你好，RSP-ResNet-50-E300给出的模型是训练好的权重吗？

[lucky85712]

我用于其他的航空数据集目标检测，性能与imagenet给出的预训练模型的性能相差很多，想问一下这个百度网盘给出的权重是已经训练好的吗？

[DotWang]

@lucky85712 是预训练好的权重，如果精度差很多的话，麻烦检查一下权重是不是真的加载进去了

[lhmtrinh]

Hi @DotWang I have the same problem:
I loaded the model checkpoint as follow and attempt to predict on several images of MillionAid dataset. They are not correct. Can you please help?

I verified that the weights are indeed loaded:

Many many thanks.

[lucky85712]

Sorry, I had some issues loading this model before, but later on, I didn't use this weight anymore, so I may not be able to help you.

…

------------------ 原始邮件 ------------------ 发件人: "ViTAE-Transformer/RSP" ***@***.***>; 发送时间: 2023年11月23日(星期四) 晚上9:29 ***@***.***>; ***@***.******@***.***>; 主题: Re: [ViTAE-Transformer/RSP] 你好，RSP-ResNet-50-E300给出的模型是训练好的权重吗？ (Issue #8) Hi I have the same problem: I loaded the model checkpoint as follow and attempt to predict on several images of MillionAid dataset. They are not correct. Can you please help? Many many thanks. — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you were mentioned.Message ID: ***@***.***>

[lhmtrinh]

did you end up using anything else? :)

[lucky85712]

resnet50.pth

…

------------------ 原始邮件 ------------------ 发件人: "ViTAE-Transformer/RSP" ***@***.***>; 发送时间: 2023年11月23日(星期四) 晚上9:35 ***@***.***>; ***@***.******@***.***>; 主题: Re: [ViTAE-Transformer/RSP] 你好，RSP-ResNet-50-E300给出的模型是训练好的权重吗？ (Issue #8) did you end up using anything else? :) — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you were mentioned.Message ID: ***@***.***>

[lhmtrinh]

@lucky85712 Ok thanks. It won't help me much, however.

[DotWang]

@lhmtrinh https://github.com/ViTAE-Transformer/RSP/tree/main/Object%20Detection

我不确定你是拿什么程序跑的，我们是拿OBBDetection，这里有config和log

[littlelight-hhr]

嗨，我有同样的问题：我按如下方式加载了模型检查点，并尝试对 MillionAid 数据集的几张图像进行预测。它们是不正确的。你能帮忙吗？

我验证了权重确实已加载：

非常感谢。

Hi, do you solve the problem?

[WenLinLliu]

我也发现给出来的预训练权重相比于imagenet差很多

[DotWang]

@lucky85712 @lhmtrinh @littlelight-hhr @WenLinLliu

After checking, we have found that the issue is caused by the saved checkpoint, which contains model weight, optimizer parameter. and so on, as shown in https://github.com/ViTAE-Transformer/RSP/blob/main/Scene%20Recognition/main.py/#L214

So, when using the uploaded checkpoint, e.g., RSP-ResNet-50-E300.pth, please extract the model weight manually

In addition, we have validated the performance training on a single GPU, where the batch size for DOTA is 2, as presented in the RSP paper.