Vadim Kantorov
Vadim Kantorov
cudnnR4 doesn't choose the optimal algorithm in the fully-connected mode, even with cudnn.benchmark = true, which results in ~20x slower backward pass compared to MatConvNet. Torch: ``` --TORCH OUTPUT (in...
Hi, I'm looking to teach BrowserFS understand [JupyterLab](https://github.com/jupyterlab/jupyterlab)'s [API](https://github.com/jupyterlab/jupyterlab/blob/master/packages/services/src/contents/index.ts#L51-L145) for fetching files from the server (at the end of the day I'd like to mount this FS to Emscripten's FS)....
### Context - [X] This issue is not a bug report. *(please use a different template for reporting a bug)* - [X] This issue is not a duplicate of an...
Drop torchvision 0.2.2 from conda-forge as it's often a silent indicator of major package conflicts
### 🐛 Describe the bug Running the following triggers install of torchvision 0.2.2. We need a stable conda install recipes and suggestions that deal with common bugged dependencies such as...
### 🚀 The feature I think the `io` namespace makes more sense, and it's where `torchvision.io.read_image` is located... ### Motivation, pitch N/A ### Alternatives _No response_ ### Additional context _No...
It would be nice to have them available in the standard library (both on CPU/GPU) rather than rolling your own. Kornia seems to have implemented it: https://kornia.readthedocs.io/en/latest/color.html, skimage as well:...
### 📚 The doc issue Does Swin impl in this repo support arbitrary dynamic execution-time-defined input resolution (same as other backbones)? Initially Swin was trained to support only one resolution,...
I have PyTorch installed from a nightly build. I think maybe it now needs a C++14 options in compiler. I install ctcdecode as `pip install git+https://github.com/parlance/ctcdecode` [log.txt](https://github.com/parlance/ctcdecode/files/3973806/log.txt)
I found a useful tutorial: https://b3d.interplanety.org/en/render-from-console-only-on-the-specified-gpu-devices-2/ This is useful for big GPU machines which may be shared by multiple people, specifying the devices is useful. Btw does Blender respect CUDA_VISIBLE_DEVICES...
According to https://github.com/pytorch/pytorch/issues/78475#issuecomment-1142559982, https://github.com/espnet/espnet/blob/5fa6dcc4e649dc66397c629d0030d09ecef36b80/espnet/nets/pytorch_backend/ctc.py#L65 should probably not have any effect on the underlying cudnn CTC loss algorithm.