, in_pos, out_pos, voxel_size):
29+
return SparseConvFunc.apply(self, feat, in_pos, out_pos, voxel_size)
30+
31+
32+
class SparseConvTransposeFunc(torch.autograd.Function):
33+
@staticmethod
34+
def symbolic(g, cls, feat, in_pos, out_pos, voxel_size):
35+
kernel = cls.state_dict()["kernel"]
36+
offset = cls.state_dict()["offset"]
37+
kernel = g.op("Constant", value_t=kernel)
38+
offset = g.op("Constant", value_t=offset)
39+
return g.op("org.open3d::SparseConvTranspose", feat, in_pos, out_pos, kernel, offset)
40+
41+
@staticmethod
42+
def forward(self, cls, feat, in_pos, out_pos, voxel_size):
43+
return cls.origin_forward(feat, in_pos, out_pos, voxel_size)
44+
45+
46+
class SparseConvTransposeONNX(SparseConvTranspose):
47+
"""
48+
This is a support class which helps export network with SparseConvTranspose in ONNX format.
49+
"""
50+
def __init__(self, *args, **kwargs):
51+
super().__init__(*args, **kwargs)
52+
self.origin_forward = super().forward
53+
54+
def forward(self, feat, in_pos, out_pos, voxel_size):
55+
return SparseConvTransposeFunc.apply(self, feat, in_pos, out_pos, voxel_size)

‎run_tests.py‎

Lines changed: 0 additions & 118 deletions
This file was deleted.

‎tests/requirements.txt‎

Lines changed: 5 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,5 @@
1+
torch==1.8.1
2+
torchvision==0.9.1
3+
open3d==0.14.1
4+
tensorboard
5+
pytest

0 commit comments

Comments
(0)

AltStyle によって変換されたページ (->オリジナル) /