pykernsformer 0.0.4

Kernel attention implementation of Pytorch TransformerEncoderLayer

pykernsformer

The pykernsformer module extends the torch.nn.TransformerEncoderLayer class to include custom attention formulas.

Installation

You can install the pykernsformer package using pip as

pip install pykernsformer

Usage

pykernsformer comes with the following built in attention kernels.

pykernsformer	Attention	Formula	Citation
attention	Regular	$softmax(\frac{QK^T}{\sqrt{d_k}})V$	Vaswani et al.
attention_linear	Linear	$\frac{QK^T}{\sum_k QK^T}V$
attention_periodic	Periodic	$softmax(-\frac{2\sin^2(\pi\frac{\sqrt{2 - 2q_ik_j^T}}{p})}{\sqrt{d_k}})V$
attention_LP	Locally Periodic	$softmax(-\frac{2\sin^2(\pi\frac{\sqrt{2 - 2\hat{q}_i\hat{k}_j^T}}{p})}{\sqrt{d_k}} + \frac{{q_i}{k_j^T}}{\sqrt{d_k}})V$
attention_RQ	Rational Quadratic	$\frac{\left( 1 + \frac{1}{\alpha \sqrt{d_k}} - \frac{2QK^T}{2 \alpha \sqrt{d_k}} \right)^{-\alpha}}{\sum_k \left( 1 + \frac{1}{\alpha \sqrt{d_k}} - \frac{2QK^T}{2 \alpha \sqrt{d_k}} \right)^{-\alpha}}V$

You can also implement your own attention function with the following signature:

def attention_custom(query, key, value, mask=None, dropout=None):

    [...]

    p_attn = [...] # the attention matrix

    [...]

    return torch.matmul(p_attn, value), p_attn