RuntimeError: expected scalar type Float but found Half

[ghost]

When I ran train_net.py, I got an error

RuntimeError: expected scalar type Float but found Half (data_ptr at /home/cy/anaconda3/envs/render/lib/python3.6/s ite-packages/torch/include/ATen/core/TensorMethods.h:6321)
frame #0: c10::Error::Error(c10::SourceLocation, std::string const&) + 0x47 (0x7f32dc94a627 in /home/cy/anaconda3/envs/NHR /lib/python3.6/site-packages/torch/lib/libc10.so)
frame #1: float* at::Tensor::data_ptr() const + 0xa60 (0x7f32bbb33480 in /home/cy/anaconda3/envs/NHR/lib/python3.6/ site-packages/pointnet2-0.0.0-py3.6-linux-x86_64.egg/pointnet2_cuda.cpython-36m-x86_64-linux-gnu.so)
frame #2: group_points_wrapper_fast(int, int, int, int, int, at::Tensor, at::Tensor, at::Tensor) + 0x27 (0x7f32bbb34427 in /home/cy/anaconda3/envs/NHR/lib/python3.6/site-packages/pointnet2-0.0.0-py3.6-linux-x86_64.egg/pointnet2_cuda.cpython-36m -x86_64-linux-gnu.so)
frame #3: + 0x14947 (0x7f32bbb2c947 in /home/cy/anaconda3/envs/NHR/lib/python3.6/site-packages/pointnet 2-0.0.0-py3.6-linux-x86_64.egg/pointnet2_cuda.cpython-36m-x86_64-linux-gnu.so)
frame #4: + 0x11a06 (0x7f32bbb29a06 in /home/cy/anaconda3/envs/NHR/lib/python3.6/site-packages/pointnet 2-0.0.0-py3.6-linux-x86_64.egg/pointnet2_cuda.cpython-36m-x86_64-linux-gnu.so)

frame #11: THPFunction_apply(_object*, _object*) + 0xa1f (0x7f330e1f6e3f in /home/cy/anaconda3/envs/NHR/lib/python3.6/site -packages/torch/lib/libtorch_python.so)

then I found that the error occurred on the line 40 of pointnet2_modules.py, which is
new_features = self.mlps[i](new_features).
The type of ‘new_features’ as input variable is 'torch.float32', but the type of output of 'self.mlps' is 'torch.float16'.

[wuminye]

This error is usually seen when latest apex module is used. In our implementation, we use an older version (https://github.com/NVIDIA/apex/tree/088985936518be7e25795a30d8ab33affa9db6ed). I suggest downloading and recompiling the old one.

An alternative is to turn off the apex. In such a way, you should remove apex stuffs in train_net.py and engine/trainer.py.

[ghost]

This error is usually seen when latest apex module is used. In our implementation, we use an older version (https://github.com/NVIDIA/apex/tree/088985936518be7e25795a30d8ab33affa9db6ed). I suggest downloading and recompiling the old one.

An alternative is to turn off the apex. In such a way, you should remove apex stuffs in train_net.py and engine/trainer.py.

我修改了pointnet2_modules.py的第154行new_features = self.mlp(new_features).type(torch.float32)和第40行new_features = self.mlps[i](new_features).type(torch.float32)，修改之后可以正常运行，请问这样解决是否合理呢？

[wuminye]

This error is usually seen when latest apex module is used. In our implementation, we use an older version (https://github.com/NVIDIA/apex/tree/088985936518be7e25795a30d8ab33affa9db6ed). I suggest downloading and recompiling the old one.
An alternative is to turn off the apex. In such a way, you should remove apex stuffs in train_net.py and engine/trainer.py.

我修改了pointnet2_modules.py的第154行new_features = self.mlp(new_features).type(torch.float32)和第40行new_features = self.mlps[i](new_features).type(torch.float32)，修改之后可以正常运行，请问这样解决是否合理呢？

Nice try 👍
I don't think it's an efficient way, but it may work.