Results 8 issues of Paul Willot

Hello, thanks for working on pytorch-summary. Models are often created on CPU, or not on the first cuda device. Defaulting to the current model device might be helpful.

Seems like the label colors are inverted in this figure: ![image](https://www.mmlab-ntu.com/project/knet/img/knet_sem_seg.png)

Avoid calling `str.maketrans` repeatedly, and de-duplication white spaces with join+split ([split act on the same whitespaces](https://docs.python.org/3/library/stdtypes.html#str.split)) About ~~x1.5~~ x3 speedup, on a tiny and already fast function but this is...

Hi, first thanks for this nice paper! For plotting in this repo, what would you think of using a different [colormap designed for depth](https://ai.googleblog.com/2019/08/turbo-improved-rainbow-colormap-for.html)? It make it much easier to...

Hi, thanks for the useful lib. Small change to inherit from MutableMapping to improve downstream type hints, for example to more easily accept either a dict or a RedisDict.

The python requirements currently constrain to numpy

`cat /etc/os-release` is not always available, the prisma client broke when trying to deploy in an Ubuntu 22.04 VM. This uses the standard library's `platform` instead. Also fix the SSL...

Hi, thanks for the useful tool, `imghdr`, used [in the resizer](https://github.com/rom1504/img2dataset/blob/main/img2dataset/resizer.py#L148), has been [removed in python3.13](https://docs.python.org/3/library/imghdr.html#module-imghdr). We could use one of the recommended external library, but as we only care...