pcnn | 使用pcnn网络进行关系分类，中文注解，有数据集，可跑通。附带绘图

I'm currently trying to implement a piecewise max pooling operation in Tensorflow, as described here. Given a sentence, I want to divide it to three different portions and max pool all of those portions separately, so that I'd end up with 3 different values instead of 1.

More concretely, I have a tensor training of shape [batch_len, 1, sentence_len, feature_len]. I also have another tensor splits of shape [batch_len, 2], where the first element of any row is the index to split off the first portion, and the second element is the index to split off the last portion. I want to index the training tensor in a way that divides it into tree parts based on the value index values provided in the splits tensor.

We cannot simply index the training tensor using the other tensor, as we have different lengths for the first, second, and third portions for different examples. I could loop through all the training data and do it that way, but that would be horribly inefficient. I want to make this as efficient as possible.

Note: since they will be max pooled, I'm fine with having 3 different tensors of shape [batch_len, 1, sentence_len, feature_len], where in the first tensor, only the elements in the first portion of each sentence has values, and the others have zero. The second tensor would only have values in the middle part, and so on.