sorenlassen / einsum-experiment

Experimental implementation of torch/onnx einsum

einsum

Experimental implementation of torch/onnx einsum

Install

pip3 install -U einsum

Or develop locally:

git clone https://github.com/sorenlassen/einsum-experiment ~/einsum
cd ~/einsum
python3 setup.py develop

Usage

import einsum

print('TODO')

Tests

Run einsum's test suite:

pip3 install pytest
pytest

or

python3 setup.py develop # enables the test file to import einsum
python3 tests/test_einsum.py

Type check with mypy:

pip3 install mypy
python3 -m mypy src/einsum/lib.py

Release

To publish a new release to pypi:

pip3 install python-semantic-release

# verify tests pass.
pytest

# bump the version number, add a git tag.
semantic-release version --patch # or --minor, or --major

# push the new version number and tag to github.
git push && git push --tags

# publish to pypi.
poetry build
poetry publish

About

pyproject.toml was generated with mkpylib. setup.py was generated with poetry-gen-setup-py.

einsum-experiment's Issues

Broadcast problem

I was testing the examples at https://www.tensorflow.org/api_docs/python/tf/einsum and noticed an interesting failure case (search for "Einsum will broadcast over axes covered by the ellipsis" on that page):

import tensorflow as tf
import einsum
s = tf.random.normal(shape=[11, 1, 5, 3])
t = tf.random.normal(shape=[1, 7, 3, 2])
e =  tf.einsum('...ij,...jk->...ik', s, t)
print(e.shape)
e2 =  einsum.einsum('...ij,...jk->...ik', s.numpy(), t.numpy())

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/Users/shawn/ml/einsum/src/einsum/lib.py", line 228, in einsum
    return einsum_execute(spec, list(tensors))
  File "/Users/shawn/ml/einsum/src/einsum/lib.py", line 223, in einsum_execute
    return einsum_tensor_frompos(fn, spec.output.shape)
  File "/Users/shawn/ml/einsum/src/einsum/lib.py", line 31, in einsum_tensor_frompos
    return einsum_tensor(recurse([], shape))
  File "/Users/shawn/ml/einsum/src/einsum/lib.py", line 29, in recurse
    return [ recurse(pos + [i], remainder) for i in range(shape[0]) ]
  File "/Users/shawn/ml/einsum/src/einsum/lib.py", line 29, in <listcomp>
    return [ recurse(pos + [i], remainder) for i in range(shape[0]) ]
  File "/Users/shawn/ml/einsum/src/einsum/lib.py", line 29, in recurse
    return [ recurse(pos + [i], remainder) for i in range(shape[0]) ]
  File "/Users/shawn/ml/einsum/src/einsum/lib.py", line 29, in <listcomp>
    return [ recurse(pos + [i], remainder) for i in range(shape[0]) ]
  File "/Users/shawn/ml/einsum/src/einsum/lib.py", line 29, in recurse
    return [ recurse(pos + [i], remainder) for i in range(shape[0]) ]
  File "/Users/shawn/ml/einsum/src/einsum/lib.py", line 29, in <listcomp>
    return [ recurse(pos + [i], remainder) for i in range(shape[0]) ]
  File "/Users/shawn/ml/einsum/src/einsum/lib.py", line 29, in recurse
    return [ recurse(pos + [i], remainder) for i in range(shape[0]) ]
  File "/Users/shawn/ml/einsum/src/einsum/lib.py", line 29, in <listcomp>
    return [ recurse(pos + [i], remainder) for i in range(shape[0]) ]
  File "/Users/shawn/ml/einsum/src/einsum/lib.py", line 26, in recurse
    return fn(*pos)
  File "/Users/shawn/ml/einsum/src/einsum/lib.py", line 221, in fn
    return recurse(in_only_idxs)
  File "/Users/shawn/ml/einsum/src/einsum/lib.py", line 218, in recurse
    acc += recurse(tail)
  File "/Users/shawn/ml/einsum/src/einsum/lib.py", line 210, in recurse
    prod *= tensor.item(*pos)
IndexError: index 1 is out of bounds for axis 1 with size 1

The example right above it works fine:

import tensorflow as tf
import einsum
import numpy as np
s = tf.random.normal(shape=[11, 7, 5, 3])
t = tf.random.normal(shape=[11, 7, 3, 2])
e =  tf.einsum('...ij,...jk->...ik', s, t)
print(e.shape)
e2 = einsum.einsum('...ij,...jk->...ik', s.numpy(), t.numpy())
np.allclose(e, e2) # True

So the problem seems related to broadcasting axes covered by ellipses.