Hard requirement of torch==1.2.0?

[RomainSabathe]

Hello, I'm revisiting the dependencies for my project and I haven't done that in a long time. I have been using torch 1.1.0 for quite a while; I am confident my code works with it. However it seems like torchvision now requires torch 1.2.0. Here's the output from poetry (pipenv gave a more verbose but essentially same output):

Updating dependencies
Resolving dependencies... (0.1s)
                                                                                                                     
[SolverProblemError]                                                                                  
Because torchvision (0.4.0) depends on torch (1.2.0)                                                               
 and no versions of torchvision match >0.4.0,<0.5.0, torchvision (>=0.4.0,<0.5.0) requires torch (1.2.0).          
So, because my_project depends on both torch (=1.1.0) and torchvision (^0.4.0), version solving failed.  

Is this intended or is this a 'bug'? Am I missing something? Thanks.

[soumith]

this is not intended to be a bug. you can use torchvision 0.3.0 to use with pytorch 1.1.0

[Multihuntr]

The Pypi description for 0.4.0 still says:

TorchVision requires PyTorch 1.1 or newer.

But if you have PyTorch 1.1 installed, then installing pip install torchvision==0.4.0 with miniconda will remove PyTorch 1.1 and install PyTorch 1.2

[soumith]

whoops, the pypi description needs to be fixed.

[sbelharbi]

hi,
i am installing pytorch 1.3.0 using pip install torch torchvision
i am getting ERROR: torchvision 0.4.0 has requirement torch==1.2.0, but you'll have torch 1.3.0 which is incompatible.

but,

it succeeds to install!
Successfully installed torch-1.3.0 torchvision-0.4.0

$pip freeze
...
torch==1.3.0
torchvision==0.4.0
...

in the release note it says Note: torchvision 0.4 requires PyTorch 1.2 or newer

pypi says TorchVision requires PyTorch 1.1 or newer.

the doc issue mentioned above may be still going on.
thanks

[fmassa]

@soumith do you know how we update the pypi readme?

@sbelharbi if you install PyTorch 1.3.0, you need torchvision 0.4.1, or else you'll face import errors.
torchvision 0.4.0 works with PyTorch 1.2.0

[sbelharbi]

for now, quick import didn't raise an issue.
i'll try it out this week and see.
in case of issues, i'll downgrade to torch.1.2.0 and wait for torchvision.0.4.1.
thanks

In [1]: import torch

In [2]: import torchvision

In [3]: torch.__version__
Out[3]: '1.3.0'

In [4]: import torchvision.transforms.functional as TF

$ pip freeze | grep -w "torchvision"
torchvision==0.4.0

[sbelharbi]

first try leads to cuda incompatibility (requires cuda 10.1, i have 10.0). my fault.
i get torch.1.3.0 and torchvision.0.4.1 from wheels.
thanks for the heads up.

torch==1.3.0+cu100
torchvision==0.4.1+cu100

[soumith]

@fmassa you upload a source distribution python setup.py sdist and then twine upload the created file, then it will use the description etc. from setup.py to populate the README. If I remember correctly.

[Multihuntr]

It uses the long_description field from the setup.py, which uses your README.rst: https://github.com/pytorch/vision/blob/master/setup.py#L203

But I can see that that should have been updated as of v0.4.1? But it still says "TorchVision requires PyTorch 1.1 or newer." on pypi for v0.4.1.post2? Something weird is going on.

Btw; I looked this up when I came across this before (I was curious), and there's a giant thread of people complaining that you can't upload a new description without uploading the whole thing again: pypi/warehouse#2170