Funnel Transformer¶
Overview¶
The Funnel Transformer model was proposed in the paper Funnel-Transformer: Filtering out Sequential Redundancy for Efficient Language Processing. It is a bidirectional transformer model, like BERT, but with a pooling operation after each block of layers, a bit like in traditional convolutional neural networks (CNN) in computer vision.
The abstract from the paper is the following:
With the success of language pretraining, it is highly desirable to develop more efficient architectures of good scalability that can exploit the abundant unlabeled data at a lower cost. To improve the efficiency, we examine the much-overlooked redundancy in maintaining a full-length token-level presentation, especially for tasks that only require a single-vector presentation of the sequence. With this intuition, we propose Funnel-Transformer which gradually compresses the sequence of hidden states to a shorter one and hence reduces the computation cost. More importantly, by re-investing the saved FLOPs from length reduction in constructing a deeper or wider model, we further improve the model capacity. In addition, to perform token-level predictions as required by common pretraining objectives, Funnel-Transformer is able to recover a deep representation for each token from the reduced hidden sequence via a decoder. Empirically, with comparable or fewer FLOPs, Funnel-Transformer outperforms the standard Transformer on a wide variety of sequence-level prediction tasks, including text classification, language understanding, and reading comprehension.
Tips:
Since Funnel Transformer uses pooling, the sequence length of the hidden states changes after each block of layers. The base model therefore has a final sequence length that is a quarter of the original one. This model can be used directly for tasks that just require a sentence summary (like sequence classification or multiple choice). For other tasks, the full model is used; this full model has a decoder that upsamples the final hidden states to the same sequence length as the input.
The Funnel Transformer checkpoints are all available with a full version and a base version. The first ones should be used for
FunnelModel,FunnelForPreTraining,FunnelForMaskedLM,FunnelForTokenClassificationand class:~transformers.FunnelForQuestionAnswering. The second ones should be used forFunnelBaseModel,FunnelForSequenceClassificationandFunnelForMultipleChoice.
The original code can be found here.