Olivier Breuleux
Olivier Breuleux
This is a rough rewrite that wasn't fully tested, there's still a breakpoint lying in the code somewhere, but once cleaned up it should make things more efficient.
Myia is currently not able to handle aliased tensors in data structures. This issue can crop up in the Pytorch frontend, in code like this: ```python class LinearSeq(torch.nn.Module): def __init__(self,...
The equality mechanism used by `intern` does not work on dicts, because it discards key information, e.g. `{'x': 6}` and `{'y': 6}` would be considered equal.
This issue pops up in commit 740731cba4577f463785a9a96fed45dd73704224 for test `test_infer::test_raise_rec`. It is not deterministic, so you may have to run the test multiple times. The error is `Too many parents...
AbstractFunction can currently hold a set of possible functions. This functionality is redundant with an AbstractUnion of AbstractFunctions. It would be therefore better if AbstractFunction held a single function, and...
When using the inferrer with max_depth > 1, meaning that we propagate values for more than one level, the specializer produces erroneous graphs. The problem occurs with `test_pow10` and also...
This is a follow up to a discussion in #30, I'm making an issue out of it so that it isn't all buried over there (feel free to copy your...
## Purpose Provide an interface, akin to `from milatools.datasets.torch import ImageNet`, to acquire and use certain datasets in a *location aware* way. For example, when running on the Mila cluster,...