Luo Wenyang

Question about RWKV formula

2

In the [first formula](https://github.com/BlinkDL/RWKV-LM/blob/main/RWKV-formula.png) in README, RWKV is rewritten into recurrent form by letting $W_n=(n-1)w$. Is there a particular reason for using $n-1$ instead of $n$? The latter is more...

The classifier-free guidance equation of diffusion models [here](https://github.com/labmlai/annotated_deep_learning_paper_implementations/blob/b05c9e0c57c6223b8f59dc11be114b97896b0481/labml_nn/diffusion/stable_diffusion/sampler/__init__.py#L50) is wrong, which is $$\epsilon_\theta(x_t, c) = s\epsilon_\text{cond}(x_t, c) + (s - 1)\epsilon_\text{cond}(x_t, c_u).$$ However, the correct equation is given in [the...

This solves issue #32 by limiting the number of threads to no more than 16.

`VideoCap::open()` sets the number of threads to whatever is returned by `std::thread::hardware_concurrency()` in [this line](https://github.com/LukasBommes/mv-extractor/blob/995ed82d38f2081f2a41fb69e570295ffc1fc2d4/src/mvextractor/video_cap.cpp#L115). However, on machines with many cores the number of threads will exceed the maximum recommended...

Is this repo still actively maintained?

1

This repo has not accepted pull requests for a long time, is this repo still maintained?

Fix issue 1721 by always initializing process group.

1

This pull request fixes issue #1721, where single GPU training/inference may error if the worker uses `torch.dist`. In summary, it makes sure the default process group is always initialized as...

Currently the [`launch`](https://github.com/Megvii-BaseDetection/YOLOX/blob/ac58e0a5e68e57454b7b9ac822aced493b553c53/yolox/core/launch.py#L39) function defined in [/yolox/core/launch.py](https://github.com/Megvii-BaseDetection/YOLOX/blob/main/yolox/core/launch.py) directly call the worker `func` when `world_size == 1`. However, this causes an issue that, if the training/inference worker uses and distributed facility...

How to train on GPUs with restricted VRAM

1

It seems that the code cannot run on mid-end GPUs, such as NVIDIA RTX 2060s, without OOM. How can we modify the code to reduce the VRAM consumption per GPU?

Please share the training recipe.

1

The training recipe (GPU model, how many query videos per task, how many tasks per batch, how many iterations/tasks/batches in total, optimizer options, learning rate policy, etc.) is not available...

[Feature Request] Allows registering custom trackers to internal tracker type registry

2

Currently `accelerate` maintains an internal known tracker registry [`LOGGER_TYPE_TO_CLASS`](https://github.com/huggingface/accelerate/blob/2c767338f29989e54ce93b3036b41e02013af7a7/src/accelerate/tracking.py#L960), which is used for tracker initializing, filtering, among other things. However, this makes working with custom trackers annoying in user code....