Awesome
Routing Transformer
<img src="./routing_attention.png" width="500px"></img>
A fully featured implementation of <a href="https://arxiv.org/pdf/2003.05997.pdf">Routing Transformer</a>. The paper proposes using k-means to route similar queries / keys into the same cluster for attention.
131k tokens
Install
$ pip install routing_transformer
Usage
A simple language model
import torch
from routing_transformer import RoutingTransformerLM
model = RoutingTransformerLM(
num_tokens = 20000,
dim = 512,
heads = 8,
depth = 12,
max_seq_len = 8192,
causal = True, # auto-regressive or not
emb_dim = 128, # embedding factorization, from Albert
weight_tie = False, # weight tie layers, from Albert
tie_embedding = False, # multiply final embeddings with token weights for logits
dim_head = 64, # be able to fix the dimension of each head, making it independent of the embedding dimension and the number of heads
attn_dropout = 0.1, # dropout after attention
attn_layer_dropout = 0., # dropout after self attention layer
ff_dropout = 0.1, # feedforward dropout
layer_dropout = 0., # layer dropout
window_size = 128, # target window size of each cluster
n_local_attn_heads = 4, # number of local attention heads
reversible = True, # reversible networks for memory savings, from Reformer paper
ff_chunks = 10, # feed forward chunking, from Reformer paper
ff_glu = True, # use GLU variant in feedforward
pkm_layers = (4, 7), # specify layers to use product key memory. paper shows 1 or 2 modules near the middle of the transformer is best
pkm_num_keys = 128, # defaults to 128, but can be increased to 256 or 512 as memory allows
moe_layers = (3, 6), # specify which layers to use mixture of experts
moe_num_experts = 4, # number of experts in the mixture of experts layer, defaults to 4. increase for adding more parameters to model
moe_loss_coef = 1e-2, # the weight for the auxiliary loss in mixture of experts to keep expert usage balanced
num_mem_kv = 8, # number of memory key/values to append to each cluster of each head, from the 'All-Attention' paper. defaults to 1 in the causal case for unshared QK to work
use_scale_norm = False, # use scale norm, simplified normalization from 'Transformers without Tears' paper
use_rezero = False, # use Rezero with no normalization
shift_tokens = True # shift tokens by one along sequence dimension, for a slight improvement in convergence
).cuda()
x = torch.randint(0, 20000, (1, 8192)).long().cuda()
input_mask = torch.ones_like(x).bool().cuda()
y, aux_loss = model(x, input_mask = input_mask) # (1, 8192, 20000)
aux_loss.backward() # add auxiliary loss to main loss before backprop
A simple transformer
import torch
from routing_transformer import RoutingTransformer
model = RoutingTransformer(
dim = 512,
heads = 8,
depth = 12,
max_seq_len = 8192,
window_size = 128,
n_local_attn_heads = 4
).cuda()
x = torch.randn(1, 8192, 512).cuda()
input_mask = torch.ones(1, 8192).bool().cuda()
y, aux_loss = model(x, input_mask = input_mask) # (1, 8192, 512)
aux_loss.backward() # add auxiliary loss to main loss before backprop
Encoder Decoder
To use a full encoder, decoder, simply import the RoutingTransformerEncDec class. Save for the dim keyword, all other keywords will be either prepended with enc_ or dec_ for the encoder and decoder RoutingTransformerLM class respectively.
import torch
from routing_transformer import RoutingTransformerEncDec
model = RoutingTransformerEncDec(
dim=512,
enc_num_tokens = 20000,
enc_depth = 4,
enc_heads = 8,
enc_max_seq_len = 4096,
enc_window_size = 128,
dec_num_tokens = 20000,
dec_depth = 4,
dec_heads = 8,
dec_max_seq_len = 4096,
dec_window_size = 128,
dec_reversible = True
).cuda()
src = torch.randint(0, 20000, (1, 4096)).cuda()
tgt = torch.randint(0, 20000, (1, 4096)).cuda()
src_mask = torch.ones_like(src).bool().cuda()
tgt_mask = torch.ones_like(tgt).bool().cuda()
loss, aux_loss = model(src, tgt, enc_input_mask = src_mask, dec_input_mask = tgt_mask, return_loss = True, randomly_truncate_sequence = True)
loss.backward()
aux_loss.backward()
# do your training, then to sample up to 2048 tokens based on the source sequence
src = torch.randint(0, 20000, (1, 4096)).cuda()
start_tokens = torch.ones(1, 1).long().cuda() # assume starting token is 1
sample = model.generate(src, start_tokens, seq_len = 2048, eos_token = 2) # (1, <= 2048, 20000)
Product Key Memory
To see the benefits of using PKM, the learning rate of the values must be set higher than the rest of the parameters. (Recommended to be 1e-2)
You can follow the instructions here to set it correctly https://github.com/lucidrains/product-key-memory#learning-rates
Kmeans Hyperparameters
kmeans_ema_decay = {defaults to 0.999}
This is the exponential moving average decay for updating the k-means. The lower this is, the faster the means will adjust, but at the cost of stability.
commitment_factor = {defaults to 1e-4}
The weight of the auxiliary loss that encourages tokens to get closer (commit) to the k-mean centroids that were chosen for them.
Updating kmeans manually
The following instructions will allow you to update the kmeans manually. By default the kmeans are updated automatically on every backward pass.
import torch
from routing_transformer import RoutingTransformerLM, AutoregressiveWrapper
model = RoutingTransformerLM(
num_tokens = 20000,
dim = 1024,
heads = 8,
depth = 6,
window_size = 256,
max_seq_len = 8192,
causal = True,
_register_kmeans_update = False # set to False to disable auto-updating
)
model = AutoregressiveWrapper(model)
x = torch.randint(0, 20000, (1, 8192))
loss = model(x, return_loss = True)
loss.backward()
# update kmeans with this call
model.update_kmeans()
Issues
This architecture has trouble generalizing to shorter sequence lengths when decoding tokens from 1 -> maximum sequence length. The simplest and surest solution is to randomly truncate the sequence during training. This helps the network and the kmeans generalize to variable number of tokens, at the cost of prolonged training.
If you are priming the network with the full sequence length at start, then you will not face this problem, and you can skip this training procedure.
import torch
from routing_transformer import RoutingTransformerLM, AutoregressiveWrapper
model = RoutingTransformerLM(
num_tokens = 20000,
dim = 1024,
heads = 8,
depth = 12,
window_size = 256,
max_seq_len = 8192,
causal = True
)
model = AutoregressiveWrapper(model)
x = torch.randint(0, 20000, (1, 8192))
loss = model(x, return_loss = True, randomly_truncate_sequence = True) # (1, 8192, 20000)
Appreciation
Special thanks to <a href="https://github.com/AranKomat">Aran Komatsuzaki</a> for bootstrapping the initial implementation in Pytorch that evolved into this library.
Citation
@misc{roy*2020efficient,
title = {Efficient Content-Based Sparse Attention with Routing Transformers},
author = {Aurko Roy* and Mohammad Taghi Saffar* and David Grangier and Ashish Vaswani},
year = {2020},
url = {https://arxiv.org/pdf/2003.05997.pdf}
}
@misc{shazeer2020glu,
title = {GLU Variants Improve Transformer},
author = {Noam Shazeer},
year = {2020},
url = {https://arxiv.org/abs/2002.05202}
}
@inproceedings{kitaev2020reformer,
title = {Reformer: The Efficient Transformer},
author = {Nikita Kitaev and Lukasz Kaiser and Anselm Levskaya},
booktitle = {International Conference on Learning Representations},
year = {2020},
url = {https://openreview.net/forum?id=rkgNKkHtvB}
}
@inproceedings{fan2020reducing,
title ={Reducing Transformer Depth on Demand with Structured Dropout},
author ={Angela Fan and Edouard Grave and Armand Joulin},
booktitle ={International Conference on Learning Representations},
year ={2020},
url ={https://openreview.net/forum?id=SylO2yStDr}
}
@misc{lan2019albert,
title = {ALBERT: A Lite BERT for Self-supervised Learning of Language Representations},
author = {Zhenzhong Lan and Mingda Chen and Sebastian Goodman and Kevin Gimpel and Piyush Sharma and Radu Soricut},
year = {2019},
url = {https://arxiv.org/abs/1909.11942}
}
@misc{lample2019large,
title = {Large Memory Layers with Product Keys},
author = {Guillaume Lample and Alexandre Sablayrolles and Marc'Aurelio Ranzato and Ludovic Denoyer and Hervé Jégou},
year = {2019},
eprint = {1907.05242},
archivePrefix = {arXiv}
}
@article{DBLP:journals/corr/abs-1907-01470,
author = {Sainbayar Sukhbaatar and
Edouard Grave and
Guillaume Lample and
Herv{\'{e}} J{\'{e}}gou and
Armand Joulin},
title = {Augmenting Self-attention with Persistent Memory},
journal = {CoRR},
volume = {abs/1907.01470},
year = {2019},
url = {http://arxiv.org/abs/1907.01470}
}
@misc{bhojanapalli2020lowrank,
title = {Low-Rank Bottleneck in Multi-head Attention Models},
author = {Srinadh Bhojanapalli and Chulhee Yun and Ankit Singh Rawat and Sashank J. Reddi and Sanjiv Kumar},
year = {2020},
eprint = {2002.07028}
}
@article{1910.05895,
author = {Toan Q. Nguyen and Julian Salazar},
title = {Transformers without Tears: Improving the Normalization of Self-Attention},
year = {2019},
eprint = {arXiv:1910.05895},
doi = {10.5281/zenodo.3525484},
}
@misc{bachlechner2020rezero,
title = {ReZero is All You Need: Fast Convergence at Large Depth},
author = {Thomas Bachlechner and Bodhisattwa Prasad Majumder and Huanru Henry Mao and Garrison W. Cottrell and Julian McAuley},
year = {2020},
url = {https://arxiv.org/abs/2003.04887}
}
@misc{vaswani2017attention,
title = {Attention Is All You Need},
author = {Ashish Vaswani and Noam Shazeer and Niki Parmar and Jakob Uszkoreit and Llion Jones and Aidan N. Gomez and Lukasz Kaiser and Illia Polosukhin},
year = {2017},
eprint = {1706.03762},
archivePrefix = {arXiv},
primaryClass = {cs.CL}
}
@software{peng_bo_2021_5196578,
author = {PENG Bo},
title = {BlinkDL/RWKV-LM: 0.01},
month = {aug},
year = {2021},
publisher = {Zenodo},
version = {0.01},
doi = {10.5281/zenodo.5196578},
url = {https://doi.org/10.5281/zenodo.5196578}
}