search

valeoai / rangevit

Apache License 2.0
74 stars 6 forks source link

I cannot reproduce the results of the paper #6

Open zc-zhao opened 9 months ago

zc-zhao commented 9 months ago

When i initialize RangeViT’s backbone with ViTs pretrained on supervised ImageNet21k classification and fine-tuned on supervised image segmentation on Cityscapes with Segmenter (entry Cityscapes),I cannot reproduce the results of the paper regarding the nuScenes dataset . My training configuration is as follows: two GeForce RTX 4090 ,the batch-size of each GPU is sixteen .Other configurations follow your config file. The environment is as follows: pytorch==2.0.0,torchvision==1.15.1,cuda==11.7 As a result, the miou only reached 74.27. Could you tell me why this situation happened?

YangParky commented 8 months ago

Hi, have you tested his published model?

zc-zhao commented 8 months ago

Hi, have you tested his published model?

Yes,i have tested his published model and obtained a structure consistent with the paper.But the results of my own training did not meet the requirements of the paper.