Fix corner cases in reshape and improve its robustness #72
Conversation
gausshj
commented
Oct 24, 2025
gausshj
commented
- Fix corner cases for reshape none edge into (1, 1, ...) shape
- Add input new shape validity check
- Add corner cases in reshape
Codecov Report✅ All modified and coverable lines are covered by tests. 📢 Thoughts on this report? Let us know! |
| merging_sign: list[tuple[int, torch.Tensor]] = [] | ||
|
|
||
| original_self_is_scalar = self.tensor.dim() == 0 | ||
| if original_self_is_scalar: |
There was a problem hiding this comment.
把这个判断放最外面挺好的，不过为啥只有original是scalar的判断，没有target是scalar的判断？
There was a problem hiding this comment.
放在外面确实好很多，现在已经在607dfd8把target判断放在了外面。代码复杂度和可读性大大提高了。
| return even, odd | ||
|
|
||
| @staticmethod | ||
| def calculate_even_odd(edges: tuple[tuple[int, int], ...]) -> tuple[int, int]: |
There was a problem hiding this comment.
这两个函数有一个就行，不用封装得这么狠，再说，你这个函数也没用到上面那个函数。
There was a problem hiding this comment.
这个函数原则上可以被上面那个reorder_indices覆盖功能, 但是你在结果是(1, 0)的情况下, 希望做简单/快速的判断. 所以我建议: 直接写一个 is_xxx_valid 之类的函数, 或者直接判断 edges_only.count((0, 1)) % 2 == 0 .
There was a problem hiding this comment.
这里这个函数后面开发会不会也有用？我在svd的测试中也用到了类似的函数，这里实现了后续可以直接调用。
grassmann_tensor/tensor.py
Outdated
| def calculate_even_odd(edges: tuple[tuple[int, int], ...]) -> tuple[int, int]: | ||
| even, odd = 1, 0 | ||
| for e, o in edges: | ||
| even, odd = even * e + odd * o, even * o + odd * e |
There was a problem hiding this comment.
试试用functool reduce写吧，其他地方是用reduce写的，保持下一致性
hzhangxyz
left a comment
hzhangxyz
left a comment
There was a problem hiding this comment.
内容都没啥问题, 只是看着代码习惯不大好, 比如eo你过个一个月就认不出来是even odd的意思了...
grassmann_tensor/tensor.py
Outdated
| f"Ambiguous integer dim {item} from scalar. " | ||
| "Use explicit (even, odd) pairs, or only use 1 for trivial edges." | ||
| ) | ||
| normalized.append((1, 0)) |
There was a problem hiding this comment.
normalized 这个变量的名字是怎么来到? 这个的含义是 归一化的, 你这里是new_shape_list ?
There was a problem hiding this comment.
部分参考了GPT，这里是因为输入new_shape是tuple[int | tuple[int, int], ...]，但是在“标量情况”中（self.tensor.dim() == 0），不允许这种混合，normarized是类型规范化的意思。
| return even, odd | ||
|
|
||
| @staticmethod | ||
| def calculate_even_odd(edges: tuple[tuple[int, int], ...]) -> tuple[int, int]: |
There was a problem hiding this comment.
这个函数原则上可以被上面那个reorder_indices覆盖功能, 但是你在结果是(1, 0)的情况下, 希望做简单/快速的判断. 所以我建议: 直接写一个 is_xxx_valid 之类的函数, 或者直接判断 edges_only.count((0, 1)) % 2 == 0 .
那我要不加上注释或者把变量名改成易读懂的类型？ |
- Fix corner cases when reshaping a None edge into (1,1,…) while preserving arrow/parity. - Validate and normalize new_shape; reject incompatible dims early. - Move `target` check outside the while-loop to reduce nesting and improve readability. - Unify `_calculate_even_odd` and `calculate_even_odd` via `functools.reduce`. - Add/extend corner-case tests for reshape to prevent regressions.
gausshj
force-pushed
the
fix/reshape-with-none-edge
branch
from
27e75dc to
f4979fa
Compare
gausshj
changed the title
gausshj
changed the title
