JIT problem with torchvision 0.5

[Borda]

🐛 Bug

There are some JIT problems with newly released torchvision 0.5
in #687 we freeze version to <0.5 but in future, we want to support all torchvisions
Maybe it is just a temporal bug in torchvision and they will handle it...

To Reproduce

#687 (comment)

Environment

https://app.circleci.com/jobs/github/Borda/pytorch-lightning/839

[9dogs]

Based on pytorch/vision#1383 (comment) torchvision now requires "recent enough" version of PyTorch. All tests are green with pytorch==1.4.0 and torchvision==0.5.0.
The problem is when pytorch==1.3.0 is installed, pip install torchvision installs torchvision==0.5.0 which is incompatible. Ah, I love dependency issues...

[Borda]

Then the problem is torchvision, it shall enforce torch >= 1.4.0

[9dogs]

But it does (starting from 0.4.0):

# pkginfo -f requires_dist torchvision-0.5.0+cpu-cp37-cp37m-linux_x86_64.whl
requires_dist: ['numpy', 'six', 'torch (==1.4.0)', 'pillow (>=4.1.1)', "scipy ; extra == 'scipy'"]
# pkginfo -f requires_dist torchvision-0.4.2+cpu-cp37-cp37m-linux_x86_64.whl
requires_dist: ['numpy', 'six', 'torch (==1.3.1)', 'pillow (>=4.1.1)', "scipy ; extra == 'scipy'"]
# pkginfo -f requires_dist torchvision-0.4.1+cpu-cp37-cp37m-linux_x86_64.whl
requires_dist: ['numpy', 'six', 'torch (==1.3.0)', 'pillow (>=4.1.1)', "scipy ; extra == 'scipy'"]
# pkginfo -f requires_dist torchvision-0.4.0+cu92-cp37-cp37m-manylinux1_x86_64.whl
requires_dist: ['numpy', 'six', 'torch (==1.2.0)', 'pillow (>=4.1.1)', "scipy ; extra == 'scipy'"]

My problem is I cannot freeze dependencies via pip-tools with torch==1.4.0 since pytorch-lightning limits torchvision version.
Dirty solution is to manipulate torch version through torchvision, i.e. omit torch dependency in requirements.txt completely and set torchvision>=0.4.0.

[neggert]

Do we need the torchvision dependency at all? IIRC it is only used for tests and examples.

[williamFalcon]

1. idk why we would limit the version
2. not needed. We do need to add to the examples that people need to have torchvision and torch installed.

[Borda]

we are using MNIST dataset in tests from torchvision.datasets import MNIST and transforms from torchvision import transforms

[williamFalcon]

so add torchvision to the tests requirements

[Borda]

but it is in package pytorch_lightning.testing.base_model so we would need to move it also out of the package or make own MNIT...

[kartikJ-9]

I have torch version 0.4.1, and torchvision->inception.py says that there is no module named call Optional in torch.jit.annotations....I want to use torch version 0.4.1 because I want to use it along with cuda 9.2 for my program. And as per my understanding, torchvision requires torch 1.4. How do I solve this dependency problem?

[Borda]

are you talking about torch==0.4.1 or torchvision=0.4.1 this is not related to lightning, you shall ask at torch/vision