Li Xuhong

icon indicating a website link https://sites.google.com/view/lixuhong

icon company Baidu Inc. icon location Beijing icon location Focusing on interpretable deep learning and transfer learning.

Results 6 issues of Li Xuhong

Assets Uploading Place

7 comment

Some training details for cityscapes

Can I ask for some training detail? The model for this > PSPNet101 on cityscapes valset (mIoU/pAcc): 79.70/96.38 (ss) and 80.91/96.59 (ms) is pre-trained only on ImageNet or on ImageNet...

Unsupported gpu architecture

2 comment

Environment: Ubuntu 18.04, CUDA 11.0, PyTorch 1.7.1, A100. When installed from source by `python setup.py install`, an nvcc fatal error raised: ``` nvcc fatal : Unsupported gpu architecture 'compute_80' RuntimeError:...

customized input data

## 🚀 Feature Input data processing needs to be flexible and can be customized. ## Motivation and Additional context Currently, images are normalized according to natural images, like ImageNet statistics....

enhancement

Gradient computation changed since paddle2.4

Thanks @CarlZhang1020 for reporting the issue and the [PR](https://github.com/PaddlePaddle/InterpretDL/pull/51). See the modification in the [PR](https://github.com/PaddlePaddle/InterpretDL/pull/51) for example of gradcam. > grad method changed in paddle2.4, need to check more code...

Plan to release models? Considering using CAM/IntGrad/SmoothGrad to visualize some images

Very interesting work! Do authors have any plan to release one model? I am thinking whether using the input feature importance methods (CAM/IntGrad/SmoothGrad) helps to locate the patterns that distinguish...