[Masaki Kozuki, chainer] The installation guides of Chainer & CuPy are enough I think.
- chainer: https://docs.chainer.org/en/stable/install.html#enable-cuda-cudnn-support
- cupy: https://docs-cupy.chainer.org/en/stable/install.html#install-cupy
For cuda10.0 environment, I think the below is enough according to the above guides.
$ pip install -U setuptools pip
$ pip install cupy-cuda100
$ pip install chainerchainerx.is_available() is also returning False.
F.sigmoid_cross_entropy](https://docs.chainer.org/en/stable/reference/generated/chainer.functions.sigmoid_cross_entropy.html#chainer.functions.sigmoid_cross_entropy) not enough?
[Tommi Kerola, chainer] https://github.com/chainer/chainer/blob/master/chainer/variable.py#L1393
This method uses :data:
gradas the initial error array. User can manually set a gradient array before calling this method.
So if you setvector.grad = xp.ones(shape), I think you can backprop vector losses.
[Seiya Tokui, chainer] It is a difficult question repeatedly raised by us, too! I really want to know the essential factor of their success.
The obvious factors that I came up with are: 1) their location (closer to the top-notch researchers/labs, other important partners like NVIDIA, etc., while Chainer is developed in Japan), 2) strong marketing (I heard there was a closing beta period involving researchers outside Facebook. Their website is also fancy from the beginning, maybe attracting young researchers who are the core of DL community), 3) being backed by a big company. I do not think these points explain everything, though.
nm --demangle libchainerx.so I see a bunch of them, which do not seem to be included in CMakeLists.txt. What have I missed? Now I am quiet curious with the compiling process..
This is the tracking issue for ChainerX op coverage. There is a link to a spreadsheet that list up supported ops: chainer/chainer#6423
There is an example code here:
https://github.com/chainer/chainer/tree/master/chainerx_cc/examples/mnist
It's a deep learning example, though.
Currently there's no way to disable backprop code in compilation, but technically it should be doable. It would be great if you could submit a feature request as an issue.
You can find all of the CUDA code under
chainerx_cc/chainerx/cuda.
https://github.com/numfocus/gsoc#organizations-confirmed-under-numfocus-umbrella
https://github.com/cupy/cupy/wiki/GSoC-2020-Project-Ideas
