search

danfenghong / IEEE_TGRS_SpectralFormer

Danfeng Hong, Zhu Han, Jing Yao, Lianru Gao, Bing Zhang, Antonio Plaza, Jocelyn Chanussot. Spectralformer: Rethinking hyperspectral image classification with transformers, IEEE Transactions on Geoscience and Remote Sensing (TGRS), 2021
224 stars 38 forks source link

reproduction of results #4

Open MichealRay opened 2 years ago

MichealRay commented 2 years ago

Hi danfeng, Very good job. Congraqulations! And thank you for sharing codes. I have a problem when I want to reproduce results of Houston2013. The training detail log is shown as below: Running Time: 1737.75

Final result: OA: 0.7679 | AA: 0.7864 | Kappa: 0.7486 [0.83095916 0.82894737 0.80594059 0.94886364 0.99621212 0.93706294 0.80970149 0.51377018 0.74032106 0.44980695 0.66223909 0.79634966 0.73333333 0.90688259 0.83509514]

Parameter: dataset: Houston mode: CAF gpu_id: 0 seed: 0 batch_size: 64 test_freq: 5 patches: 7 band_patches: 3 epoches: 1000 learning_rate: 0.0005 gamma: 0.9 weight_decay: 0

And the result in your paper is shown as below: houston2013_result

As you can see there is a big gap between OA: 0.7679 | AA: 0.7864 | Kappa: 0.7486 and OA: 0.8801 | AA: 0.8891 | Kappa: 0.8699

So what I miss? Could you please help me?

jingyao16 commented 2 years ago

Hi, @MichealRay . Thanks for your interest. You can help yourself by spending more time reading the implementation details in the paper. More supplementary files will be uploaded soon for the inexperienced to reproduce all the results. Please kindly keep an eye on our repository.