Dynamic pooling (Or variable-size pooling?)

[celinaum]

Is there a way of performing dynamic pooling in Keras? (1D CNN)
Like Goldberg elaborates: Splitting the resulting output from the convolution (conv layer) into r groups, thus apply pooling separately on each group, where the groups can be of different sizes.
Or is there a way of performing variable-size pooling?

Example:
Given a text of sentences, text = [sentence1, sentence2, sentence3], where sentence1 is of length 20, sentence2 of length 15, and sentence3 of length 18 (counted in words).
Is there a way of grouping the convolution associated with the words from sentence1, sentence2, and sentence3? Thus apply pooling at each group.
...
conv = Conv1D(filters=100 kernel_size=2, strides=1, padding="valid", activation="relu")(training)
pool_0 = MaxPooling1D(pool_size=20, stride=20)(conv[0])
pool_1 = MaxPooling1D(pool_size=15, stride=15)(conv[1])
pool_2 = MaxPooling1D(pool_size=18, stride=18)(conv[2])
traning = concatenate([pool_0, pool_1, pool_2)
...

Or performing variable-size pooling, a structure like:
...
conv = Conv1D(filters=100 kernel_size=2, strides=1, padding="valid", activation="relu")(training)
pool = MaxPooling1D(pool_size=[20, 15, 18], stride=[20, 15, 18])(conv)
traning = concatenate([pool[0], pool[1], pool[2])
...

Thanks.

[rjurney]

I need this as well to implement XML-CNN.