search

gaopengcuhk / Stable-Pix2Seq

A full-fledged version of Pix2Seq
Apache License 2.0
235 stars 20 forks source link

Performance Difference Compared With Original Paper？ #5

Closed xmu-xiaoma666 closed 2 years ago

xmu-xiaoma666 commented 2 years ago

Thank you for your implementation!!! Can you share Performance Difference Compared With Original Paper and checkpoint?

gaopengcuhk commented 2 years ago

I am still running experiments.

Yuliang-Liu commented 2 years ago

Is there any update of the performance?

gaopengcuhk commented 2 years ago

By adding new augmentation, stable-pix2seq could achieve 34mAP at 218 epoches

xmu-xiaoma666 commented 2 years ago

By adding new augmentation, stable-pix2seq could achieve 34mAP at 218 epoches

The paper reports 45.0mAP？ image

gaopengcuhk commented 2 years ago

The reported performance is 43.0 mAP. The paper lack details about the implementation. We are trying our best to replicate the results reported in the paper. As the training process is slow, it may take several months to replicate the results. We will update the results on this GitHub once we have progress,

gaopengcuhk commented 2 years ago

256 epoch 36mAP. Beam search will addition improve 1-2 mAP. The estimated mAP is 37-38mAP at 256 epoch.

seanzhuh commented 2 years ago

Will you release the beam search code?

gaopengcuhk commented 2 years ago

All code including beam search and data augmentation will be released.

baiyongrui commented 2 years ago

Does Pix2Seq converge slower than DETR?

gaopengcuhk commented 2 years ago

Due to different learning schedules, it's hard to compare the convergence speed between DETR and Pix2Seq.

gaopengcuhk commented 2 years ago

We released the pretrained model at https://github.com/gaopengcuhk/Pretrained-Pix2Seq.