PyTorch nightly with conda

To install PyTorch nightly from pytorch.otg:

 conda install pytorch torchvision cudatoolkit=10.0 -c pytorch-nightly

To install PyTorch nightly from anaconda

conda install -c pytorch pytorch-nightly

So, which one is correct? the latter one installs pytorch-nightly from pytorch channel, the first one install pytorch from pytorch-nightly channels!! confusing

The first one is the new nightly build. In doubt, stick to the install instructions on the install instructions on the website.

I use the first version. There seems now it always downgrades pytorch if I also include torchvision. nightly torchvision seems to be depending on an older version of nightly torch, also the stable torchvision raises an error when I try it with the nightly pytorch.

If I use conda update to update the nightly I get:

Updating pytorch is constricted by

torchvision -> requires pytorch==1.3.0.dev20190917

So I will have to remove torchvision first, and then I can only install the stable torchvision (which does not work against nightly pytorch)

I would assume the nightly binaries of torchvision and PyTorch should be matching, i.e. both should have the same build date. If some torchvision builds were skipped (for whatever reason), you might have to downgrade PyTorch to the latest matching nightly build.
@fmassa might know more about it.

@fmassa pytorch.org suggests

conda install pytorch torchvision cudatoolkit=10.0 -c pytorch-nightly

To get the latest nightly packages for PyTorch and torchvision.
So can I assume

conda update pytorch torchvision -c pytorch-nightly

Is the right way to update them both?

I currently cannot update both (torchvision forced PyTorch to donwgrade):

The following packages will be UPDATED:
  torchvision        pytorch::torchvision-0.3.0-py36_cu10.~ --> pytorch-nightly::torchvision-0.5.0.dev20190917-py36_cu100

The following packages will be DOWNGRADED:

  pytorch            1.3.0.dev20191002-py3.6_cuda10.0.130_~ --> 1.3.0.dev20190917-py3.6_cuda10.0.130_cudnn7.6.2_0

I just wan to know if this is normal or my setup is wrong. Also, I cannot load latest released torchvison against nightly PyTorch:

ImportError: /home/ehazar/miniconda3/envs/py3_night/lib/python3.6/site-packages/torchvision/_C.cpython-36m-x86_64-linux-gnu.so: undefined symbol: _ZN3c1011CPUTensorIdEv