Toshiki Kataoka
Toshiki Kataoka
Do the args require grads?
It seems `gym.Wrapper` assumes the base environment is a subclass of `gym.Env`. For example, ``` gym.wrappers.TimeLimit(chainerrl.envs.abc.ABC(), max_episode_steps=2) ``` raises `AttributeError: 'ABC' object has no attribute 'metadata'`.
Should Chainer support "complex32" whose real/imag part is `float16`? It seems NumPy hasn't yet support "complex32".
Definitely. Users should see `cupy.ndarray` is printed instead of `cupy.core.core.ndarray`. For developers, the change will not impact much because there's no `from ... import *` in CuPy.
Calling `numpy.einsum_path` can eliminate the copied code (`cupy/linalg/_einsum_opt.py`). My concern is about the performance. `cupy.einsum` will have to parse args twice, while `numpy.einsum` parses them only once by using hidden...
> However, what's concerning is before arriving this block, CuPy did several simplifications that mutate the inputs, which makes reusing `einsum_path` very challenging. As a result I'd suggest we defer...
Technically yes. It seems similar to the question "should `cupy.einsum` support genuine `optimize=False` einsum?" `cupy.einsum`'s actual path is a subdivision of the given path (if we ignore the execution order...
`cupy.is_available()` can be used. https://github.com/cupy/cupy/blob/9608557bf2f7361940d52e8bc5db69de036689fe/cupy/__init__.py#L35-L36
The exponent bias does not matter because the code is normalized by `code /= code.max()`. However the subnormals are still different (`0.0625` is quite small), probably because of a bug...