Zhouxia Wang comments

Results 30 comments of


                                            Zhouxia Wang

failed to install matmul_cuda

I can install the matmul_cuda succeededly. However, I meet Segmentation fault. Anyone knows how to fix it?

关于数据集和代码的一些问题

Hi， I am sorry that there was some wrong in the old model, and I have updated it in the same place. Download and try it.

关于数据集和代码的一些问题

https://pan.baidu.com/s/13tvWT5FmfvIFaBRE9nq1WQ#list/path=%2FSR%2Flist Get them from here.

Training settings in the paper differ from the code

Please follow the setting described in the paper. Note that the learning rate set in the config is not the actual learning rate. It will be divided by the number...

When use the HQ dictionary for MHCA.

MHCA (implemented as MultiHeadAttnBlock in RestoreFormer.modules.vqvae.vqvae_arch.py ) has been merged into the decoder (MultiHeadDecoderTransformer in RestoreFormer.modules.vqvae.vqvae_arch.py).

HQ_Dictionary model and RestoreFormer model use different number of attention heads

Thanks for your reminder. I have revised it to 8.

The training data is [FFHQ](https://github.com/NVlabs/ffhq-dataset) which contains 70000 HD images. The corresponding degraded images are synthesized with RestoreFormer/data/ffhq_dagradation_dataset.py

关于数据集

https://www.mpi-inf.mpg.de/departments/computer-vision-and-multimodal-computing/research/human-activity-recognition/social-relation-recognition/

Training time of the RestoreFormer++ model

The training details are outlined below: - Step 1: Achieve ROHQD. We fine-tune the model provided in [taming-transformers](https://github.com/CompVis/taming-transformers) using FFHQ with an input size of 512x512. It is important to...

Training time of the RestoreFormer++ model

4 V100s with batchsize = 4 for each gpu.