Type annotation of the return value should be Union[dict[str, np.ndarray], dict[str, torch.Tensor]] instead of dict[str, Union[np.ndarray, torch.Tensor]]. The latter annotation accepts the return value like:
{
"a": np.array(...),
"b": torch.Tensor(...),
}
but it is actually invalid.
Furthermore, the design itself, where the return type of a function can change dynamically depending on the detach flag passed at initialization, makes it difficult for the editor to benefit from type inference. One example is the following (in #81).
The .encode() method of a BaseEncoder subclass must return the type dict[str, torch.Tensor] by design. SimpleEncoder initializes FeatureExtractor with detach=False at initialization time, so the type constraints of .encode() are satisfied at runtime, but that information is not determined at the type annotation level.
ShuntaroAoki
commented
1 month ago
https://github.com/KamitaniLab/bdpy/blob/016d4aedb11585f302f9e96d927496438320b915/bdpy/dl/torch/torch.py#L59-L60
Type annotation of the return value should be
Union[dict[str, np.ndarray], dict[str, torch.Tensor]]instead ofdict[str, Union[np.ndarray, torch.Tensor]]. The latter annotation accepts the return value like:but it is actually invalid.
Furthermore, the design itself, where the return type of a function can change dynamically depending on the detach flag passed at initialization, makes it difficult for the editor to benefit from type inference. One example is the following (in #81).
https://github.com/KamitaniLab/bdpy/blob/f67228e64c7b4fa8877234d2b9c4822fcb6ed069/bdpy/recon/torch/modules/encoder.py#L89-L103
The
.encode()method of aBaseEncodersubclass must return the typedict[str, torch.Tensor]by design.SimpleEncoderinitializesFeatureExtractorwithdetach=Falseat initialization time, so the type constraints of.encode()are satisfied at runtime, but that information is not determined at the type annotation level.