Compute static shape types in outputs of Join
#164
Conversation
| if not builtins.all(x.ndim == len(out_shape) for x in tensors): | ||
| raise TypeError( | ||
| "Only tensors with the same number of dimensions can be joined" | ||
| ) |
There was a problem hiding this comment.
Not relevant for the len(tensors) == 1 branch
| f = self.join_op(0, a, b, c, d, e) | ||
| fb = tuple(s == 1 for s in f.type.shape) | ||
| assert not fb[0] and fb[1] and fb[2] and fb[3] and not fb[4] and fb[5] | ||
| assert f.type.shape == (5, 1, 1, 1, None, 1) | ||
| assert fb == (False, True, True, True, False, True) |
There was a problem hiding this comment.
I kept the broadcastability assers even though this information is redundant from the current interpretation of shape == 1.
If in the future we re-introduce a broadcastable attribute, this should help updating the test.
michaelosthege
force-pushed
the
join-shapes
branch
4 times, most recently
from
9278ef7 to
d8896ea
Compare
| - gcc_linux-64 | ||
| - gxx_linux-64 | ||
| - numpy | ||
| - scipy | ||
| - six | ||
| - sphinx>=5.1.0 | ||
| - sphinx>=5.1.0,<6 |
There was a problem hiding this comment.
I would also add this to https://github.com/pymc-devs/pymc-sphinx-theme, otherwise we'll soon have the same issue in pymc, pymc-examples... (if we aren't already). We use the theme straight from github so the effect will be immediate
There was a problem hiding this comment.
michaelosthege
force-pushed
the
join-shapes
branch
from
513c722 to
78ce076
Compare
| # Other dims must match exactly, | ||
| # or if a mix of None and ? the output will be ? | ||
| # otherwise the input shapes are incompatible. | ||
| if len(inset) == 1: |
There was a problem hiding this comment.
Does join allow broadcasting?
There was a problem hiding this comment.
It doesn't look like it does - if so it wasn't tested, and NumPy does not allow broadcasting inputs to np.concatenate either
ricardoV94
changed the title
Join
ricardoV94
changed the title
JoinJoin
Motivation for these changes
Making static shapes propagate through
Joins such aspt.concatenate.Closes #163
Implementation details
The previous double-
forimplementation was rather tricky to refactor, which is why I went with a simple matrix-based approach.The first commit makes the corresponding tests and their output easier to read.
Checklist
New features
Joinnow propagate static shape information more reliably.Maintenance