关于数据集和代码的一些问题

[lileilai]

你好～ 在恢复实验的过程中，碰到了问题，期望你能解答一些。 根据您给定的代码和预训练的模型，在数据集PIPA上进行测试，出现如下的错误: self.class.name, "\n\t".join(error_msgs))) RuntimeError: Error(s) in loading state_dict for GRM: Missing key(s) in state_dict: "ggnn.reason_fc_x.bias", "ggnn.reason_fc_x.weight", "ggnn.reason_fc_y.bias", "ggnn.reason_fc_y.weight", "ggnn.reason_fc2.bias", "ggnn.reason_fc2.weight". size mismatch for classifier.1.weight: copying a param of torch.Size([4096, 41472]) from checkpoint, where the shape is torch.Size([4096, 49152]) in current model. size mismatch for classifier.4.bias: copying a param of torch.Size([1]) from checkpoint, where the shape is torch.Size([16]) in current model. size mismatch for classifier.4.weight: copying a param of torch.Size([1, 4096]) from checkpoint, where the shape is torch.Size([16, 4096]) in current model.

预训练模型和您给定的代码有出入，请问是这样吗？ 如果是的话，能提供一份最新的模型代码吗？

[wzhouxiff]

Hi， I am sorry that there was some wrong in the old model, and I have updated it in the same place. Download and try it.

[lileilai]

师兄 /师姐 ，你好～

目前在follow您发表在ijcai上关于社会关系抽取的这篇文章，打算作为毕业论文，有一个问题，希望能得到您的帮助

PIPA数据集中没有原图的person-pair的两个box坐标，只是给了single_body的图片。

---

在您给的测试集中的数据

72157624551655535_4870147689.jpg 375 333 749 767 863 183 1023 548 8

请问您能提供下PIPA中peroson-pair (body)的两个box坐标的这个数据集(train/val)吗？

谢谢～～～

------------------ Original ------------------ From: "wzhouxiff"notifications@github.com; Date: Tue, Jan 8, 2019 03:43 PM To: "wzhouxiff/SR"SR@noreply.github.com; Cc: "leilai li"lillai@mail2.sysu.edu.cn; "Author"author@noreply.github.com; Subject: Re: [wzhouxiff/SR] 关于数据集和代码的一些问题 (#2)

Hi， I am sorry that there was some wrong in the old model, and I have updated it in the same place. Download and try it.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or mute the thread.

[wzhouxiff]

https://pan.baidu.com/s/13tvWT5FmfvIFaBRE9nq1WQ#list/path=%2FSR%2Flist

Get them from here.

[lileilai]

https://pan.baidu.com/s/13tvWT5FmfvIFaBRE9nq1WQ#list/path=%2FSR%2Flist

Get them from here.

i notice that there does not existing the the training code. could you provide it ? because i could not know the details of your model, such as the learning rate, number of epoch, optimization method. And it‘s also not mentioned in the experiment of paper.

Looking forward to your reply again～～

[lileilai]

Hi~ sorry to disturb you again~~ i try to coding a training function to reproduce your experiment result, and the result as follows:

---

frozen: resnet101,vgg16 batch_size=25 epch=7 lr=0.0001

tp: {0: 141, 1: 115, 2: 111} p: {0: 231, 1: 205, 2: 184} r: {0: 226, 1: 190, 2: 204} precision: {0: 0.6103896103896104, 1: 0.5609756097560976, 2: 0.6032608695652174} recall: {0: 0.6238938053097345, 1: 0.6052631578947368, 2: 0.5441176470588235}

• Prec@1 58.598 * Loss 0.9104 i still not achive the result as same your result... can you provide some advices. Thanks you

Looking foward your reply again

[stzhang1994]

Hope to get the training code of your model too. Especially the training hyper-parameters and opitimization method, thanks.