Created by: lcy-seso

Problem:

• Currently, sequence Pooling in PaddlePaddle is performed sequence-wise. After pooling, a sequence is transformed into a non-sequence.
• This is not enough in building a deep one-dimensional convolution model, such as models for NLP or speech tasks, in which pooling layer may be followed by a new convolution layer.
• According to current design, such a model is impossible, because output of pooling is no longer a sequence, that cannot be input to a new convolution layer.

It will be helpful if the sequence Pooling layer provides a stride parameter, which controls the scope of pooling operation.

It may work like this:

• If stride is set -1, it equals to sequence-wise pooling as it does now, otherwise, pooling is performed upon a small local area.
• After pooling with stride n (n is smaller than sequence length), a long sequence will be shorten.