Closed kuaiqushangzixiba closed 2 years ago
If I comment " batch_first=True" out, it appears File "site-packages/torch/nn/functional.py", line 4790, in multi_head_attention_forward attn_output_weights = torch.bmm(q, k.transpose(1, 2)) RuntimeError: CUDA error: device-side assert triggered
@kuaiqushangzixiba It can be solved by update the pytorch > 1.7 which support batch_first parameters.
1.python train.py --cfg_file cfgs/kitti_models/voxset.yaml File "/OpenPCDet/pcdet/models/backbones_3d/vfe/voxset.py", line 168, in init self.mhsa = nn.MultiheadAttention(dim, num_heads=1, batch_first=True) TypeError: init() got an unexpected keyword argument 'batch_first' 2.python setup.py build_dist --inplace error: invalid command 'build_dist'
how can i fit it ?