Code Repository for the paper: "Affect2MM: Affective Analysis of Multimedia Content Using Emotion Causality".
Training and Testing Codes for all three datasets.
- SENDv1 Dataset
- MovieGraphs Dataset
- LIRIS-ACCEDE Dataset
- current
- GC 输出全部是0
- co-attention 权重输出全部是average weights
-
- (32,10) attention 是对时间维度
- 模型最后一层输出如下,模型似乎没有学习能力
-
- predict_last......
- tensor(-0.0411, 0.1370, -0.1754, -0.0580, -0.0578, 0.1431, 0.1562, -0.0996,0.1616
-
- dec_out .....
- 1.3158e-01, -2.6054e-02, 2.1075e-01, 1.3412e-01, 7.6706e-02, 1.9719e-01, 2.0428e-01, 5.6338e-02, 4.5805e-02, 6.5245e-02,
-
- 通过查阅资料,模型最后一层通过softmax之后输出有负值很正常
- KLDivLoss 没有随着epoch而减小,并没有什么变化
-
- 已经修改,看结果如何!!!
- 完成
-
criterion 未添加
-
- MultiLabelMarginLoss
- MultiLabelSoftMarginLoss
-
GCN + label correlation未添加
-
- model 相对简单
-
dominant emotion loss founction
-
pred = torch.from_numpy(pred).float()
-
label = torch.from_numpy(label).float()
-
crition1 = torch.nn.BCEWithLogitsLoss()
-
- loss1 = crition1(pred, label)
-
- print(loss1) #0.7193
-
- crition2 = torch.nn.MultiLabelSoftMarginLoss()
-
- loss2 = crition2(pred, label)
-
- print(loss2) #0.7193
-
- 两个criterion达到的效果一样
-
已经添加
-
dominant emotion criterion 未添加
-
- top 1 top 3 top 5
- precision
- 根据代码,跟其保持一致即可
-
以上全部完成
-
对于co-attention模型:
-
修改模型:只得到 context vector,不再经过decoder layer
-
完成:但是也没有觉得结果变好
-
关于 top-3 OR_3、CR_3 的结果不如 OR、CR 结果好:
-
learning spatial regularization with image-level supervision for multi-label image classification 文章中的结果
-
top-3 label 的确低于 all label