A fully featured Transformer that mixes (QKᵀ)V local attention with Q(KᵀV) global attention (scales linearly with respect to sequence length) for efficient long-range language modeling.
$ pip install linear-attention-transformerLanguage model
import torch
from linear_attention_transformer import LinearAttentionTransformerLM
model = LinearAttentionTransformerLM(
num_tokens = 20000,
dim = 512,
heads = 8,
depth = 1,
max_seq_len = 8192,
causal = True, # auto-regressive or not
ff_dropout = 0.1, # dropout for feedforward
attn_layer_dropout = 0.1, # dropout right after self-attention layer
attn_dropout = 0.1, # dropout post-attention
emb_dim = 128, # embedding factorization, to save on memory
dim_head = 128, # be able to fix the dimension of each head, making it independent of the embedding dimension and the number of heads
blindspot_size = 64, # this gives the q(kv) attention a blindspot of 64 tokens back in the causal case, but gives back an order of magnitude return in memory savings. should be paired with local attention of at least a window size of this setting. setting this to 1 will allow for full q(kv) attention of past
n_local_attn_heads = 4, # number of local attention heads for (qk)v attention. this can be a tuple specifying the exact number of local attention heads at that depth
local_attn_window_size = 128, # receptive field of the local attention
reversible = True, # use reversible nets, from Reformer paper
ff_chunks = 2, # feedforward chunking, from Reformer paper
ff_glu = True, # use GLU variant for feedforward
attend_axially = False, # will fold the sequence by the local attention window size, and do an extra strided attention followed by a feedforward with the cheap q(kv) attention
shift_tokens = True # add single token shifting, for great improved convergence
).cuda()
x = torch.randint(0, 20000, (1, 8192)).cuda()
model(x) # (1, 8192, 512)Transformer
import torch
from linear_attention_transformer import LinearAttentionTransformer
model = LinearAttentionTransformer(
dim = 512,
heads = 8,
depth = 1,
max_seq_len = 8192,
n_local_attn_heads = 4
).cuda()
x = torch.randn(1, 8192, 512).cuda()
model(x) # (1, 8192, 512)Encoder / decoder
import torch
from linear_attention_transformer import LinearAttentionTransformerLM
enc = LinearAttentionTransformerLM(
num_tokens = 20000,
dim = 512,
heads = 8,
depth = 6,
max_seq_len = 4096,
reversible = True,
n_local_attn_heads = 4,
return_embeddings = True
).cuda()
dec = LinearAttentionTransformerLM(
num_tokens = 20000,
dim = 512,
heads = 8,
depth = 6,
causal = True,
max_seq_len = 4096,
reversible = True,
receives_context = True,
n_local_attn_heads = 4
).cuda()
src = torch.randint(0, 20000, (1, 4096)).cuda()
src_mask = torch.ones_like(src).bool().cuda()
tgt = torch.randint(0, 20000, (1, 4096)).cuda()
tgt_mask = torch.ones_like(tgt).bool().cuda()
context = enc(src, input_mask = src_mask)
logits = dec(tgt, context = context, input_mask = tgt_mask, context_mask = src_mask)Linformer is another variant of attention with linear complexity championed by Facebook AI. It only works with non-autoregressive models of a fixed sequence length. If your problem satisfies that criteria, you may choose to try it out.
from linear_attention_transformer import LinearAttentionTransformerLM, LinformerSettings
settings = LinformerSettings(k = 256)
enc = LinearAttentionTransformerLM(
num_tokens = 20000,
dim = 512,
heads = 8,
depth = 6,
max_seq_len = 4096,
linformer_settings = settings
).cuda()You can also used Linformer for the contextual attention layer, if the contextual keys are of a fixed sequence length.
from linear_attention_transformer import LinearAttentionTransformerLM, LinformerContextSettings
settings = LinformerContextSettings(
seq_len = 2048,
k = 256
)
dec = LinearAttentionTransformerLM(
num_tokens = 20000,
dim = 512,
heads = 8,
depth = 6,
max_seq_len = 4096,
causal = True,
context_linformer_settings = settings,
receives_context = True
).cuda()This repository also contains a concise implementation of this efficient attention for images
import torch
from linear_attention_transformer.images import ImageLinearAttention
attn =ImageLinearAttention(
chan = 32,
heads = 8,
key_dim = 64 # can be decreased to 32 for more memory savings
)
img = torch.randn(1, 32, 256, 256)
attn(img) # (1, 32, 256, 256)@inproceedings{katharopoulos-et-al-2020,
author = {Katharopoulos, A. and Vyas, A. and Pappas, N. and Fleuret, F.},
title = {Transformers are RNNs: Fast Autoregressive Transformers with Linear Attention},
booktitle = {Proceedings of the International Conference on Machine Learning (ICML)},
year = {2020},
url = {https://arxiv.org/abs/2006.16236}
}@article{shen2019efficient,
author = {Zhuoran Shen and
Mingyuan Zhang and
Haiyu Zhao and
Shuai Yi and
Hongsheng Li},
title = {Efficient Attention: Attention with Linear Complexities},
journal = {CoRR},
volume = {abs/1812.01243},
year = {2018},
url = {http://arxiv.org/abs/1812.01243}
}@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}
}@misc{shazeer2020glu,
title = {GLU Variants Improve Transformer},
author = {Noam Shazeer},
year = {2020},
url = {https://arxiv.org/abs/2002.05202}
}@misc{wang2020linformer,
title = {Linformer: Self-Attention with Linear Complexity},
author = {Sinong Wang and Belinda Z. Li and Madian Khabsa and Han Fang and Hao Ma},
year = {2020},
eprint = {2006.04768}
}@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}
}@techreport{zhuiyiroformer,
title = {RoFormer: Transformer with Rotary Position Embeddings - ZhuiyiAI},
author = {Jianlin Su},
year = {2021},
url = "https://github.com/ZhuiyiTechnology/roformer",
}