Fix all warnings in pt1.6 (#72)

* Fix (torch.nonzero): Fix  warning of torch.nonzero and bug of nms_iou

* Fix contiguous warning of tensor

* Update CI with pt1.6

* update build action

* recover ci

* Add force cuda

* Use mmcv-full==1.0.5

* Fix nonzero warning

* Update CI config

* update links

* fix unit tests

* fix unit tests

* Lock pytorch version in CI

* Lock pytorch version in CI

* Lock pytorch version in CI

* Try specify torchlink

* Try specify torchlink

* Only add 3.7 for CI

* fix synctax error

* Use 1.3.0

* Use 1.3.0 and specify mmcv

* give up 1.3

* fix cublas issue

* fix cublas issue

.github/workflows/build.yml

@@ -38,13 +38,19 @@ jobs:
     strategy:
       matrix:
         python-version: [3.6, 3.7]
-        torch: [1.3.0, 1.5.0]
+        torch: [1.3.0, 1.5.0+cu101, 1.6.0+cu101]
         include:
           - torch: 1.3.0
             torchvision: 0.4.2
+            mmcv: 1.3.0+cu101
             cuda_arch: "6.0"
-          - torch: 1.5.0
-            torchvision: 0.6.0
+          - torch: 1.5.0+cu101
+            torchvision: 0.6.0+cu101
+            mmcv: 1.5.0+cu101
+            cuda_arch: "7.0"
+          - torch: 1.6.0+cu101
+            mmcv: 1.6.0+cu101
+            torchvision: 0.7.0+cu101
             cuda_arch: "7.0"
 
     steps:
@@ -63,6 +69,8 @@ jobs:
           sudo apt update -qq
           sudo apt install -y cuda-${CUDA_SHORT/./-} cuda-cufft-dev-${CUDA_SHORT/./-}
           sudo apt clean
+          sudo cp /usr/local/cuda-10.2/include/* /usr/local/cuda/include
+          sudo cp -r /usr/local/cuda-10.2/lib64/* /usr/local/cuda/lib64/
           export CUDA_HOME=/usr/local/cuda-${CUDA_SHORT}
           export LD_LIBRARY_PATH=${CUDA_HOME}/lib64:${CUDA_HOME}/include:${LD_LIBRARY_PATH}
           export PATH=${CUDA_HOME}/bin:${PATH}
@@ -74,7 +82,7 @@ jobs:
         run: pip install torch==${{matrix.torch}} torchvision==${{matrix.torchvision}} -f https://download.pytorch.org/whl/torch_stable.html
       - name: Install mmdet3d dependencies
         run: |
-          pip install mmcv-full==latest+torch${{matrix.torch}}+cu101 -f https://openmmlab.oss-accelerate.aliyuncs.com/mmcv/dist/index.html
+          pip install mmcv-full==latest+torch${{matrix.mmcv}} -f https://openmmlab.oss-accelerate.aliyuncs.com/mmcv/dist/index.html
           pip install -q git+https://github.com/open-mmlab/mmdetection.git
           pip install -r requirements.txt
       - name: Build and install

.github/workflows/deploy.yml

@@ -9,7 +9,7 @@ jobs:
     steps:
       - uses: actions/checkout@v2
       - name: Set up Python 3.7
-        uses: actions/setup-python@v1
+        uses: actions/setup-python@v2
         with:
           python-version: 3.7
       - name: Install torch

README.md

@@ -14,7 +14,7 @@ Documentation: https://mmdetection3d.readthedocs.io/
 
 ## Introduction
 
-The master branch works with **PyTorch 1.3 to 1.5**.
+The master branch works with **PyTorch 1.3 to 1.6**.
 
 MMDetection3D is an open source object detection toolbox based on PyTorch, towards the next-generation platform for general 3D detection. It is
 a part of the OpenMMLab project developed by [MMLab](http://mmlab.ie.cuhk.edu.hk/).

configs/mvxnet/dv_mvx-fpn_second_secfpn_adamw_2x8_80e_kitti-3d-3class.py

@@ -249,6 +249,7 @@ total_epochs = 40
 dist_params = dict(backend='nccl')
 log_level = 'INFO'
 work_dir = None
-load_from = './pretrain_mmdet/mvx_faster_rcnn_detectron2-caffe_20e_coco-pretrain_gt-sample_kitti-3-class_moderate-79.3_20200207-a4a6a3c7.pth'  # noqa
+# You may need to download the model first is the network is unstable
+load_from = 'https://openmmlab.oss-accelerate.aliyuncs.com/mmdetection3d/pretrain_models/mvx_faster_rcnn_detectron2-caffe_20e_coco-pretrain_gt-sample_kitti-3-class_moderate-79.3_20200207-a4a6a3c7.pth'  # noqa
 resume_from = None
 workflow = [('train', 1)]

mmdet3d/core/bbox/coders/partial_bin_based_bbox_coder.py

@@ -114,22 +114,22 @@ class PartialBinBasedBBoxCoder(BaseBBoxCoder):
 
         # decode objectness score
         end += 2
-        results['obj_scores'] = preds_trans[..., start:end]
+        results['obj_scores'] = preds_trans[..., start:end].contiguous()
         start = end
 
         # decode center
         end += 3
         # (batch_size, num_proposal, 3)
-        results['center'] = base_xyz + preds_trans[..., start:end]
+        results['center'] = base_xyz + preds_trans[..., start:end].contiguous()
         start = end
 
         # decode direction
         end += self.num_dir_bins
-        results['dir_class'] = preds_trans[..., start:end]
+        results['dir_class'] = preds_trans[..., start:end].contiguous()
         start = end
 
         end += self.num_dir_bins
-        dir_res_norm = preds_trans[..., start:end]
+        dir_res_norm = preds_trans[..., start:end].contiguous()
         start = end
 
         results['dir_res_norm'] = dir_res_norm
@@ -137,7 +137,7 @@ class PartialBinBasedBBoxCoder(BaseBBoxCoder):
 
         # decode size
         end += self.num_sizes
-        results['size_class'] = preds_trans[..., start:end]
+        results['size_class'] = preds_trans[..., start:end].contiguous()
         start = end
 
         end += self.num_sizes * 3
@@ -147,13 +147,13 @@ class PartialBinBasedBBoxCoder(BaseBBoxCoder):
             [batch_size, num_proposal, self.num_sizes, 3])
         start = end
 
-        results['size_res_norm'] = size_res_norm
+        results['size_res_norm'] = size_res_norm.contiguous()
         mean_sizes = preds.new_tensor(self.mean_sizes)
         results['size_res'] = (
             size_res_norm * mean_sizes.unsqueeze(0).unsqueeze(0))
 
         # decode semantic score
-        results['sem_scores'] = preds_trans[..., start:]
+        results['sem_scores'] = preds_trans[..., start:].contiguous()
 
         return results
 

mmdet3d/models/dense_heads/vote_head.py

@@ -157,7 +157,8 @@ class VoteHead(nn.Module):
                 torch.randint(0, num_seed, (batch_size, self.num_proposal)),
                 dtype=torch.int32)
         else:
-            raise NotImplementedError
+            raise NotImplementedError(
+                f'Sample mode {sample_mod} is not supported!')
 
         vote_aggregation_ret = self.vote_aggregation(vote_points,
                                                      vote_features,
@@ -261,7 +262,7 @@ class VoteHead(nn.Module):
             (batch_size, proposal_num, self.num_sizes))
         one_hot_size_targets.scatter_(2, size_class_targets.unsqueeze(-1), 1)
         one_hot_size_targets_expand = one_hot_size_targets.unsqueeze(
-            -1).repeat(1, 1, 1, 3)
+            -1).repeat(1, 1, 1, 3).contiguous()
         size_residual_norm = torch.sum(
             bbox_preds['size_res_norm'] * one_hot_size_targets_expand, 2)
         box_loss_weights_expand = box_loss_weights.unsqueeze(-1).repeat(

mmdet3d/models/model_utils/vote_module.py

@@ -88,13 +88,13 @@ class VoteModule(nn.Module):
 
         votes = votes.transpose(2, 1).view(batch_size, num_seed,
                                            self.vote_per_seed, -1)
-        offset = votes[:, :, :, 0:3]
-        res_feats = votes[:, :, :, 3:]
+        offset = votes[:, :, :, 0:3].contiguous()
+        res_feats = votes[:, :, :, 3:].contiguous()
 
-        vote_points = (seed_points.unsqueeze(2) + offset).contiguous()
+        vote_points = seed_points.unsqueeze(2) + offset
         vote_points = vote_points.view(batch_size, num_vote, 3)
-        vote_feats = (seed_feats.transpose(2, 1).unsqueeze(2) +
-                      res_feats).contiguous()
+        vote_feats = seed_feats.permute(
+            0, 2, 1).unsqueeze(2).contiguous() + res_feats
         vote_feats = vote_feats.view(batch_size, num_vote,
                                      feat_channels).transpose(2,
                                                               1).contiguous()

mmdet3d/models/roi_heads/bbox_heads/parta2_bbox_head.py

@@ -246,7 +246,7 @@ class PartA2BboxHead(nn.Module):
         # transform to sparse tensors
         sparse_shape = part_feats.shape[1:4]
         # (non_empty_num, 4) ==> [bs_idx, x_idx, y_idx, z_idx]
-        sparse_idx = part_feats.sum(dim=-1).nonzero()
+        sparse_idx = part_feats.sum(dim=-1).nonzero(as_tuple=False)
 
         part_features = part_feats[sparse_idx[:, 0], sparse_idx[:, 1],
                                    sparse_idx[:, 2], sparse_idx[:, 3]]
@@ -600,7 +600,8 @@ class PartA2BboxHead(nn.Module):
             class_scores_keep = box_probs[:, k] >= score_thresh[k]
 
             if class_scores_keep.int().sum() > 0:
-                original_idxs = class_scores_keep.nonzero().view(-1)
+                original_idxs = class_scores_keep.nonzero(
+                    as_tuple=False).view(-1)
                 cur_boxes_for_nms = boxes_for_nms[class_scores_keep]
                 cur_rank_scores = box_probs[class_scores_keep, k]
 

mmdet3d/models/roi_heads/part_aggregation_roi_head.py

@@ -260,7 +260,8 @@ class PartAggregationROIHead(Base3DRoIHead):
                     # gather assign_results in different class into one result
                     batch_num_gts += cur_assign_res.num_gts
                     # gt inds (1-based)
-                    gt_inds_arange_pad = gt_per_cls.nonzero().view(-1) + 1
+                    gt_inds_arange_pad = gt_per_cls.nonzero(
+                        as_tuple=False).view(-1) + 1
                     # pad 0 for indice unassigned
                     gt_inds_arange_pad = F.pad(
                         gt_inds_arange_pad, (1, 0), mode='constant', value=0)

tests/test_losses.py

@@ -50,19 +50,22 @@ def test_chamfer_disrance():
 
     assert torch.allclose(loss_source, torch.tensor(219.5936))
     assert torch.allclose(loss_target, torch.tensor(22.3705))
-    assert (indices1 == indices1.new_tensor([[0, 4, 4, 4, 4, 2, 4, 4, 4, 3],
-                                             [0, 1, 0, 1, 0, 4, 2, 0, 0,
-                                              1]])).all()
-    assert (indices2 == indices2.new_tensor([[0, 0, 0, 0, 0], [0, 3, 6, 0,
-                                                               0]])).all()
+
+    expected_inds1 = [[0, 4, 4, 4, 4, 2, 4, 4, 4, 3],
+                      [0, 1, 0, 1, 0, 4, 2, 0, 0, 1]]
+    expected_inds2 = [[0, 4, 4, 4, 4, 2, 4, 4, 4, 3],
+                      [0, 1, 0, 1, 0, 3, 2, 0, 0, 1]]
+    assert (torch.equal(indices1, indices1.new_tensor(expected_inds1))
+            or torch.equal(indices1, indices1.new_tensor(expected_inds2)))
+    assert torch.equal(indices2,
+                       indices2.new_tensor([[0, 0, 0, 0, 0], [0, 3, 6, 0, 0]]))
 
     loss_source, loss_target, indices1, indices2 = chamfer_distance(
         source, target, reduction='sum')
 
     assert torch.allclose(loss_source, torch.tensor(219.5936))
     assert torch.allclose(loss_target, torch.tensor(22.3705))
-    assert (indices1 == indices1.new_tensor([[0, 4, 4, 4, 4, 2, 4, 4, 4, 3],
-                                             [0, 1, 0, 1, 0, 4, 2, 0, 0,
-                                              1]])).all()
+    assert (torch.equal(indices1, indices1.new_tensor(expected_inds1))
+            or torch.equal(indices1, indices1.new_tensor(expected_inds2)))
     assert (indices2 == indices2.new_tensor([[0, 0, 0, 0, 0], [0, 3, 6, 0,
                                                                0]])).all()