A standalone library for adding rotary embeddings to transformers in Tesnorflow, following its success as relative positional encoding. Specifically it will make rotating information into any axis of a tensor easy and efficient, whether they be fixed positional or learned. This library will give you state of the art results for positional embedding, at little costs.
My gut also tells me there is something more to rotations that can be exploited in artificial neural networks.
An implemented version of Pytorch is available in this repository.
This version is written by converting to the version of Pytorch.
The three functions of rearrange, irearrange and repeat have been written due to the incompatibility of the einops library with tensorflow 2.x.
$ pip install rotary-embedding-tensorflowimport tensorflow as tf
from rotary_embedding_tensorflow import apply_rotary_emb, RotaryEmbedding
# instantiate the positional embedding in your transformer and pass to all your attention layers
pos_emb = RotaryEmbedding(dim = 32)
# generate the rotations
freqs = pos_emb(tf.range(1024), cache_key = 1024) # cache with a key that is the sequence length, so that it does not need to recompute
# mock queries and keys
q = tf.random.normal((1, 1024, 64)) # queries - (batch, seq len, dimension of head)
k = tf.random.normal((1, 1024, 64)) # keys
# apply the rotations to your queries and keys after the heads have been split out, but prior to the dot product and subsequent softmax (attention)
freqs = freqs[None, ...] # expand dimension for batch dimension
q = apply_rotary_emb(freqs, q)
k = apply_rotary_emb(freqs, k)
# then do your attention with your queries (q) and keys (k)If you do all the steps above correctly, you should see a dramatic improvement during training
For easy use of 2d axial relative positional embedding, ie. vision transformers
import tensorflow as tf
from rotary_embedding_tensorflow import apply_rotary_emb, RotaryEmbedding, broadcat
pos_emb = RotaryEmbedding(
dim = 32,
freqs_for = 'pixel'
)
# queries and keys for frequencies to be rotated into
q = tf.random.normal((1, 256, 256, 64))
k = tf.random.normal((1, 256, 256, 64))
# get frequencies for each axial
# -1 to 1 has been shown to be a good choice for images and audio
freqs_h = pos_emb(tf.linspace(-1, 1, num = 256), cache_key = 256)
freqs_w = pos_emb(tf.linspace(-1, 1, num = 256), cache_key = 256)
# concat the frequencies along each axial
# broadcat function makes this easy without a bunch of expands
freqs = broadcat((freqs_h[None, :, None, :], freqs_w[None, None, :, :]), dim = -1)
# rotate in frequencies
q = apply_rotary_emb(freqs, q)
k = apply_rotary_emb(freqs, k)For injecting learned rotations into a network. Experiments pending
Update: doesn't seem to do anything -_-, will keep trying...
import tensorflow as tf
from tensorflow.keras import layers
from rotary_embedding_tensorflow import apply_learned_rotations
x = tf.random.normal((1, 1024, 512))
# you can only rotate in (dim // 2) values
# ex. for 512, you can only rotate in 256 values
# say you have two sets of learned rotations of 128 values each
rots1 = layers.Dense(128)(x)
rots2 = layers.Dense(128)(x)
# you rotate in 256 (128 x 2) at first
x = apply_learned_rotations(rots1, x, start_index = 0)
# then you start at index 256 and rotate in the last (128 x 2)
x = apply_learned_rotations(rots2, x, start_index = 256)
# you could also concat the rotations together and pass it in all at once
rots = tf.concat((rots1, rots2), axis = -1)
x = apply_learned_rotations(rots, x)@misc{su2021roformer,
title = {RoFormer: Enhanced Transformer with Rotary Position Embedding},
author = {Jianlin Su and Yu Lu and Shengfeng Pan and Bo Wen and Yunfeng Liu},
year = {2021},
eprint = {2104.09864},
archivePrefix = {arXiv},
primaryClass = {cs.CL}
}
@misc{rotary-embedding-torch,
title = {Rotary Embeddings - Pytorch},
author = {Phil Wang (lucidrains)},
year = {2021},
url = {https://github.com/lucidrains/rotary-embedding-torch},
publisher = {Github},
}