Ma Zerun
Ma Zerun
In MMCLS, we use `permute` + `F.layer_norm` to implement `LayerNorm2d`. https://github.com/open-mmlab/mmclassification/blob/d2e505415040bf5329ab218bb6fe3d899f176cd5/mmcls/models/backbones/convnext.py#L35-L40 However, in ConvNeXt official repo, they use a more intuitional implementation. ```python elif self.data_format == "channels_first": u = x.mean(1,...
To develop a better image classification toolbox for everyone. Welcome any contributions and suggestions. In this issue, **you can either**: - Suggest a new feature by leaving a comment. The...
## Motivation To support using a single backbone to perform multiple classification tasks. ## Modification This PR is one part of the multi-task support plan, and it depends on #675...
## Motivation Some optimizer constructors set different learning rate for layers of different depth. ## Modification Add `get_layer_depth` method for `VisionTransformer`, `ConvNeXt` and `SwinTransformer`. ## Use cases (Optional) **Vision Transformer**...
## Motivation A config feature improvement draft. You can directly use the variables in the base config as normal variables virtually in the Python format config file. ## Modification Parse...
**This is a PR for mmcls 1.x** ## Motivation Support MViT ## Modification As the title
I tried to run the test of HRFormer on ImageNet-1k, but the test result was strange. The top-1 accuracy is about 2.0% **Test command** ```bash bash run_eval.sh hrt/hrt_tiny ~/Downloads/hrt_tiny_imagenet_pretrained_top1_786.pth ~/data/imagenet...
## Motivation In the original new config implementation, we use AST transform to manipulate the import statements. It's a little obscure, and hard to debug. ## Modification Python provide a...
### Your current environment VLLM version: 0.85 ### 🐛 Describe the bug The command to serve Qwen3-32b: ``` VLLM_USE_V1=0 vllm serve Qwen/Qwen3-32B --served-model-name qwen3-32b -tp 4 --trust-remote-code --enable-reasoning --reasoning-parser deepseek_r1...