Attribute error with scipy.misc.logsumexp and autograd about ssm

[slinderman]

Hi Shruthi, Yes, I’m waiting for the autograd fix to make it to pypi. Please try installing autograd from github master. Thanks, Scott

…

On Jul 26, 2019, at 8:24 AM, r-shruthi11 ***@***.***> wrote: Hi, I think related to this issue (HIPS/autograd#501) here, I'm having trouble importing ssm because I get an attribute error trying to import logsumexp from scipy.misc. Would re-installing autograd from the master branch be the solution here? Thanks! Shruthi — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or mute the thread.

[r-shruthi11]

Thanks, Scott. I've tried that but I still see the same issue.

----> 1 import ssm

~/Desktop/Projects/state-space-models/ssm/ssm/ssm/init.py in
1 # Default imports for SSM
2
----> 3 from .hmm import *
4 from .lds import *

~/Desktop/Projects/state-space-models/ssm/ssm/ssm/hmm.py in
7
8 from ssm.optimizers import adam_step, rmsprop_step, sgd_step, convex_combination
----> 9 from ssm.primitives import hmm_normalizer, hmm_expected_states, hmm_filter, hmm_sample, viterbi
10 from ssm.util import ensure_args_are_lists, ensure_args_not_none,
11 ensure_slds_args_not_none, ensure_variational_args_are_lists, \

~/Desktop/Projects/state-space-models/ssm/ssm/ssm/primitives.py in
2 import autograd.numpy as np
3 import autograd.numpy.random as npr
----> 4 from autograd.scipy.misc import logsumexp
5 from autograd.scipy.linalg import cholesky_banded, solve_banded, solveh_banded
6 from autograd.extend import primitive, defvjp

ImportError: cannot import name 'logsumexp'

[slinderman]

My guess is that Python is still loading the pip installed version of Autograd. You can check by running

import autograd
autograd.__file__

and seeing what path it's pointing to. Make sure it's the path from Github.

That said, I just double checked Autograd and it looks like @mattjj just uploaded the latest Github master to PyPI. You should be able to get it by running

pip install autograd --upgrade

[r-shruthi11]

Hi Scott,

I tried both - setting the autograd path to point to the Github directory and upgrading my install of autograd but I still see the same error when I try to import logsumexp from autograd.scipy.misc. However, importing logsumexp from autograd.scipy.special seems to work.

pip install autograd --upgrade
Requirement already up-to-date: autograd in /Users/shrumac2/miniconda3/envs/ssmodels/lib/python3.5/site-packages (1.3)
Requirement not upgraded as not directly required: numpy>=1.12 in /Users/shrumac2/miniconda3/envs/ssmodels/lib/python3.5/site-packages (from autograd) (1.16.4)
Requirement not upgraded as not directly required: future>=0.15.2 in /Users/shrumac2/miniconda3/envs/ssmodels/lib/python3.5/site-packages (from autograd) (0.17.1)

python
Python 3.5.6 |Anaconda, Inc.| (default, Aug 26 2018, 16:30:03)
[GCC 4.2.1 Compatible Clang 4.0.1 (tags/RELEASE_401/final)] on darwin
Type "help", "copyright", "credits" or "license" for more information.

from autograd.scipy.misc import logsumexp
Traceback (most recent call last):
File "", line 1, in
ImportError: cannot import name 'logsumexp'
from autograd.scipy.special import logsumexp

logsumexp([1.0, 2.0, 3.0])
3.4076059644443806

[slinderman]

Sorry for the delay. Quick update: I reran the Travis integrations and they are also failing in the same way you describe. Unfortunately I haven't been able to reproduce the problem locally yet, and I'm currently tied up at a conference so it might take a few more days.

In the meantime, it would be helpful to also know more about your configuration: what version of numpy/scipy are you on? I'm sure there's a simple explanation for all this...

[r-shruthi11]

Thanks for the update, Scott. Here's the info:

autograd==1.3
certifi==2019.6.16
cycler==0.10.0
Cython==0.29.13
future==0.17.1
joblib==0.13.2
kiwisolver==1.1.0
matplotlib==3.1.1
numpy==1.17.0
pyparsing==2.4.1.1
python-dateutil==2.8.0
scikit-learn==0.21.2
scipy==1.3.0
six==1.12.0
tqdm==4.32.2

[slinderman]

Thanks for your patience. The fix was simple in the end: I had to change the imports to from autograd.scipy.special import logsumexp to work with the latest versions of scipy and autograd. Fix is here 57950a8.

Note: I also made these changes in the master branch of pybasicbayes and pyhsmm to fix the nosetests.