Description
Currently trying to export a model from the Torch Points 3D framework (PointNet2) to ONNX to then get to TensorRT where I can load and run inference in C++. There is an unsupported function three_interpolate that would require me to define it as a custom ONNX operator and then I would need to define it as a custom plugin in TensorRT. Originally I attempted to go into a simpler route of setting up Torch Points 3D in the Jetson Xavier but had dependency issues that lead me to consider this as the main option.
Would it be possible to implement this as a custom plugin (three interpolate related functions)?
This is the original error when trying to export:
/venv/lib/python3.8/site-packages/MinkowskiEngine/__init__.py:36: UserWarning: The environment variable `OMP_NUM_THREADS` not set. MinkowskiEngine will automatically set `OMP_NUM_THREADS=16`. If you want to set `OMP_NUM_THREADS` manually, please export it on the command line before running a python script. e.g. `export OMP_NUM_THREADS=12; python your_program.py`. It is recommended to set it below 24.
warnings.warn(
/venv/lib/python3.8/site-packages/hydra/core/utils.py:214: UserWarning:
Using config_path to specify the config name is deprecated, specify the config name via config_name
See https://hydra.cc/docs/next/upgrades/0.11_to_1.0/config_path_changes
warnings.warn(category=UserWarning, message=msg)
[2021-11-30 13:19:59,226][__main__][INFO] - DEVICE : cuda
[2021-11-30 13:19:59,226][torch_points3d.metrics.model_checkpoint][INFO] - Loading checkpoint from /models/pointnet2_charlesssg_GY4_wo_normals_2021-11-15_14-35-40.pt
DATASET PROPS: {'feature_dimension': 0, 'num_classes': 5, 'class_to_segments': {'hotstab': [1], 'wetmate': [2], 'manifold': [3], 'kettlebell': [4]}}
DATA CONFIG: {'class': 'graveyard4.graveyard4Dataset', 'task': 'segmentation', 'dataroot': 'data', 'normal': False, 'first_subsampling': 0.02, 'use_category': True, 'pre_transforms': [{'transform': 'NormalizeScale'}, {'transform': 'GridSampling3D', 'params': {'size': '${data.first_subsampling}'}}], 'train_transforms': [{'transform': 'FixedPoints', 'lparams': [32768]}, {'transform': 'RandomNoise', 'params': {'sigma': 0.01, 'clip': 0.05}}], 'test_transforms': [{'transform': 'FixedPoints', 'lparams': [32768]}], 'val_transforms': '${data.test_transforms}'}
[2021-11-30 13:19:59,436][torch_points3d.models.segmentation.pointnet2][INFO] - Using category information for the predictions with 4 categories
[2021-11-30 13:19:59,474][torch_points3d.metrics.model_checkpoint][INFO] - Available weights : ['latest', 'loss_seg', 'acc', 'macc', 'miou']
[2021-11-30 13:19:59,474][torch_points3d.metrics.model_checkpoint][INFO] - Model loaded from pointnet2_charlesssg_GY4_wo_normals_2021-11-15_14-35-40.pt:best_miou.
[2021-11-30 13:19:59,515][torch_points3d.core.schedulers.bn_schedulers][INFO] - Setting batchnorm momentum at 0.1
[2021-11-30 13:19:59,515][__main__][INFO] - PointNet2_D(
(model): UnetSkipConnectionBlock(
(down): PointNetMSGDown(
(mlps): ModuleList(
(0): MLP2D(
(0): Conv2D(
(0): Conv2d(3, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): LeakyReLU(negative_slope=0.01)
)
(1): Conv2D(
(0): Conv2d(64, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): LeakyReLU(negative_slope=0.01)
)
(2): Conv2D(
(0): Conv2d(64, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): LeakyReLU(negative_slope=0.01)
)
)
)
)
(submodule): UnetSkipConnectionBlock(
(down): PointNetMSGDown(
(mlps): ModuleList(
(0): MLP2D(
(0): Conv2D(
(0): Conv2d(131, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): LeakyReLU(negative_slope=0.01)
)
(1): Conv2D(
(0): Conv2d(128, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): LeakyReLU(negative_slope=0.01)
)
(2): Conv2D(
(0): Conv2d(128, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): LeakyReLU(negative_slope=0.01)
)
)
)
)
(submodule): UnetSkipConnectionBlock(
(inner): GlobalDenseBaseModule: 725248 (aggr=max, MLP2D(
(0): Conv2D(
(0): Conv2d(259, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): LeakyReLU(negative_slope=0.01)
)
(1): Conv2D(
(0): Conv2d(256, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): LeakyReLU(negative_slope=0.01)
)
(2): Conv2D(
(0): Conv2d(512, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): LeakyReLU(negative_slope=0.01)
)
))
(up): DenseFPModule: 394240 (MLP2D(
(0): Conv2D(
(0): Conv2d(1280, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): LeakyReLU(negative_slope=0.01)
)
(1): Conv2D(
(0): Conv2d(256, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): LeakyReLU(negative_slope=0.01)
)
))
)
(up): DenseFPModule: 131840 (MLP2D(
(0): Conv2D(
(0): Conv2d(384, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): LeakyReLU(negative_slope=0.01)
)
(1): Conv2D(
(0): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): LeakyReLU(negative_slope=0.01)
)
))
)
(up): DenseFPModule: 49920 (MLP2D(
(0): Conv2D(
(0): Conv2d(128, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): LeakyReLU(negative_slope=0.01)
)
(1): Conv2D(
(0): Conv2d(128, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): LeakyReLU(negative_slope=0.01)
)
(2): Conv2D(
(0): Conv2d(128, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): LeakyReLU(negative_slope=0.01)
)
))
)
(FC_layer): Seq(
(0): Conv1D(
(0): Conv1d(132, 128, kernel_size=(1,), stride=(1,), bias=False)
(1): BatchNorm1d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): LeakyReLU(negative_slope=0.01)
)
(1): Dropout(p=0.5, inplace=False)
(2): Conv1D(
(0): Conv1d(128, 5, kernel_size=(1,), stride=(1,))
)
)
)
[2021-11-30 13:19:59,518][__main__][INFO] - Model size = 1398981
[2021-11-30 13:19:59,566][__main__][INFO] - Dataset: graveyard4Dataset
e[0;95mtrain_pre_batch_collate_transform e[0m= None
e[0;95mval_pre_batch_collate_transform e[0m= None
e[0;95mtest_pre_batch_collate_transform e[0m= None
e[0;95mpre_transform e[0m= Compose([
NormalizeScale(),
GridSampling3D(grid_size=0.02, quantize_coords=False, mode=mean),
])
e[0;95mtest_transform e[0m= Compose([
FixedPoints(32768, replace=True),
])
e[0;95mtrain_transform e[0m= Compose([
FixedPoints(32768, replace=True),
RandomNoise(sigma=0.01, clip=0.05),
])
e[0;95mval_transform e[0m= Compose([
FixedPoints(32768, replace=True),
])
e[0;95minference_transform e[0m= Compose([
NormalizeScale(),
GridSampling3D(grid_size=0.02, quantize_coords=False, mode=mean),
FixedPoints(32768, replace=True),
])
Size of e[0;95mtrain_dataset e[0m= 4
Size of e[0;95mtest_dataset e[0m= 4
Size of e[0;95mval_dataset e[0m= 4
e[0;95mBatch size =e[0m 16
0%| | 0/1 [00:00<?, ?it/s]DIST SHAPE: torch.Size([4, 512, 3])
IDX SHAPE: torch.Size([4, 512, 3])
DIST SHAPE: torch.Size([4, 32768, 3])
IDX SHAPE: torch.Size([4, 32768, 3])
graph(%pos : Float(4:98304, 32768:3, 3:1, requires_grad=0, device=cuda:0),
%category : Long(4:32768, 32768:1, requires_grad=0, device=cuda:0),
%FC_layer.2.0.weight : Float(5:128, 128:1, 1:1, requires_grad=1, device=cuda:0),
%FC_layer.2.0.bias : Float(5:1, requires_grad=1, device=cuda:0),
%402 : Float(64:3, 3:1, 1:1, 1:1, requires_grad=0, device=cuda:0),
%403 : Float(64:1, requires_grad=0, device=cuda:0),
%405 : Float(64:64, 64:1, 1:1, 1:1, requires_grad=0, device=cuda:0),
%406 : Float(64:1, requires_grad=0, device=cuda:0),
%408 : Float(128:64, 64:1, 1:1, 1:1, requires_grad=0, device=cuda:0),
%409 : Float(128:1, requires_grad=0, device=cuda:0),
%411 : Float(128:131, 131:1, 1:1, 1:1, requires_grad=0, device=cuda:0),
%412 : Float(128:1, requires_grad=0, device=cuda:0),
%414 : Float(128:128, 128:1, 1:1, 1:1, requires_grad=0, device=cuda:0),
%415 : Float(128:1, requires_grad=0, device=cuda:0),
%417 : Float(256:128, 128:1, 1:1, 1:1, requires_grad=0, device=cuda:0),
%418 : Float(256:1, requires_grad=0, device=cuda:0),
%420 : Float(256:259, 259:1, 1:1, 1:1, requires_grad=0, device=cuda:0),
%421 : Float(256:1, requires_grad=0, device=cuda:0),
%423 : Float(512:256, 256:1, 1:1, 1:1, requires_grad=0, device=cuda:0),
%424 : Float(512:1, requires_grad=0, device=cuda:0),
%426 : Float(1024:512, 512:1, 1:1, 1:1, requires_grad=0, device=cuda:0),
%427 : Float(1024:1, requires_grad=0, device=cuda:0),
%429 : Float(256:1280, 1280:1, 1:1, 1:1, requires_grad=0, device=cuda:0),
%430 : Float(256:1, requires_grad=0, device=cuda:0),
%432 : Float(256:256, 256:1, 1:1, 1:1, requires_grad=0, device=cuda:0),
%433 : Float(256:1, requires_grad=0, device=cuda:0),
%435 : Float(256:384, 384:1, 1:1, 1:1, requires_grad=0, device=cuda:0),
%436 : Float(256:1, requires_grad=0, device=cuda:0),
%438 : Float(128:256, 256:1, 1:1, 1:1, requires_grad=0, device=cuda:0),
%439 : Float(128:1, requires_grad=0, device=cuda:0),
%441 : Float(128:128, 128:1, 1:1, 1:1, requires_grad=0, device=cuda:0),
%442 : Float(128:1, requires_grad=0, device=cuda:0),
%444 : Float(128:128, 128:1, 1:1, 1:1, requires_grad=0, device=cuda:0),
%445 : Float(128:1, requires_grad=0, device=cuda:0),
%447 : Float(128:128, 128:1, 1:1, 1:1, requires_grad=0, device=cuda:0),
%448 : Float(128:1, requires_grad=0, device=cuda:0),
%450 : Float(128:132, 132:1, 1:1, requires_grad=0, device=cuda:0),
%451 : Float(128:1, requires_grad=0, device=cuda:0),
%452 : Long(1:1, requires_grad=0, device=cpu),
%453 : Long(1:1, requires_grad=0, device=cpu),
%454 : Long(1:1, requires_grad=0, device=cpu),
%455 : Long(1:1, requires_grad=0, device=cpu),
%456 : Long(1:1, requires_grad=0, device=cpu),
%457 : Long(1:1, requires_grad=0, device=cpu),
%458 : Long(1:1, requires_grad=0, device=cpu),
%459 : Long(1:1, requires_grad=0, device=cpu),
%460 : Long(1:1, requires_grad=0, device=cpu),
%461 : Long(1:1, requires_grad=0, device=cpu),
%462 : Long(1:1, requires_grad=0, device=cpu),
%463 : Long(1:1, requires_grad=0, device=cpu),
%464 : Long(1:1, requires_grad=0, device=cpu),
%465 : Long(1:1, requires_grad=0, device=cpu),
%466 : Long(1:1, requires_grad=0, device=cpu),
%467 : Long(1:1, requires_grad=0, device=cpu),
%468 : Long(1:1, requires_grad=0, device=cpu),
%469 : Long(1:1, requires_grad=0, device=cpu),
%470 : Long(1:1, requires_grad=0, device=cpu),
%471 : Long(1:1, requires_grad=0, device=cpu),
%472 : Long(1:1, requires_grad=0, device=cpu),
%473 : Long(1:1, requires_grad=0, device=cpu),
%474 : Long(1:1, requires_grad=0, device=cpu),
%475 : Long(1:1, requires_grad=0, device=cpu),
%476 : Long(1:1, requires_grad=0, device=cpu),
%477 : Long(1:1, requires_grad=0, device=cpu),
%478 : Long(1:1, requires_grad=0, device=cpu),
%479 : Long(1:1, requires_grad=0, device=cpu),
%480 : Long(1:1, requires_grad=0, device=cpu)):
%107 : Float(4:98304, 32768:3, 3:1, requires_grad=0, device=cuda:0) = onnx::Cast[to=1](%pos) # /workdir/forward_scripts/../torch_points3d/models/segmentation/pointnet2.py:152:0
%108 : Long(4:32768, 32768:1, requires_grad=0, device=cuda:0) = onnx::Cast[to=7](%category) # /workdir/forward_scripts/../torch_points3d/models/segmentation/pointnet2.py:153:0
%109 : Int(4:512, 512:1, 1:1, requires_grad=0, device=cuda:0) = onnx::Constant[value=<Tensor>]()
%110 : Tensor = onnx::Shape(%107)
%111 : Tensor = onnx::Constant[value={2}]()
%112 : Long(device=cpu) = onnx::Gather[axis=0](%110, %111) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:81:0
%117 : Tensor = onnx::Unsqueeze[axes=[0]](%112)
%118 : Tensor = onnx::Concat[axis=0](%452, %453, %117)
%121 : Tensor = onnx::Unsqueeze[axes=[0]](%112)
%122 : Tensor = onnx::Concat[axis=0](%454, %455, %121)
%123 : Tensor = onnx::Shape(%118)
%124 : Tensor = onnx::ConstantOfShape[value={1}](%123)
%125 : Tensor = onnx::Expand(%109, %124)
%126 : Int(4:1536, 512:3, 3:1, requires_grad=0, device=cuda:0) = onnx::Tile(%125, %122) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:81:0
%127 : Long(4:1536, 512:3, 3:1, requires_grad=0, device=cuda:0) = onnx::Cast[to=7](%126) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:81:0
%128 : Tensor = onnx::Constant[value= 0 1 [ CPULongType{2} ]]()
%129 : Tensor = onnx::Constant[value={1}]()
%130 : Tensor = onnx::Shape(%107)
%131 : Tensor = onnx::Gather[axis=0](%130, %129)
%132 : Tensor = onnx::OneHot[axis=1](%127, %131, %128)
%133 : Tensor = onnx::Cast[to=1](%132)
%134 : Tensor = onnx::Unsqueeze[axes=[2]](%107)
%135 : Tensor = onnx::Mul(%134, %133)
%136 : Float(4:1536, 512:3, 3:1, requires_grad=0, device=cuda:0) = onnx::ReduceSum[axes=[1], keepdims=0](%135) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:82:0
%137 : Float(4:98304, 3:32768, 32768:1, requires_grad=0, device=cuda:0) = onnx::Transpose[perm=[0, 2, 1]](%107) # /workdir/forward_scripts/../torch_points3d/modules/pointnet2/dense.py:37:0
%138 : Long(4:32768, 1:32768, 32768:1, requires_grad=0, device=cuda:0) = onnx::Constant[value=<Tensor>]()
%139 : Tensor = onnx::Shape(%137)
%140 : Tensor = onnx::Constant[value={1}]()
%141 : Long(device=cpu) = onnx::Gather[axis=0](%139, %140) # /venv/lib/python3.8/site-packages/torch_points_kernels/torchpoints.py:144:0
%145 : Tensor = onnx::Unsqueeze[axes=[0]](%141)
%147 : Tensor = onnx::Concat[axis=0](%456, %145, %457)
%149 : Tensor = onnx::Unsqueeze[axes=[0]](%141)
%151 : Tensor = onnx::Concat[axis=0](%458, %149, %459)
%152 : Tensor = onnx::Shape(%147)
%153 : Tensor = onnx::ConstantOfShape[value={1}](%152)
%154 : Tensor = onnx::Expand(%138, %153)
%155 : Long(4:98304, 3:32768, 32768:1, requires_grad=0, device=cuda:0) = onnx::Tile(%154, %151) # /venv/lib/python3.8/site-packages/torch_points_kernels/torchpoints.py:144:0
%156 : Tensor = onnx::Constant[value= 0 1 [ CPULongType{2} ]]()
%157 : Tensor = onnx::Constant[value={2}]()
%158 : Tensor = onnx::Shape(%137)
%159 : Tensor = onnx::Gather[axis=0](%158, %157)
%160 : Tensor = onnx::OneHot[axis=2](%155, %159, %156)
%161 : Tensor = onnx::Cast[to=1](%160)
%162 : Tensor = onnx::Unsqueeze[axes=[3]](%137)
%163 : Tensor = onnx::Mul(%162, %161)
%164 : Float(4:98304, 3:32768, 32768:1, requires_grad=0, device=cuda:0) = onnx::ReduceSum[axes=[2], keepdims=0](%163) # /venv/lib/python3.8/site-packages/torch_points_kernels/torchpoints.py:145:0
%166 : Tensor = onnx::Shape(%137)
%167 : Tensor = onnx::Constant[value={1}]()
%168 : Long(device=cpu) = onnx::Gather[axis=0](%166, %167) # /venv/lib/python3.8/site-packages/torch_points_kernels/torchpoints.py:146:0
%172 : Tensor = onnx::Unsqueeze[axes=[0]](%168)
%175 : Tensor = onnx::Concat[axis=0](%460, %172, %461, %462)
%176 : Float(4:98304, 3:32768, 512:64, 64:1, requires_grad=0, device=cuda:0) = onnx::Reshape(%164, %175) # /venv/lib/python3.8/site-packages/torch_points_kernels/torchpoints.py:146:0
%177 : Float(4:1536, 3:1, 512:3, requires_grad=0, device=cuda:0) = onnx::Transpose[perm=[0, 2, 1]](%136) # /workdir/forward_scripts/../torch_points3d/modules/pointnet2/dense.py:39:0
%178 : Float(4:1536, 3:1, 512:3, 1:1, requires_grad=0, device=cuda:0) = onnx::Unsqueeze[axes=[3]](%177) # /workdir/forward_scripts/../torch_points3d/modules/pointnet2/dense.py:39:0
%179 : Float(4:98304, 3:32768, 512:64, 64:1, requires_grad=0, device=cuda:0) = onnx::Sub(%176, %178)
%401 : Float(4:2097152, 64:32768, 512:64, 64:1, requires_grad=0, device=cuda:0) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%179, %402, %403)
%182 : Float(4:2097152, 64:32768, 512:64, 64:1, requires_grad=0, device=cuda:0) = onnx::LeakyRelu[alpha=0.01](%401) # /venv/lib/python3.8/site-packages/torch/nn/functional.py:1309:0
%404 : Float(4:2097152, 64:32768, 512:64, 64:1, requires_grad=0, device=cuda:0) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%182, %405, %406)
%185 : Float(4:2097152, 64:32768, 512:64, 64:1, requires_grad=0, device=cuda:0) = onnx::LeakyRelu[alpha=0.01](%404) # /venv/lib/python3.8/site-packages/torch/nn/functional.py:1309:0
%407 : Float(4:4194304, 128:32768, 512:64, 64:1, requires_grad=0, device=cuda:0) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%185, %408, %409)
%188 : Float(4:4194304, 128:32768, 512:64, 64:1, requires_grad=0, device=cuda:0) = onnx::LeakyRelu[alpha=0.01](%407) # /venv/lib/python3.8/site-packages/torch/nn/functional.py:1309:0
%189 : Float(4:65536, 128:512, 512:1, 1:1, requires_grad=0, device=cuda:0) = onnx::MaxPool[kernel_shape=[1, 64], pads=[0, 0, 0, 0], strides=[1, 64]](%188) # /venv/lib/python3.8/site-packages/torch/nn/functional.py:585:0
%190 : Float(4:65536, 128:512, 512:1, requires_grad=0, device=cuda:0) = onnx::Squeeze[axes=[3]](%189) # /workdir/forward_scripts/../torch_points3d/modules/pointnet2/dense.py:74:0
%191 : Float(4:65536, 128:512, 512:1, requires_grad=0, device=cuda:0) = onnx::Concat[axis=1](%190) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:88:0
%192 : Int(4:128, 128:1, 1:1, requires_grad=0, device=cuda:0) = onnx::Constant[value=<Tensor>]()
%193 : Tensor = onnx::Shape(%136)
%194 : Tensor = onnx::Constant[value={2}]()
%195 : Long(device=cpu) = onnx::Gather[axis=0](%193, %194) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:81:0
%200 : Tensor = onnx::Unsqueeze[axes=[0]](%195)
%201 : Tensor = onnx::Concat[axis=0](%463, %464, %200)
%204 : Tensor = onnx::Unsqueeze[axes=[0]](%195)
%205 : Tensor = onnx::Concat[axis=0](%465, %466, %204)
%206 : Tensor = onnx::Shape(%201)
%207 : Tensor = onnx::ConstantOfShape[value={1}](%206)
%208 : Tensor = onnx::Expand(%192, %207)
%209 : Int(4:384, 128:3, 3:1, requires_grad=0, device=cuda:0) = onnx::Tile(%208, %205) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:81:0
%210 : Long(4:384, 128:3, 3:1, requires_grad=0, device=cuda:0) = onnx::Cast[to=7](%209) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:81:0
%211 : Tensor = onnx::Constant[value= 0 1 [ CPULongType{2} ]]()
%212 : Tensor = onnx::Constant[value={1}]()
%213 : Tensor = onnx::Shape(%136)
%214 : Tensor = onnx::Gather[axis=0](%213, %212)
%215 : Tensor = onnx::OneHot[axis=1](%210, %214, %211)
%216 : Tensor = onnx::Cast[to=1](%215)
%217 : Tensor = onnx::Unsqueeze[axes=[2]](%136)
%218 : Tensor = onnx::Mul(%217, %216)
%219 : Float(4:384, 128:3, 3:1, requires_grad=0, device=cuda:0) = onnx::ReduceSum[axes=[1], keepdims=0](%218) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:82:0
%220 : Float(4:1536, 3:512, 512:1, requires_grad=0, device=cuda:0) = onnx::Transpose[perm=[0, 2, 1]](%136) # /workdir/forward_scripts/../torch_points3d/modules/pointnet2/dense.py:37:0
%221 : Long(4:8192, 1:8192, 8192:1, requires_grad=0, device=cuda:0) = onnx::Constant[value=<Tensor>]()
%222 : Tensor = onnx::Shape(%220)
%223 : Tensor = onnx::Constant[value={1}]()
%224 : Long(device=cpu) = onnx::Gather[axis=0](%222, %223) # /venv/lib/python3.8/site-packages/torch_points_kernels/torchpoints.py:144:0
%228 : Tensor = onnx::Unsqueeze[axes=[0]](%224)
%230 : Tensor = onnx::Concat[axis=0](%467, %228, %468)
%232 : Tensor = onnx::Unsqueeze[axes=[0]](%224)
%234 : Tensor = onnx::Concat[axis=0](%469, %232, %470)
%235 : Tensor = onnx::Shape(%230)
%236 : Tensor = onnx::ConstantOfShape[value={1}](%235)
%237 : Tensor = onnx::Expand(%221, %236)
%238 : Long(4:24576, 3:8192, 8192:1, requires_grad=0, device=cuda:0) = onnx::Tile(%237, %234) # /venv/lib/python3.8/site-packages/torch_points_kernels/torchpoints.py:144:0
%239 : Tensor = onnx::Constant[value= 0 1 [ CPULongType{2} ]]()
%240 : Tensor = onnx::Constant[value={2}]()
%241 : Tensor = onnx::Shape(%220)
%242 : Tensor = onnx::Gather[axis=0](%241, %240)
%243 : Tensor = onnx::OneHot[axis=2](%238, %242, %239)
%244 : Tensor = onnx::Cast[to=1](%243)
%245 : Tensor = onnx::Unsqueeze[axes=[3]](%220)
%246 : Tensor = onnx::Mul(%245, %244)
%247 : Float(4:24576, 3:8192, 8192:1, requires_grad=0, device=cuda:0) = onnx::ReduceSum[axes=[2], keepdims=0](%246) # /venv/lib/python3.8/site-packages/torch_points_kernels/torchpoints.py:145:0
%249 : Tensor = onnx::Shape(%220)
%250 : Tensor = onnx::Constant[value={1}]()
%251 : Long(device=cpu) = onnx::Gather[axis=0](%249, %250) # /venv/lib/python3.8/site-packages/torch_points_kernels/torchpoints.py:146:0
%255 : Tensor = onnx::Unsqueeze[axes=[0]](%251)
%258 : Tensor = onnx::Concat[axis=0](%471, %255, %472, %473)
%259 : Float(4:24576, 3:8192, 128:64, 64:1, requires_grad=0, device=cuda:0) = onnx::Reshape(%247, %258) # /venv/lib/python3.8/site-packages/torch_points_kernels/torchpoints.py:146:0
%260 : Float(4:384, 3:1, 128:3, requires_grad=0, device=cuda:0) = onnx::Transpose[perm=[0, 2, 1]](%219) # /workdir/forward_scripts/../torch_points3d/modules/pointnet2/dense.py:39:0
%261 : Float(4:384, 3:1, 128:3, 1:1, requires_grad=0, device=cuda:0) = onnx::Unsqueeze[axes=[3]](%260) # /workdir/forward_scripts/../torch_points3d/modules/pointnet2/dense.py:39:0
%262 : Float(4:24576, 3:8192, 128:64, 64:1, requires_grad=0, device=cuda:0) = onnx::Sub(%259, %261)
%263 : Long(4:8192, 1:8192, 8192:1, requires_grad=0, device=cuda:0) = onnx::Constant[value=<Tensor>]()
%264 : Tensor = onnx::Shape(%191)
%265 : Tensor = onnx::Constant[value={1}]()
%266 : Long(device=cpu) = onnx::Gather[axis=0](%264, %265) # /venv/lib/python3.8/site-packages/torch_points_kernels/torchpoints.py:144:0
%270 : Tensor = onnx::Unsqueeze[axes=[0]](%266)
%272 : Tensor = onnx::Concat[axis=0](%474, %270, %475)
%274 : Tensor = onnx::Unsqueeze[axes=[0]](%266)
%276 : Tensor = onnx::Concat[axis=0](%476, %274, %477)
%277 : Tensor = onnx::Shape(%272)
%278 : Tensor = onnx::ConstantOfShape[value={1}](%277)
%279 : Tensor = onnx::Expand(%263, %278)
%280 : Long(4:1048576, 128:8192, 8192:1, requires_grad=0, device=cuda:0) = onnx::Tile(%279, %276) # /venv/lib/python3.8/site-packages/torch_points_kernels/torchpoints.py:144:0
%281 : Tensor = onnx::Constant[value= 0 1 [ CPULongType{2} ]]()
%282 : Tensor = onnx::Constant[value={2}]()
%283 : Tensor = onnx::Shape(%191)
%284 : Tensor = onnx::Gather[axis=0](%283, %282)
%285 : Tensor = onnx::OneHot[axis=2](%280, %284, %281)
%286 : Tensor = onnx::Cast[to=1](%285)
%287 : Tensor = onnx::Unsqueeze[axes=[3]](%191)
%288 : Tensor = onnx::Mul(%287, %286)
%289 : Float(4:1048576, 128:8192, 8192:1, requires_grad=0, device=cuda:0) = onnx::ReduceSum[axes=[2], keepdims=0](%288) # /venv/lib/python3.8/site-packages/torch_points_kernels/torchpoints.py:145:0
%291 : Tensor = onnx::Shape(%191)
%292 : Tensor = onnx::Constant[value={1}]()
%293 : Long(device=cpu) = onnx::Gather[axis=0](%291, %292) # /venv/lib/python3.8/site-packages/torch_points_kernels/torchpoints.py:146:0
%297 : Tensor = onnx::Unsqueeze[axes=[0]](%293)
%300 : Tensor = onnx::Concat[axis=0](%478, %297, %479, %480)
%301 : Float(4:1048576, 128:8192, 128:64, 64:1, requires_grad=0, device=cuda:0) = onnx::Reshape(%289, %300) # /venv/lib/python3.8/site-packages/torch_points_kernels/torchpoints.py:146:0
%302 : Float(4:1073152, 131:8192, 128:64, 64:1, requires_grad=0, device=cuda:0) = onnx::Concat[axis=1](%262, %301) # /workdir/forward_scripts/../torch_points3d/modules/pointnet2/dense.py:47:0
%410 : Float(4:1048576, 128:8192, 128:64, 64:1, requires_grad=0, device=cuda:0) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%302, %411, %412)
%305 : Float(4:1048576, 128:8192, 128:64, 64:1, requires_grad=0, device=cuda:0) = onnx::LeakyRelu[alpha=0.01](%410) # /venv/lib/python3.8/site-packages/torch/nn/functional.py:1309:0
%413 : Float(4:1048576, 128:8192, 128:64, 64:1, requires_grad=0, device=cuda:0) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%305, %414, %415)
%308 : Float(4:1048576, 128:8192, 128:64, 64:1, requires_grad=0, device=cuda:0) = onnx::LeakyRelu[alpha=0.01](%413) # /venv/lib/python3.8/site-packages/torch/nn/functional.py:1309:0
%416 : Float(4:2097152, 256:8192, 128:64, 64:1, requires_grad=0, device=cuda:0) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%308, %417, %418)
%311 : Float(4:2097152, 256:8192, 128:64, 64:1, requires_grad=0, device=cuda:0) = onnx::LeakyRelu[alpha=0.01](%416) # /venv/lib/python3.8/site-packages/torch/nn/functional.py:1309:0
%312 : Float(4:32768, 256:128, 128:1, 1:1, requires_grad=0, device=cuda:0) = onnx::MaxPool[kernel_shape=[1, 64], pads=[0, 0, 0, 0], strides=[1, 64]](%311) # /venv/lib/python3.8/site-packages/torch/nn/functional.py:585:0
%313 : Float(4:32768, 256:128, 128:1, requires_grad=0, device=cuda:0) = onnx::Squeeze[axes=[3]](%312) # /workdir/forward_scripts/../torch_points3d/modules/pointnet2/dense.py:74:0
%314 : Float(4:32768, 256:128, 128:1, requires_grad=0, device=cuda:0) = onnx::Concat[axis=1](%313) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:88:0
%315 : Float(4:384, 3:128, 128:1, requires_grad=0, device=cuda:0) = onnx::Transpose[perm=[0, 2, 1]](%219) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:182:0
%316 : Float(4:33152, 259:128, 128:1, requires_grad=0, device=cuda:0) = onnx::Concat[axis=1](%314, %315) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:184:0
%317 : Float(4:33152, 259:128, 128:1, 1:1, requires_grad=0, device=cuda:0) = onnx::Unsqueeze[axes=[3]](%316) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:184:0
%419 : Float(4:32768, 256:128, 128:1, 1:1, requires_grad=0, device=cuda:0) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%317, %420, %421)
%320 : Float(4:32768, 256:128, 128:1, 1:1, requires_grad=0, device=cuda:0) = onnx::LeakyRelu[alpha=0.01](%419) # /venv/lib/python3.8/site-packages/torch/nn/functional.py:1309:0
%422 : Float(4:65536, 512:128, 128:1, 1:1, requires_grad=0, device=cuda:0) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%320, %423, %424)
%323 : Float(4:65536, 512:128, 128:1, 1:1, requires_grad=0, device=cuda:0) = onnx::LeakyRelu[alpha=0.01](%422) # /venv/lib/python3.8/site-packages/torch/nn/functional.py:1309:0
%425 : Float(4:131072, 1024:128, 128:1, 1:1, requires_grad=0, device=cuda:0) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%323, %426, %427)
%326 : Float(4:131072, 1024:128, 128:1, 1:1, requires_grad=0, device=cuda:0) = onnx::LeakyRelu[alpha=0.01](%425) # /venv/lib/python3.8/site-packages/torch/nn/functional.py:1309:0
%327 : Float(4:131072, 1024:128, 128:1, requires_grad=0, device=cuda:0) = onnx::Squeeze[axes=[3]](%326) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:187:0
%328 : Float(4:1024, 1024:1, requires_grad=0, device=cuda:0) = onnx::ReduceMax[axes=[-1], keepdims=0](%327) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:187:0
%329 : Float(4:1024, 1024:1, 1:1, requires_grad=0, device=cuda:0) = onnx::Unsqueeze[axes=[2]](%328) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:194:0
%330 : Tensor = onnx::Shape(%329)
%331 : Tensor = onnx::Constant[value={0}]()
%332 : Long(device=cpu) = onnx::Gather[axis=0](%330, %331) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:153:0
%333 : Tensor = onnx::Shape(%329)
%334 : Tensor = onnx::Constant[value={1}]()
%335 : Long(device=cpu) = onnx::Gather[axis=0](%333, %334) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:153:0
%336 : Tensor = onnx::Shape(%219)
%337 : Tensor = onnx::Constant[value={1}]()
%338 : Long(device=cpu) = onnx::Gather[axis=0](%336, %337) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:153:0
%339 : Tensor = onnx::Unsqueeze[axes=[0]](%332)
%340 : Tensor = onnx::Unsqueeze[axes=[0]](%335)
%341 : Tensor = onnx::Unsqueeze[axes=[0]](%338)
%342 : Tensor = onnx::Concat[axis=0](%339, %340, %341)
%343 : Tensor = onnx::Constant[value={-1}]()
%344 : Tensor = onnx::Reshape(%342, %343)
%345 : Tensor = onnx::Shape(%344)
%346 : Tensor = onnx::ConstantOfShape[value={1}](%345)
%347 : Long(requires_grad=0, device=cpu) = onnx::Constant[value={-1}]()
%348 : LongTensor = onnx::Mul(%346, %347)
%349 : Tensor = onnx::Equal(%344, %348)
%350 : Tensor = onnx::Cast[to=9](%349)
%351 : Tensor = onnx::Where(%350, %346, %344)
%352 : Float(4:1024, 1024:1, 128:0, requires_grad=0, device=cuda:0) = onnx::Expand(%329, %351) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:153:0
%353 : Float(4:163840, 1280:128, 128:1, requires_grad=0, device=cuda:0) = onnx::Concat[axis=1](%352, %314) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:122:0
%354 : Float(4:163840, 1280:128, 128:1, 1:1, requires_grad=0, device=cuda:0) = onnx::Unsqueeze[axes=[3]](%353) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:124:0
%428 : Float(4:32768, 256:128, 128:1, 1:1, requires_grad=0, device=cuda:0) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%354, %429, %430)
%357 : Float(4:32768, 256:128, 128:1, 1:1, requires_grad=0, device=cuda:0) = onnx::LeakyRelu[alpha=0.01](%428) # /venv/lib/python3.8/site-packages/torch/nn/functional.py:1309:0
%431 : Float(4:32768, 256:128, 128:1, 1:1, requires_grad=0, device=cuda:0) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%357, %432, %433)
%360 : Float(4:32768, 256:128, 128:1, 1:1, requires_grad=0, device=cuda:0) = onnx::LeakyRelu[alpha=0.01](%431) # /venv/lib/python3.8/site-packages/torch/nn/functional.py:1309:0
%361 : Float(4:32768, 256:128, 128:1, requires_grad=0, device=cuda:0) = onnx::Squeeze[axes=[3]](%360) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:129:0
%362 : Int(4:1536, 512:3, 3:1, requires_grad=0, device=cuda:0) = onnx::Constant[value=<Tensor>]()
%363 : Float(4:1536, 512:3, 3:1, requires_grad=0, device=cuda:0) = onnx::Constant[value=<Tensor>]()
%364 : Float(4:131072, 256:512, 512:1, requires_grad=0, device=cuda:0) = ^ThreeInterpolate()(%361, %362, %363) # /venv/lib/python3.8/site-packages/torch_points_kernels/torchpoints.py:126:0
%365 : Float(4:196608, 384:512, 512:1, requires_grad=0, device=cuda:0) = onnx::Concat[axis=1](%364, %191) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:122:0
%366 : Float(4:196608, 384:512, 512:1, 1:1, requires_grad=0, device=cuda:0) = onnx::Unsqueeze[axes=[3]](%365) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:124:0
%434 : Float(4:131072, 256:512, 512:1, 1:1, requires_grad=0, device=cuda:0) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%366, %435, %436)
%369 : Float(4:131072, 256:512, 512:1, 1:1, requires_grad=0, device=cuda:0) = onnx::LeakyRelu[alpha=0.01](%434) # /venv/lib/python3.8/site-packages/torch/nn/functional.py:1309:0
%437 : Float(4:65536, 128:512, 512:1, 1:1, requires_grad=0, device=cuda:0) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%369, %438, %439)
%372 : Float(4:65536, 128:512, 512:1, 1:1, requires_grad=0, device=cuda:0) = onnx::LeakyRelu[alpha=0.01](%437) # /venv/lib/python3.8/site-packages/torch/nn/functional.py:1309:0
%373 : Float(4:65536, 128:512, 512:1, requires_grad=0, device=cuda:0) = onnx::Squeeze[axes=[3]](%372) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:129:0
%374 : Int(4:98304, 32768:3, 3:1, requires_grad=0, device=cuda:0) = onnx::Constant[value=<Tensor>]()
%375 : Float(4:98304, 32768:3, 3:1, requires_grad=0, device=cuda:0) = onnx::Constant[value=<Tensor>]()
%376 : Float(4:4194304, 128:32768, 32768:1, requires_grad=0, device=cuda:0) = ^ThreeInterpolate()(%373, %374, %375) # /venv/lib/python3.8/site-packages/torch_points_kernels/torchpoints.py:126:0
%377 : Float(4:4194304, 128:32768, 32768:1, 1:1, requires_grad=0, device=cuda:0) = onnx::Unsqueeze[axes=[3]](%376) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:124:0
%440 : Float(4:4194304, 128:32768, 32768:1, 1:1, requires_grad=0, device=cuda:0) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%377, %441, %442)
%380 : Float(4:4194304, 128:32768, 32768:1, 1:1, requires_grad=0, device=cuda:0) = onnx::LeakyRelu[alpha=0.01](%440) # /venv/lib/python3.8/site-packages/torch/nn/functional.py:1309:0
%443 : Float(4:4194304, 128:32768, 32768:1, 1:1, requires_grad=0, device=cuda:0) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%380, %444, %445)
%383 : Float(4:4194304, 128:32768, 32768:1, 1:1, requires_grad=0, device=cuda:0) = onnx::LeakyRelu[alpha=0.01](%443) # /venv/lib/python3.8/site-packages/torch/nn/functional.py:1309:0
%446 : Float(4:4194304, 128:32768, 32768:1, 1:1, requires_grad=0, device=cuda:0) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%383, %447, %448)
%386 : Float(4:4194304, 128:32768, 32768:1, 1:1, requires_grad=0, device=cuda:0) = onnx::LeakyRelu[alpha=0.01](%446) # /venv/lib/python3.8/site-packages/torch/nn/functional.py:1309:0
%387 : Float(4:4194304, 128:32768, 32768:1, requires_grad=0, device=cuda:0) = onnx::Squeeze[axes=[3]](%386) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:129:0
%388 : Tensor = onnx::Constant[value= 0 1 [ CPULongType{2} ]]()
%389 : Tensor = onnx::Constant[value={4}]()
%390 : Long(4:131072, 32768:4, 4:1, requires_grad=0, device=cuda:0) = onnx::OneHot[axis=-1](%108, %389, %388) # /workdir/forward_scripts/../torch_points3d/models/segmentation/pointnet2.py:236:0
%391 : Float(4:131072, 32768:4, 4:1, requires_grad=0, device=cuda:0) = onnx::Cast[to=1](%390) # /workdir/forward_scripts/../torch_points3d/models/segmentation/pointnet2.py:236:0
%392 : Float(4:131072, 4:1, 32768:4, requires_grad=0, device=cuda:0) = onnx::Transpose[perm=[0, 2, 1]](%391) # /workdir/forward_scripts/../torch_points3d/models/segmentation/pointnet2.py:236:0
%393 : Float(4:4325376, 132:32768, 32768:1, requires_grad=0, device=cuda:0) = onnx::Concat[axis=1](%387, %392) # /workdir/forward_scripts/../torch_points3d/models/segmentation/pointnet2.py:237:0
%449 : Float(4:4194304, 128:32768, 32768:1, requires_grad=0, device=cuda:0) = onnx::Conv[dilations=[1], group=1, kernel_shape=[1], pads=[0, 0], strides=[1]](%393, %450, %451)
%396 : Float(4:4194304, 128:32768, 32768:1, requires_grad=0, device=cuda:0) = onnx::LeakyRelu[alpha=0.01](%449) # /venv/lib/python3.8/site-packages/torch/nn/functional.py:983:0
%397 : Float(4:163840, 5:32768, 32768:1, requires_grad=0, device=cuda:0) = onnx::Conv[dilations=[1], group=1, kernel_shape=[1], pads=[0, 0], strides=[1]](%396, %FC_layer.2.0.weight, %FC_layer.2.0.bias) # /venv/lib/python3.8/site-packages/torch/nn/modules/conv.py:258:0
%398 : Float(4:163840, 32768:5, 5:1, requires_grad=0, device=cuda:0) = onnx::Transpose[perm=[0, 2, 1]](%397) # /workdir/forward_scripts/../torch_points3d/models/segmentation/pointnet2.py:239:0
%399 : Tensor = onnx::Constant[value=-1 5 [ CPULongType{2} ]]()
%output_labels : Float(131072:5, 5:1, requires_grad=0, device=cuda:0) = onnx::Reshape(%398, %399) # /workdir/forward_scripts/../torch_points3d/models/segmentation/pointnet2.py:239:0
return (%output_labels)
/venv/lib/python3.8/site-packages/torch/onnx/symbolic_opset9.py:1703: UserWarning: ONNX export unsqueeze with negative axis -1 might cause the onnx model to be incorrect. Negative axis is not supported in ONNX. Axis is converted to 3 based on input shape at export time. Passing an tensor of different rank in execution will be incorrect.
warnings.warn("ONNX export unsqueeze with negative axis " + str(dim) +
/venv/lib/python3.8/site-packages/torch/onnx/symbolic_opset9.py:572: UserWarning: ONNX export squeeze with negative axis -1 might cause the onnx model to be incorrect. Negative axis is not supported in ONNX. Axis is converted to 3 based on input shape at export time. Passing an tensor of different rank in execution will be incorrect.
warnings.warn("ONNX export squeeze with negative axis " + str(squeeze_dim) +
/venv/lib/python3.8/site-packages/torch/onnx/symbolic_opset9.py:598: UserWarning: This model contains a squeeze operation on dimension 3. If the model is intended to be used with dynamic input shapes, please use opset version 11 to export the model.
warnings.warn("This model contains a squeeze operation on dimension " + str(squeeze_dim) + ". If the model is " +
/venv/lib/python3.8/site-packages/torch/onnx/symbolic_opset9.py:1703: UserWarning: ONNX export unsqueeze with negative axis -1 might cause the onnx model to be incorrect. Negative axis is not supported in ONNX. Axis is converted to 2 based on input shape at export time. Passing an tensor of different rank in execution will be incorrect.
warnings.warn("ONNX export unsqueeze with negative axis " + str(dim) +
0%| | 0/1 [00:00<?, ?it/s]
Traceback (most recent call last):
File "forward_scripts/checkpoint_export2.py", line 137, in <module>
main()
File "/venv/lib/python3.8/site-packages/hydra/main.py", line 32, in decorated_main
_run_hydra(
File "/venv/lib/python3.8/site-packages/hydra/_internal/utils.py", line 346, in _run_hydra
run_and_report(
File "/venv/lib/python3.8/site-packages/hydra/_internal/utils.py", line 201, in run_and_report
raise ex
File "/venv/lib/python3.8/site-packages/hydra/_internal/utils.py", line 198, in run_and_report
return func()
File "/venv/lib/python3.8/site-packages/hydra/_internal/utils.py", line 347, in <lambda>
lambda: hydra.run(
File "/venv/lib/python3.8/site-packages/hydra/_internal/hydra.py", line 107, in run
return run_job(
File "/venv/lib/python3.8/site-packages/hydra/core/utils.py", line 129, in run_job
ret.return_value = task_function(task_cfg)
File "forward_scripts/checkpoint_export2.py", line 133, in main
run(model, dataset, device, cfg.output_path)
File "forward_scripts/checkpoint_export2.py", line 66, in run
torch.onnx.export(model,
File "/venv/lib/python3.8/site-packages/torch/onnx/__init__.py", line 225, in export
return utils.export(model, args, f, export_params, verbose, training,
File "/venv/lib/python3.8/site-packages/torch/onnx/utils.py", line 85, in export
_export(model, args, f, export_params, verbose, training, input_names, output_names,
File "/venv/lib/python3.8/site-packages/torch/onnx/utils.py", line 647, in _export
proto, export_map = graph._export_onnx(
RuntimeError: ONNX export failed: Couldn't export Python operator ThreeInterpolate
Defined at:
/venv/lib/python3.8/site-packages/torch_points_kernels/torchpoints.py(126): three_interpolate
/workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py(151): conv
/workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py(119): forward
/venv/lib/python3.8/site-packages/torch/nn/modules/module.py(709): _slow_forward
/venv/lib/python3.8/site-packages/torch/nn/modules/module.py(725): _call_impl
/workdir/forward_scripts/../torch_points3d/models/base_architectures/unet.py(306): forward
/venv/lib/python3.8/site-packages/torch/nn/modules/module.py(709): _slow_forward
/venv/lib/python3.8/site-packages/torch/nn/modules/module.py(725): _call_impl
/workdir/forward_scripts/../torch_points3d/models/base_architectures/unet.py(304): forward
/venv/lib/python3.8/site-packages/torch/nn/modules/module.py(709): _slow_forward
/venv/lib/python3.8/site-packages/torch/nn/modules/module.py(725): _call_impl
/workdir/forward_scripts/../torch_points3d/models/segmentation/pointnet2.py(233): forward
/venv/lib/python3.8/site-packages/torch/nn/modules/module.py(709): _slow_forward
/venv/lib/python3.8/site-packages/torch/nn/modules/module.py(725): _call_impl
/venv/lib/python3.8/site-packages/torch/jit/_trace.py(116): wrapper
/venv/lib/python3.8/site-packages/torch/jit/_trace.py(125): forward
/venv/lib/python3.8/site-packages/torch/nn/modules/module.py(727): _call_impl
/venv/lib/python3.8/site-packages/torch/jit/_trace.py(1148): _get_trace_graph
/venv/lib/python3.8/site-packages/torch/onnx/utils.py(342): _trace_and_get_graph_from_model
/venv/lib/python3.8/site-packages/torch/onnx/utils.py(379): _create_jit_graph
/venv/lib/python3.8/site-packages/torch/onnx/utils.py(409): _model_to_graph
/venv/lib/python3.8/site-packages/torch/onnx/utils.py(632): _export
/venv/lib/python3.8/site-packages/torch/onnx/utils.py(85): export
/venv/lib/python3.8/site-packages/torch/onnx/__init__.py(225): export
forward_scripts/checkpoint_export2.py(66): run
forward_scripts/checkpoint_export2.py(133): main
/venv/lib/python3.8/site-packages/hydra/core/utils.py(129): run_job
/venv/lib/python3.8/site-packages/hydra/_internal/hydra.py(107): run
/venv/lib/python3.8/site-packages/hydra/_internal/utils.py(347): <lambda>
/venv/lib/python3.8/site-packages/hydra/_internal/utils.py(198): run_and_report
/venv/lib/python3.8/site-packages/hydra/_internal/utils.py(346): _run_hydra
/venv/lib/python3.8/site-packages/hydra/main.py(32): decorated_main
forward_scripts/checkpoint_export2.py(137): <module>
Graph we tried to export:
graph(%pos : Float(4:98304, 32768:3, 3:1, requires_grad=0, device=cuda:0),
%category : Long(4:32768, 32768:1, requires_grad=0, device=cuda:0),
%FC_layer.2.0.weight : Float(5:128, 128:1, 1:1, requires_grad=1, device=cuda:0),
%FC_layer.2.0.bias : Float(5:1, requires_grad=1, device=cuda:0),
%402 : Float(64:3, 3:1, 1:1, 1:1, requires_grad=0, device=cuda:0),
%403 : Float(64:1, requires_grad=0, device=cuda:0),
%405 : Float(64:64, 64:1, 1:1, 1:1, requires_grad=0, device=cuda:0),
%406 : Float(64:1, requires_grad=0, device=cuda:0),
%408 : Float(128:64, 64:1, 1:1, 1:1, requires_grad=0, device=cuda:0),
%409 : Float(128:1, requires_grad=0, device=cuda:0),
%411 : Float(128:131, 131:1, 1:1, 1:1, requires_grad=0, device=cuda:0),
%412 : Float(128:1, requires_grad=0, device=cuda:0),
%414 : Float(128:128, 128:1, 1:1, 1:1, requires_grad=0, device=cuda:0),
%415 : Float(128:1, requires_grad=0, device=cuda:0),
%417 : Float(256:128, 128:1, 1:1, 1:1, requires_grad=0, device=cuda:0),
%418 : Float(256:1, requires_grad=0, device=cuda:0),
%420 : Float(256:259, 259:1, 1:1, 1:1, requires_grad=0, device=cuda:0),
%421 : Float(256:1, requires_grad=0, device=cuda:0),
%423 : Float(512:256, 256:1, 1:1, 1:1, requires_grad=0, device=cuda:0),
%424 : Float(512:1, requires_grad=0, device=cuda:0),
%426 : Float(1024:512, 512:1, 1:1, 1:1, requires_grad=0, device=cuda:0),
%427 : Float(1024:1, requires_grad=0, device=cuda:0),
%429 : Float(256:1280, 1280:1, 1:1, 1:1, requires_grad=0, device=cuda:0),
%430 : Float(256:1, requires_grad=0, device=cuda:0),
%432 : Float(256:256, 256:1, 1:1, 1:1, requires_grad=0, device=cuda:0),
%433 : Float(256:1, requires_grad=0, device=cuda:0),
%435 : Float(256:384, 384:1, 1:1, 1:1, requires_grad=0, device=cuda:0),
%436 : Float(256:1, requires_grad=0, device=cuda:0),
%438 : Float(128:256, 256:1, 1:1, 1:1, requires_grad=0, device=cuda:0),
%439 : Float(128:1, requires_grad=0, device=cuda:0),
%441 : Float(128:128, 128:1, 1:1, 1:1, requires_grad=0, device=cuda:0),
%442 : Float(128:1, requires_grad=0, device=cuda:0),
%444 : Float(128:128, 128:1, 1:1, 1:1, requires_grad=0, device=cuda:0),
%445 : Float(128:1, requires_grad=0, device=cuda:0),
%447 : Float(128:128, 128:1, 1:1, 1:1, requires_grad=0, device=cuda:0),
%448 : Float(128:1, requires_grad=0, device=cuda:0),
%450 : Float(128:132, 132:1, 1:1, requires_grad=0, device=cuda:0),
%451 : Float(128:1, requires_grad=0, device=cuda:0),
%452 : Long(1:1, requires_grad=0, device=cpu),
%453 : Long(1:1, requires_grad=0, device=cpu),
%454 : Long(1:1, requires_grad=0, device=cpu),
%455 : Long(1:1, requires_grad=0, device=cpu),
%456 : Long(1:1, requires_grad=0, device=cpu),
%457 : Long(1:1, requires_grad=0, device=cpu),
%458 : Long(1:1, requires_grad=0, device=cpu),
%459 : Long(1:1, requires_grad=0, device=cpu),
%460 : Long(1:1, requires_grad=0, device=cpu),
%461 : Long(1:1, requires_grad=0, device=cpu),
%462 : Long(1:1, requires_grad=0, device=cpu),
%463 : Long(1:1, requires_grad=0, device=cpu),
%464 : Long(1:1, requires_grad=0, device=cpu),
%465 : Long(1:1, requires_grad=0, device=cpu),
%466 : Long(1:1, requires_grad=0, device=cpu),
%467 : Long(1:1, requires_grad=0, device=cpu),
%468 : Long(1:1, requires_grad=0, device=cpu),
%469 : Long(1:1, requires_grad=0, device=cpu),
%470 : Long(1:1, requires_grad=0, device=cpu),
%471 : Long(1:1, requires_grad=0, device=cpu),
%472 : Long(1:1, requires_grad=0, device=cpu),
%473 : Long(1:1, requires_grad=0, device=cpu),
%474 : Long(1:1, requires_grad=0, device=cpu),
%475 : Long(1:1, requires_grad=0, device=cpu),
%476 : Long(1:1, requires_grad=0, device=cpu),
%477 : Long(1:1, requires_grad=0, device=cpu),
%478 : Long(1:1, requires_grad=0, device=cpu),
%479 : Long(1:1, requires_grad=0, device=cpu),
%480 : Long(1:1, requires_grad=0, device=cpu)):
%107 : Float(4:98304, 32768:3, 3:1, requires_grad=0, device=cuda:0) = onnx::Cast[to=1](%pos) # /workdir/forward_scripts/../torch_points3d/models/segmentation/pointnet2.py:152:0
%108 : Long(4:32768, 32768:1, requires_grad=0, device=cuda:0) = onnx::Cast[to=7](%category) # /workdir/forward_scripts/../torch_points3d/models/segmentation/pointnet2.py:153:0
%109 : Int(4:512, 512:1, 1:1, requires_grad=0, device=cuda:0) = onnx::Constant[value=<Tensor>]()
%110 : Tensor = onnx::Shape(%107)
%111 : Tensor = onnx::Constant[value={2}]()
%112 : Long(device=cpu) = onnx::Gather[axis=0](%110, %111) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:81:0
%117 : Tensor = onnx::Unsqueeze[axes=[0]](%112)
%118 : Tensor = onnx::Concat[axis=0](%452, %453, %117)
%121 : Tensor = onnx::Unsqueeze[axes=[0]](%112)
%122 : Tensor = onnx::Concat[axis=0](%454, %455, %121)
%123 : Tensor = onnx::Shape(%118)
%124 : Tensor = onnx::ConstantOfShape[value={1}](%123)
%125 : Tensor = onnx::Expand(%109, %124)
%126 : Int(4:1536, 512:3, 3:1, requires_grad=0, device=cuda:0) = onnx::Tile(%125, %122) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:81:0
%127 : Long(4:1536, 512:3, 3:1, requires_grad=0, device=cuda:0) = onnx::Cast[to=7](%126) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:81:0
%128 : Tensor = onnx::Constant[value= 0 1 [ CPULongType{2} ]]()
%129 : Tensor = onnx::Constant[value={1}]()
%130 : Tensor = onnx::Shape(%107)
%131 : Tensor = onnx::Gather[axis=0](%130, %129)
%132 : Tensor = onnx::OneHot[axis=1](%127, %131, %128)
%133 : Tensor = onnx::Cast[to=1](%132)
%134 : Tensor = onnx::Unsqueeze[axes=[2]](%107)
%135 : Tensor = onnx::Mul(%134, %133)
%136 : Float(4:1536, 512:3, 3:1, requires_grad=0, device=cuda:0) = onnx::ReduceSum[axes=[1], keepdims=0](%135) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:82:0
%137 : Float(4:98304, 3:32768, 32768:1, requires_grad=0, device=cuda:0) = onnx::Transpose[perm=[0, 2, 1]](%107) # /workdir/forward_scripts/../torch_points3d/modules/pointnet2/dense.py:37:0
%138 : Long(4:32768, 1:32768, 32768:1, requires_grad=0, device=cuda:0) = onnx::Constant[value=<Tensor>]()
%139 : Tensor = onnx::Shape(%137)
%140 : Tensor = onnx::Constant[value={1}]()
%141 : Long(device=cpu) = onnx::Gather[axis=0](%139, %140) # /venv/lib/python3.8/site-packages/torch_points_kernels/torchpoints.py:144:0
%145 : Tensor = onnx::Unsqueeze[axes=[0]](%141)
%147 : Tensor = onnx::Concat[axis=0](%456, %145, %457)
%149 : Tensor = onnx::Unsqueeze[axes=[0]](%141)
%151 : Tensor = onnx::Concat[axis=0](%458, %149, %459)
%152 : Tensor = onnx::Shape(%147)
%153 : Tensor = onnx::ConstantOfShape[value={1}](%152)
%154 : Tensor = onnx::Expand(%138, %153)
%155 : Long(4:98304, 3:32768, 32768:1, requires_grad=0, device=cuda:0) = onnx::Tile(%154, %151) # /venv/lib/python3.8/site-packages/torch_points_kernels/torchpoints.py:144:0
%156 : Tensor = onnx::Constant[value= 0 1 [ CPULongType{2} ]]()
%157 : Tensor = onnx::Constant[value={2}]()
%158 : Tensor = onnx::Shape(%137)
%159 : Tensor = onnx::Gather[axis=0](%158, %157)
%160 : Tensor = onnx::OneHot[axis=2](%155, %159, %156)
%161 : Tensor = onnx::Cast[to=1](%160)
%162 : Tensor = onnx::Unsqueeze[axes=[3]](%137)
%163 : Tensor = onnx::Mul(%162, %161)
%164 : Float(4:98304, 3:32768, 32768:1, requires_grad=0, device=cuda:0) = onnx::ReduceSum[axes=[2], keepdims=0](%163) # /venv/lib/python3.8/site-packages/torch_points_kernels/torchpoints.py:145:0
%166 : Tensor = onnx::Shape(%137)
%167 : Tensor = onnx::Constant[value={1}]()
%168 : Long(device=cpu) = onnx::Gather[axis=0](%166, %167) # /venv/lib/python3.8/site-packages/torch_points_kernels/torchpoints.py:146:0
%172 : Tensor = onnx::Unsqueeze[axes=[0]](%168)
%175 : Tensor = onnx::Concat[axis=0](%460, %172, %461, %462)
%176 : Float(4:98304, 3:32768, 512:64, 64:1, requires_grad=0, device=cuda:0) = onnx::Reshape(%164, %175) # /venv/lib/python3.8/site-packages/torch_points_kernels/torchpoints.py:146:0
%177 : Float(4:1536, 3:1, 512:3, requires_grad=0, device=cuda:0) = onnx::Transpose[perm=[0, 2, 1]](%136) # /workdir/forward_scripts/../torch_points3d/modules/pointnet2/dense.py:39:0
%178 : Float(4:1536, 3:1, 512:3, 1:1, requires_grad=0, device=cuda:0) = onnx::Unsqueeze[axes=[3]](%177) # /workdir/forward_scripts/../torch_points3d/modules/pointnet2/dense.py:39:0
%179 : Float(4:98304, 3:32768, 512:64, 64:1, requires_grad=0, device=cuda:0) = onnx::Sub(%176, %178)
%401 : Float(4:2097152, 64:32768, 512:64, 64:1, requires_grad=0, device=cuda:0) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%179, %402, %403)
%182 : Float(4:2097152, 64:32768, 512:64, 64:1, requires_grad=0, device=cuda:0) = onnx::LeakyRelu[alpha=0.01](%401) # /venv/lib/python3.8/site-packages/torch/nn/functional.py:1309:0
%404 : Float(4:2097152, 64:32768, 512:64, 64:1, requires_grad=0, device=cuda:0) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%182, %405, %406)
%185 : Float(4:2097152, 64:32768, 512:64, 64:1, requires_grad=0, device=cuda:0) = onnx::LeakyRelu[alpha=0.01](%404) # /venv/lib/python3.8/site-packages/torch/nn/functional.py:1309:0
%407 : Float(4:4194304, 128:32768, 512:64, 64:1, requires_grad=0, device=cuda:0) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%185, %408, %409)
%188 : Float(4:4194304, 128:32768, 512:64, 64:1, requires_grad=0, device=cuda:0) = onnx::LeakyRelu[alpha=0.01](%407) # /venv/lib/python3.8/site-packages/torch/nn/functional.py:1309:0
%189 : Float(4:65536, 128:512, 512:1, 1:1, requires_grad=0, device=cuda:0) = onnx::MaxPool[kernel_shape=[1, 64], pads=[0, 0, 0, 0], strides=[1, 64]](%188) # /venv/lib/python3.8/site-packages/torch/nn/functional.py:585:0
%190 : Float(4:65536, 128:512, 512:1, requires_grad=0, device=cuda:0) = onnx::Squeeze[axes=[3]](%189) # /workdir/forward_scripts/../torch_points3d/modules/pointnet2/dense.py:74:0
%191 : Float(4:65536, 128:512, 512:1, requires_grad=0, device=cuda:0) = onnx::Concat[axis=1](%190) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:88:0
%192 : Int(4:128, 128:1, 1:1, requires_grad=0, device=cuda:0) = onnx::Constant[value=<Tensor>]()
%193 : Tensor = onnx::Shape(%136)
%194 : Tensor = onnx::Constant[value={2}]()
%195 : Long(device=cpu) = onnx::Gather[axis=0](%193, %194) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:81:0
%200 : Tensor = onnx::Unsqueeze[axes=[0]](%195)
%201 : Tensor = onnx::Concat[axis=0](%463, %464, %200)
%204 : Tensor = onnx::Unsqueeze[axes=[0]](%195)
%205 : Tensor = onnx::Concat[axis=0](%465, %466, %204)
%206 : Tensor = onnx::Shape(%201)
%207 : Tensor = onnx::ConstantOfShape[value={1}](%206)
%208 : Tensor = onnx::Expand(%192, %207)
%209 : Int(4:384, 128:3, 3:1, requires_grad=0, device=cuda:0) = onnx::Tile(%208, %205) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:81:0
%210 : Long(4:384, 128:3, 3:1, requires_grad=0, device=cuda:0) = onnx::Cast[to=7](%209) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:81:0
%211 : Tensor = onnx::Constant[value= 0 1 [ CPULongType{2} ]]()
%212 : Tensor = onnx::Constant[value={1}]()
%213 : Tensor = onnx::Shape(%136)
%214 : Tensor = onnx::Gather[axis=0](%213, %212)
%215 : Tensor = onnx::OneHot[axis=1](%210, %214, %211)
%216 : Tensor = onnx::Cast[to=1](%215)
%217 : Tensor = onnx::Unsqueeze[axes=[2]](%136)
%218 : Tensor = onnx::Mul(%217, %216)
%219 : Float(4:384, 128:3, 3:1, requires_grad=0, device=cuda:0) = onnx::ReduceSum[axes=[1], keepdims=0](%218) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:82:0
%220 : Float(4:1536, 3:512, 512:1, requires_grad=0, device=cuda:0) = onnx::Transpose[perm=[0, 2, 1]](%136) # /workdir/forward_scripts/../torch_points3d/modules/pointnet2/dense.py:37:0
%221 : Long(4:8192, 1:8192, 8192:1, requires_grad=0, device=cuda:0) = onnx::Constant[value=<Tensor>]()
%222 : Tensor = onnx::Shape(%220)
%223 : Tensor = onnx::Constant[value={1}]()
%224 : Long(device=cpu) = onnx::Gather[axis=0](%222, %223) # /venv/lib/python3.8/site-packages/torch_points_kernels/torchpoints.py:144:0
%228 : Tensor = onnx::Unsqueeze[axes=[0]](%224)
%230 : Tensor = onnx::Concat[axis=0](%467, %228, %468)
%232 : Tensor = onnx::Unsqueeze[axes=[0]](%224)
%234 : Tensor = onnx::Concat[axis=0](%469, %232, %470)
%235 : Tensor = onnx::Shape(%230)
%236 : Tensor = onnx::ConstantOfShape[value={1}](%235)
%237 : Tensor = onnx::Expand(%221, %236)
%238 : Long(4:24576, 3:8192, 8192:1, requires_grad=0, device=cuda:0) = onnx::Tile(%237, %234) # /venv/lib/python3.8/site-packages/torch_points_kernels/torchpoints.py:144:0
%239 : Tensor = onnx::Constant[value= 0 1 [ CPULongType{2} ]]()
%240 : Tensor = onnx::Constant[value={2}]()
%241 : Tensor = onnx::Shape(%220)
%242 : Tensor = onnx::Gather[axis=0](%241, %240)
%243 : Tensor = onnx::OneHot[axis=2](%238, %242, %239)
%244 : Tensor = onnx::Cast[to=1](%243)
%245 : Tensor = onnx::Unsqueeze[axes=[3]](%220)
%246 : Tensor = onnx::Mul(%245, %244)
%247 : Float(4:24576, 3:8192, 8192:1, requires_grad=0, device=cuda:0) = onnx::ReduceSum[axes=[2], keepdims=0](%246) # /venv/lib/python3.8/site-packages/torch_points_kernels/torchpoints.py:145:0
%249 : Tensor = onnx::Shape(%220)
%250 : Tensor = onnx::Constant[value={1}]()
%251 : Long(device=cpu) = onnx::Gather[axis=0](%249, %250) # /venv/lib/python3.8/site-packages/torch_points_kernels/torchpoints.py:146:0
%255 : Tensor = onnx::Unsqueeze[axes=[0]](%251)
%258 : Tensor = onnx::Concat[axis=0](%471, %255, %472, %473)
%259 : Float(4:24576, 3:8192, 128:64, 64:1, requires_grad=0, device=cuda:0) = onnx::Reshape(%247, %258) # /venv/lib/python3.8/site-packages/torch_points_kernels/torchpoints.py:146:0
%260 : Float(4:384, 3:1, 128:3, requires_grad=0, device=cuda:0) = onnx::Transpose[perm=[0, 2, 1]](%219) # /workdir/forward_scripts/../torch_points3d/modules/pointnet2/dense.py:39:0
%261 : Float(4:384, 3:1, 128:3, 1:1, requires_grad=0, device=cuda:0) = onnx::Unsqueeze[axes=[3]](%260) # /workdir/forward_scripts/../torch_points3d/modules/pointnet2/dense.py:39:0
%262 : Float(4:24576, 3:8192, 128:64, 64:1, requires_grad=0, device=cuda:0) = onnx::Sub(%259, %261)
%263 : Long(4:8192, 1:8192, 8192:1, requires_grad=0, device=cuda:0) = onnx::Constant[value=<Tensor>]()
%264 : Tensor = onnx::Shape(%191)
%265 : Tensor = onnx::Constant[value={1}]()
%266 : Long(device=cpu) = onnx::Gather[axis=0](%264, %265) # /venv/lib/python3.8/site-packages/torch_points_kernels/torchpoints.py:144:0
%270 : Tensor = onnx::Unsqueeze[axes=[0]](%266)
%272 : Tensor = onnx::Concat[axis=0](%474, %270, %475)
%274 : Tensor = onnx::Unsqueeze[axes=[0]](%266)
%276 : Tensor = onnx::Concat[axis=0](%476, %274, %477)
%277 : Tensor = onnx::Shape(%272)
%278 : Tensor = onnx::ConstantOfShape[value={1}](%277)
%279 : Tensor = onnx::Expand(%263, %278)
%280 : Long(4:1048576, 128:8192, 8192:1, requires_grad=0, device=cuda:0) = onnx::Tile(%279, %276) # /venv/lib/python3.8/site-packages/torch_points_kernels/torchpoints.py:144:0
%281 : Tensor = onnx::Constant[value= 0 1 [ CPULongType{2} ]]()
%282 : Tensor = onnx::Constant[value={2}]()
%283 : Tensor = onnx::Shape(%191)
%284 : Tensor = onnx::Gather[axis=0](%283, %282)
%285 : Tensor = onnx::OneHot[axis=2](%280, %284, %281)
%286 : Tensor = onnx::Cast[to=1](%285)
%287 : Tensor = onnx::Unsqueeze[axes=[3]](%191)
%288 : Tensor = onnx::Mul(%287, %286)
%289 : Float(4:1048576, 128:8192, 8192:1, requires_grad=0, device=cuda:0) = onnx::ReduceSum[axes=[2], keepdims=0](%288) # /venv/lib/python3.8/site-packages/torch_points_kernels/torchpoints.py:145:0
%291 : Tensor = onnx::Shape(%191)
%292 : Tensor = onnx::Constant[value={1}]()
%293 : Long(device=cpu) = onnx::Gather[axis=0](%291, %292) # /venv/lib/python3.8/site-packages/torch_points_kernels/torchpoints.py:146:0
%297 : Tensor = onnx::Unsqueeze[axes=[0]](%293)
%300 : Tensor = onnx::Concat[axis=0](%478, %297, %479, %480)
%301 : Float(4:1048576, 128:8192, 128:64, 64:1, requires_grad=0, device=cuda:0) = onnx::Reshape(%289, %300) # /venv/lib/python3.8/site-packages/torch_points_kernels/torchpoints.py:146:0
%302 : Float(4:1073152, 131:8192, 128:64, 64:1, requires_grad=0, device=cuda:0) = onnx::Concat[axis=1](%262, %301) # /workdir/forward_scripts/../torch_points3d/modules/pointnet2/dense.py:47:0
%410 : Float(4:1048576, 128:8192, 128:64, 64:1, requires_grad=0, device=cuda:0) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%302, %411, %412)
%305 : Float(4:1048576, 128:8192, 128:64, 64:1, requires_grad=0, device=cuda:0) = onnx::LeakyRelu[alpha=0.01](%410) # /venv/lib/python3.8/site-packages/torch/nn/functional.py:1309:0
%413 : Float(4:1048576, 128:8192, 128:64, 64:1, requires_grad=0, device=cuda:0) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%305, %414, %415)
%308 : Float(4:1048576, 128:8192, 128:64, 64:1, requires_grad=0, device=cuda:0) = onnx::LeakyRelu[alpha=0.01](%413) # /venv/lib/python3.8/site-packages/torch/nn/functional.py:1309:0
%416 : Float(4:2097152, 256:8192, 128:64, 64:1, requires_grad=0, device=cuda:0) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%308, %417, %418)
%311 : Float(4:2097152, 256:8192, 128:64, 64:1, requires_grad=0, device=cuda:0) = onnx::LeakyRelu[alpha=0.01](%416) # /venv/lib/python3.8/site-packages/torch/nn/functional.py:1309:0
%312 : Float(4:32768, 256:128, 128:1, 1:1, requires_grad=0, device=cuda:0) = onnx::MaxPool[kernel_shape=[1, 64], pads=[0, 0, 0, 0], strides=[1, 64]](%311) # /venv/lib/python3.8/site-packages/torch/nn/functional.py:585:0
%313 : Float(4:32768, 256:128, 128:1, requires_grad=0, device=cuda:0) = onnx::Squeeze[axes=[3]](%312) # /workdir/forward_scripts/../torch_points3d/modules/pointnet2/dense.py:74:0
%314 : Float(4:32768, 256:128, 128:1, requires_grad=0, device=cuda:0) = onnx::Concat[axis=1](%313) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:88:0
%315 : Float(4:384, 3:128, 128:1, requires_grad=0, device=cuda:0) = onnx::Transpose[perm=[0, 2, 1]](%219) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:182:0
%316 : Float(4:33152, 259:128, 128:1, requires_grad=0, device=cuda:0) = onnx::Concat[axis=1](%314, %315) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:184:0
%317 : Float(4:33152, 259:128, 128:1, 1:1, requires_grad=0, device=cuda:0) = onnx::Unsqueeze[axes=[3]](%316) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:184:0
%419 : Float(4:32768, 256:128, 128:1, 1:1, requires_grad=0, device=cuda:0) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%317, %420, %421)
%320 : Float(4:32768, 256:128, 128:1, 1:1, requires_grad=0, device=cuda:0) = onnx::LeakyRelu[alpha=0.01](%419) # /venv/lib/python3.8/site-packages/torch/nn/functional.py:1309:0
%422 : Float(4:65536, 512:128, 128:1, 1:1, requires_grad=0, device=cuda:0) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%320, %423, %424)
%323 : Float(4:65536, 512:128, 128:1, 1:1, requires_grad=0, device=cuda:0) = onnx::LeakyRelu[alpha=0.01](%422) # /venv/lib/python3.8/site-packages/torch/nn/functional.py:1309:0
%425 : Float(4:131072, 1024:128, 128:1, 1:1, requires_grad=0, device=cuda:0) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%323, %426, %427)
%326 : Float(4:131072, 1024:128, 128:1, 1:1, requires_grad=0, device=cuda:0) = onnx::LeakyRelu[alpha=0.01](%425) # /venv/lib/python3.8/site-packages/torch/nn/functional.py:1309:0
%327 : Float(4:131072, 1024:128, 128:1, requires_grad=0, device=cuda:0) = onnx::Squeeze[axes=[3]](%326) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:187:0
%328 : Float(4:1024, 1024:1, requires_grad=0, device=cuda:0) = onnx::ReduceMax[axes=[-1], keepdims=0](%327) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:187:0
%329 : Float(4:1024, 1024:1, 1:1, requires_grad=0, device=cuda:0) = onnx::Unsqueeze[axes=[2]](%328) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:194:0
%330 : Tensor = onnx::Shape(%329)
%331 : Tensor = onnx::Constant[value={0}]()
%332 : Long(device=cpu) = onnx::Gather[axis=0](%330, %331) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:153:0
%333 : Tensor = onnx::Shape(%329)
%334 : Tensor = onnx::Constant[value={1}]()
%335 : Long(device=cpu) = onnx::Gather[axis=0](%333, %334) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:153:0
%336 : Tensor = onnx::Shape(%219)
%337 : Tensor = onnx::Constant[value={1}]()
%338 : Long(device=cpu) = onnx::Gather[axis=0](%336, %337) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:153:0
%339 : Tensor = onnx::Unsqueeze[axes=[0]](%332)
%340 : Tensor = onnx::Unsqueeze[axes=[0]](%335)
%341 : Tensor = onnx::Unsqueeze[axes=[0]](%338)
%342 : Tensor = onnx::Concat[axis=0](%339, %340, %341)
%343 : Tensor = onnx::Constant[value={-1}]()
%344 : Tensor = onnx::Reshape(%342, %343)
%345 : Tensor = onnx::Shape(%344)
%346 : Tensor = onnx::ConstantOfShape[value={1}](%345)
%347 : Long(requires_grad=0, device=cpu) = onnx::Constant[value={-1}]()
%348 : LongTensor = onnx::Mul(%346, %347)
%349 : Tensor = onnx::Equal(%344, %348)
%350 : Tensor = onnx::Cast[to=9](%349)
%351 : Tensor = onnx::Where(%350, %346, %344)
%352 : Float(4:1024, 1024:1, 128:0, requires_grad=0, device=cuda:0) = onnx::Expand(%329, %351) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:153:0
%353 : Float(4:163840, 1280:128, 128:1, requires_grad=0, device=cuda:0) = onnx::Concat[axis=1](%352, %314) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:122:0
%354 : Float(4:163840, 1280:128, 128:1, 1:1, requires_grad=0, device=cuda:0) = onnx::Unsqueeze[axes=[3]](%353) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:124:0
%428 : Float(4:32768, 256:128, 128:1, 1:1, requires_grad=0, device=cuda:0) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%354, %429, %430)
%357 : Float(4:32768, 256:128, 128:1, 1:1, requires_grad=0, device=cuda:0) = onnx::LeakyRelu[alpha=0.01](%428) # /venv/lib/python3.8/site-packages/torch/nn/functional.py:1309:0
%431 : Float(4:32768, 256:128, 128:1, 1:1, requires_grad=0, device=cuda:0) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%357, %432, %433)
%360 : Float(4:32768, 256:128, 128:1, 1:1, requires_grad=0, device=cuda:0) = onnx::LeakyRelu[alpha=0.01](%431) # /venv/lib/python3.8/site-packages/torch/nn/functional.py:1309:0
%361 : Float(4:32768, 256:128, 128:1, requires_grad=0, device=cuda:0) = onnx::Squeeze[axes=[3]](%360) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:129:0
%362 : Int(4:1536, 512:3, 3:1, requires_grad=0, device=cuda:0) = onnx::Constant[value=<Tensor>]()
%363 : Float(4:1536, 512:3, 3:1, requires_grad=0, device=cuda:0) = onnx::Constant[value=<Tensor>]()
%364 : Float(4:131072, 256:512, 512:1, requires_grad=0, device=cuda:0) = ^ThreeInterpolate()(%361, %362, %363) # /venv/lib/python3.8/site-packages/torch_points_kernels/torchpoints.py:126:0
%365 : Float(4:196608, 384:512, 512:1, requires_grad=0, device=cuda:0) = onnx::Concat[axis=1](%364, %191) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:122:0
%366 : Float(4:196608, 384:512, 512:1, 1:1, requires_grad=0, device=cuda:0) = onnx::Unsqueeze[axes=[3]](%365) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:124:0
%434 : Float(4:131072, 256:512, 512:1, 1:1, requires_grad=0, device=cuda:0) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%366, %435, %436)
%369 : Float(4:131072, 256:512, 512:1, 1:1, requires_grad=0, device=cuda:0) = onnx::LeakyRelu[alpha=0.01](%434) # /venv/lib/python3.8/site-packages/torch/nn/functional.py:1309:0
%437 : Float(4:65536, 128:512, 512:1, 1:1, requires_grad=0, device=cuda:0) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%369, %438, %439)
%372 : Float(4:65536, 128:512, 512:1, 1:1, requires_grad=0, device=cuda:0) = onnx::LeakyRelu[alpha=0.01](%437) # /venv/lib/python3.8/site-packages/torch/nn/functional.py:1309:0
%373 : Float(4:65536, 128:512, 512:1, requires_grad=0, device=cuda:0) = onnx::Squeeze[axes=[3]](%372) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:129:0
%374 : Int(4:98304, 32768:3, 3:1, requires_grad=0, device=cuda:0) = onnx::Constant[value=<Tensor>]()
%375 : Float(4:98304, 32768:3, 3:1, requires_grad=0, device=cuda:0) = onnx::Constant[value=<Tensor>]()
%376 : Float(4:4194304, 128:32768, 32768:1, requires_grad=0, device=cuda:0) = ^ThreeInterpolate()(%373, %374, %375) # /venv/lib/python3.8/site-packages/torch_points_kernels/torchpoints.py:126:0
%377 : Float(4:4194304, 128:32768, 32768:1, 1:1, requires_grad=0, device=cuda:0) = onnx::Unsqueeze[axes=[3]](%376) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:124:0
%440 : Float(4:4194304, 128:32768, 32768:1, 1:1, requires_grad=0, device=cuda:0) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%377, %441, %442)
%380 : Float(4:4194304, 128:32768, 32768:1, 1:1, requires_grad=0, device=cuda:0) = onnx::LeakyRelu[alpha=0.01](%440) # /venv/lib/python3.8/site-packages/torch/nn/functional.py:1309:0
%443 : Float(4:4194304, 128:32768, 32768:1, 1:1, requires_grad=0, device=cuda:0) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%380, %444, %445)
%383 : Float(4:4194304, 128:32768, 32768:1, 1:1, requires_grad=0, device=cuda:0) = onnx::LeakyRelu[alpha=0.01](%443) # /venv/lib/python3.8/site-packages/torch/nn/functional.py:1309:0
%446 : Float(4:4194304, 128:32768, 32768:1, 1:1, requires_grad=0, device=cuda:0) = onnx::Conv[dilations=[1, 1], group=1, kernel_shape=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1]](%383, %447, %448)
%386 : Float(4:4194304, 128:32768, 32768:1, 1:1, requires_grad=0, device=cuda:0) = onnx::LeakyRelu[alpha=0.01](%446) # /venv/lib/python3.8/site-packages/torch/nn/functional.py:1309:0
%387 : Float(4:4194304, 128:32768, 32768:1, requires_grad=0, device=cuda:0) = onnx::Squeeze[axes=[3]](%386) # /workdir/forward_scripts/../torch_points3d/core/base_conv/dense.py:129:0
%388 : Tensor = onnx::Constant[value= 0 1 [ CPULongType{2} ]]()
%389 : Tensor = onnx::Constant[value={4}]()
%390 : Long(4:131072, 32768:4, 4:1, requires_grad=0, device=cuda:0) = onnx::OneHot[axis=-1](%108, %389, %388) # /workdir/forward_scripts/../torch_points3d/models/segmentation/pointnet2.py:236:0
%391 : Float(4:131072, 32768:4, 4:1, requires_grad=0, device=cuda:0) = onnx::Cast[to=1](%390) # /workdir/forward_scripts/../torch_points3d/models/segmentation/pointnet2.py:236:0
%392 : Float(4:131072, 4:1, 32768:4, requires_grad=0, device=cuda:0) = onnx::Transpose[perm=[0, 2, 1]](%391) # /workdir/forward_scripts/../torch_points3d/models/segmentation/pointnet2.py:236:0
%393 : Float(4:4325376, 132:32768, 32768:1, requires_grad=0, device=cuda:0) = onnx::Concat[axis=1](%387, %392) # /workdir/forward_scripts/../torch_points3d/models/segmentation/pointnet2.py:237:0
%449 : Float(4:4194304, 128:32768, 32768:1, requires_grad=0, device=cuda:0) = onnx::Conv[dilations=[1], group=1, kernel_shape=[1], pads=[0, 0], strides=[1]](%393, %450, %451)
%396 : Float(4:4194304, 128:32768, 32768:1, requires_grad=0, device=cuda:0) = onnx::LeakyRelu[alpha=0.01](%449) # /venv/lib/python3.8/site-packages/torch/nn/functional.py:983:0
%397 : Float(4:163840, 5:32768, 32768:1, requires_grad=0, device=cuda:0) = onnx::Conv[dilations=[1], group=1, kernel_shape=[1], pads=[0, 0], strides=[1]](%396, %FC_layer.2.0.weight, %FC_layer.2.0.bias) # /venv/lib/python3.8/site-packages/torch/nn/modules/conv.py:258:0
%398 : Float(4:163840, 32768:5, 5:1, requires_grad=0, device=cuda:0) = onnx::Transpose[perm=[0, 2, 1]](%397) # /workdir/forward_scripts/../torch_points3d/models/segmentation/pointnet2.py:239:0
%399 : Tensor = onnx::Constant[value=-1 5 [ CPULongType{2} ]]()
%output_labels : Float(131072:5, 5:1, requires_grad=0, device=cuda:0) = onnx::Reshape(%398, %399) # /workdir/forward_scripts/../torch_points3d/models/segmentation/pointnet2.py:239:0
return (%output_labels)
Environment
For Training/Export:
Python – 3.8
PyTorch – 1.7.0
Docker Container in an Ubuntu System (from this dockerfile torch-points3d/Dockerfile.gpu at master · nicolas-chaulet/torch-points3d · GitHub, which is pretty much the work environment of the torch points 3d developers)
Other Steps:
ONNX to TensorRT Serialization - Jetson Xavier
Inference Baseline - Jetson Xavier (ubuntu arm64/aarch64)
Model Trained - PointNet++ from the torch points 3d framework (GitHub - nicolas-chaulet/torch-points3d: Pytorch framework for doing deep learning on point clouds.)
Thanks,
Mark