[Question] 标签一定要str类型吗，可以拿来做情感分析吗 about kashgari HOT 15 CLOSED

多标签分类的模型还没有增加，可以提供详细的数据集链接，我会看情况增加相关任务。

from kashgari.

@BrikerMan 这是数据集链接 链接：https://pan.baidu.com/s/1spbM8QuyjRmgCSa9i-P7Mw
提取码：679y
只有7天有效哦

from kashgari.

这个问题并不是单个维度的多标签，感觉需要专门建模处理才可以，我这里能支持的是 [(1, 2), (3,4),(5,)] 这种简单的多标签多分类。

from kashgari.

这样的多分类能满足需求嘛，y 是由多个标签组成的~

from kashgari.

A multi label classification problem could be solved by training a model with sigmoid as activation function and binary-crossentropy as loss function. The output is an n-dim one-hot vector to predict n possible labels.
That's it.

from kashgari.

@alexwwang @Owenscu check this #29 out, is this looks good to you? I am planning to add multi_label classification, but it seems we had to change all the classification models to support this feature.
Dataset is from http://tcci.ccf.org.cn/conference/2018/taskdata.php task 1.

from kashgari.

Is it possible to wrap up a given nn models in a multi-label task model? I mean just replace the last output layer to fit with different kind of classification tasks and allow users to determine which to choose.

Meanwhile I am considering allowing users setting hyper_parameters while initializing a model class in the model zoo.

I think these two aspects could be put together, so the flexibility and concision could both be kept.

from kashgari.

I think your solution is better than mine, please make the changes and submit a pull request~ @alexwwang

from kashgari.

Ok but I'm afraid it would take some time. Fighting the shape bug yet.

from kashgari.

Take your time, no need to rush.

from kashgari.

#34 This commit fulfilled this need. By passing sigmoid and binary_crossentropy hyper_parameters to init function of a classification model, you could get one model support multi-label vector output.

from kashgari.

@alexwwang I think multi-label is not finished yet, still need to change here

from kashgari.

@BrikerMan Yeah, the data padding work. The nearest approach, I think, maybe with the help of sklearn.preprocessing.MultiLabelBinarizer. And this tool could also deal with multi-class classification as a specific type of multi-label classification. Or just add up another switch to approach multi-label y-vector, leaving current part unchanged.
Whatever, seems there's no shortcut here in data preprocessing and prediction processing, if we want to keep the predict confidence/probability of each label/class.
How do you think?

from kashgari.

Please check out the #29, I have used sklearn.preprocessing.MultiLabelBinarizer and rewrite the data-process and predict function.

from kashgari.

Kashgari supported multi-label classification now, change y from string to a string list, then set multi_label=True While init model class.

from kashgari.