search

BR-IDL / PaddleViT

:robot: PaddleViT: State-of-the-art Visual Transformer and MLP Models for PaddlePaddle 2.0+
https://github.com/BR-IDL/PaddleViT
Apache License 2.0
1.22k stars 318 forks source link

add(cvt): impl cvt for paddle #121

Closed defensetongxue closed 2 years ago

defensetongxue commented 2 years ago

给paddle复现CvT论文。

xperzy commented 2 years ago

非常棒！请问精度是否和官方对齐？我这边readme 没有看见相关的信息

我正在改这部分，目前基本对齐，还需要进一步改代码风格和对齐更多模型@defensetongxue

xperzy commented 2 years ago

Now the cvt_13 ported weights obtains correct performance on Imagenet (81.60%).

There are still several issues should be finished before merging:

  1. The implementations in cvt.py could be simplified, please refer to our ViT implementation (here)
  2. Please convert more ported weights form the original repo and align the accuracies.
  3. If possible, please also make the training settings same as the original paper.
defensetongxue commented 2 years ago

ok,i'll simplify the model soon

defensetongxue commented 2 years ago

I have simplifed the code style and test the model cvt_13_384x384-1k(0.8275) and cvt_21_224x224-1k(0.8252).please check the code style if it can meet the requests ,thanks @xperzy

xperzy commented 2 years ago

I have tested the performances and now the results are same as official repo. So I merge this PR.