Adjustments to setup.py and README for torch Sobol

[Balandat]

Summary:
Using the torch native SobolEngine makes things a lot simpler for botorch.
ax still uses the cython SobolEngine for a lot of things outside torch land.

Differential Revision: D14946543

[eytan]

Can we hold off on committing this until closer to PyTorch 1.1 lands? I don’t want to require external testers of the software to install PyTorch off master, but rather use conda.

…

Sent from my iPhone

On Apr 15, 2019, at 6:43 PM, Max Balandat ***@***.***> wrote: Summary: Using the torch native SobolEngine makes things a lot simpler for botorch. ax still uses the cython SobolEngine for a lot of things outside torch land. Differential Revision: D14946543 You can view, comment on, or merge this pull request online at: #55 Commit Summary Use torch native SobolEngine Add `__version__` attribute to botorch package (#45) Adjustments to setup.py and README for torch Sobol File Changes M README.md (25) M botorch/__init__.py (8) M botorch/acquisition/sampler.py (2) M botorch/qmc/__init__.py (3) M botorch/qmc/normal.py (87) D botorch/qmc/sobol.pyx (1690) M botorch/utils/sampling.py (17) M setup.py (24) M test/qmc/test_normal.py (653) D test/qmc/test_sobol.py (232) M test/utils/test_sampling.py (2) Patch Links: https://github.com/facebookexternal/botorch/pull/55.patch https://github.com/facebookexternal/botorch/pull/55.diff — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or mute the thread.

[Balandat]

I don’t want to require external testers of the software to install PyTorch off master, but rather use conda.

The PyTorch nightlies are available on conda. So the setup will require running
conda install pytorch-nightly -c pytorch instead of
conda install pytorch -c pytorch

Are you worried about them having to install a different version of pytorch if they already have one set up?

[Balandat]

Hmm good thing I exported this. Looks like we're getting failures that we are not seeing internally (D14601706).

My suspicion is that the pytorch nightlies we get from pip don't contain/properly build the ATen SobolEngine. This all runs fine when using the nightlies from conda locally on my Mac.

[eytan]

If it is that easy then it’s OK. Go for it.

…

Sent from my iPhone

On Apr 15, 2019, at 6:48 PM, Max Balandat ***@***.***> wrote: I don’t want to require external testers of the software to install PyTorch off master, but rather use conda. The PyTorch nightlies are available on conda. So the setup changes from conda install pytorch-nightly -c pytorch instead of conda install pytorch -c pytorch Are you worried about them having to install a different version of pytorch if they already have one set up? — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or mute the thread.

[Balandat]

We'll need to figure out what the issues are with the nightlies in travis first anyway.

[Balandat]

@soumith, any idea why we're running into the following on travis when using the pytorch nightlies from pip (this is on linux):

AttributeError: module 'torch' has no attribute '_sobol_engine_initialize_state_'

Things run fine using the pytorch nightlies from conda on my Mac (pip installs on my Mac fail b/c of a MKL linking issues). Things also run fine internally (D14601706).

[Balandat]

Travis log for a recent failed run: https://travis-ci.com/facebookexternal/botorch/jobs/194077151

[soumith]

i cant see the log. it says cant find project. Can you paste the failure?

[Balandat]

https://phabricator.internmc.facebook.com/P62627525$1044-1045

Excerpt:

self = SobolEngine(dimension=1, scramble=True, seed=6627), dimension = 1
scramble = True, seed = 6627

�[1m    def __init__(self, dimension, scramble=False, seed=None):�[0m
�[1m        if dimension > self.MAXDIM or dimension < 1:�[0m
�[1m            raise ValueError("Supported range of dimensionality "�[0m
�[1m                             "for SobolEngine is [1, {}]".format(self.MAXDIM))�[0m
�[1m    �[0m
�[1m        self.seed = seed�[0m
�[1m        self.scramble = scramble�[0m
�[1m        self.dimension = dimension�[0m
�[1m    �[0m
�[1m        self.sobolstate = torch.zeros(dimension, self.MAXBIT, dtype=torch.long)�[0m
�[1m>       torch._sobol_engine_initialize_state_(self.sobolstate, self.dimension)�[0m
�[1m�[31mE       AttributeError: module 'torch' has no attribute '_sobol_engine_initialize_state_'�[0m

�[1m�[31m../../../virtualenv/python3.6.3/lib/python3.6/site-packages/torch/quasirandom.py�[0m:54: AttributeError

[Balandat]

Upon further digging, it seems that this failure may be caused by gpytorch's pip install re-installing the latest pytorch version over the nightlies. This should not happen.

[Balandat]

Fixed gpytorch setup.py upstream: cornellius-gp/gpytorch#663, will peg to that commit once merged into master.

[Balandat]

Argh I don't know why this is still failing...

[Balandat]

OK so even with the upstream change to the gpytorch setup we still get the following:

I think the reason is that the pip install actually calls the package torch-nightly and not torch as conda does.

I tried working around this upstream by doing some parsing in setup.py of the sorts

torch_min = "1.0.1"
# if recent dev version of PyTorch is installed, no need to install stable
try:
    import torch
    if torch.__version__ >= torch_min:
        install_requires = []
    else:
        install_requires = ["torch" + ">=" + torch_min]
except ImportError:
    install_requires = ["torch" + ">=" + torch_min]

Unfortunately. there is a bug in pip 19.0.1 that shows up in conjunction with having a pyproject.toml file (which gpytorch does) that always results in setup.py not being able to import torch: pypa/pip#6163.

I think for now I'll just re-install the nightlies after installing gpytorch in .travis.yml, and then just put the req. "torch>=1.1` upon release.