Accuracy issues

[zzk-nb]

After I reproduced it according to your method, it was a bit different from the accuracy given in the paper, so what step did go wrong?

[JoyeZLearning]

WOW！ Congratulations! It seems your result is better than mine slightly, I guess maybe the network exists some fluctuation and the initial random boxes also may influence the results.

[Chris666-sudo]

@zzk-nb How did you solve problem “Value Error：cannot convert float NaN to int ” And my result is so bad like this，maybe something wrong in my training process

[zzk-nb]

哇！ 恭喜！看来你的结果比我的稍微好一点，我想也许网络存在一些波动，最初的随机框也可能影响结果。

Hello, the result only AP50 achieved a good effect, the other accuracy difference is a little big, what is the reason? In addition, I would like to ask what number of iterations you set when training the model. My training results in my sar image data set are not ideal, is it because of the number of iterations? Looking forward to your reply.

[JoyeZLearning]

【the result only AP50 achieved a good effect, the other accuracy difference is a little big, 】You mean the specific category mAP? because the results show upside is the AP50:95. and if you want to know the category AP of 50, you can write a new code to see the AP under different threshold.

As for your own data, can you tell me what dataset you used?

The iterations I use for my dataset is 30000, approximately 78 epochs

[JoyeZLearning]

Besides, the SAR-AIRcraft1.0 dataset maybe not a perfect dataset, for I found the annotations of test dataset is so bad, so I use the val dataset to evaluate my DiffDet4SAR. and the algorithms I compared in my papar are also evaluated on val dataset.

[zzk-nb]

【结果只有AP50取得了不错的效果，其他精度相差有点大，】你是说具体类别的mAP？因为结果显示上行空间是 AP50：95。如果你想知道 AP 的类别为 50，你可以编写一个新代码来查看不同阈值下的 AP。

至于你自己的数据，你能告诉我你用的是什么数据集吗？

我用于数据集的迭代次数为 30000 次，大约 78 个 epoch

我使用的是HRSID数据集，类别只有一个，我迭代10000次的时候ap50在45左右不再增长，您的STEPS设置的是多少呢，这个会影响训练结果吗

[zzk-nb]

@zzk-nb How did you solve problem “Value Error：cannot convert float NaN to int ” And my result is so bad like this，maybe something wrong in my training process

I was carried out in accordance with the author step repetition, you tried to modify detectron2 / engine/defaults. Py the path inside, I the results of the repetition of just using the file to assess the weights of the author, the result of my training is not good.

[JoyeZLearning]

category is 1, then you should modify the num_classes and the image size in config.yaml... according to your own dataset. also the number of initial boxes（may 100 or 200 is enough?）, and also the mean and std in pre-processing? and so on For I design the parameter of the net on the basis of my aircraft dataset, when you use to your dataset, there may be some modifications you need to do.

I guess aircraft and ship have many difference, and you can reflect on the key point of ship and adjust the code to fit the ship datasets.

[zzk-nb]

category 为 1，则应修改 config.yaml 中的num_classes和图像大小...根据您自己的数据集。还有初始框的数量（可能 100 或 200 个就足够了？），以及预处理中的平均值和 STD？依此类推，因为我根据我的飞机数据集设计了网络的参数，当你使用你的数据集时，你可能需要做一些修改。

我猜飞机和船有很多区别，你可以反思船的关键点，调整代码以适应船的数据集。 Thank you very much for your reply. I would like to ask how many STEPS you set during training and whether this will affect the results of training.

[spikeeeeeeeeee]

@Chris666-sudo @zzk-nb How did you solve problem “Value Error：cannot convert float NaN to int ”, it confused me terribly, i didn't know what's wrong

[yangyahu-1994]

@Chris666-sudo @zzk-nb How did you solve problem “Value Error：cannot convert float NaN to int ”, it confused me terribly, i didn't know what's wrong

hello,have you solved this problem?