Feedback Transformer - Pytorch
Simple implementation of Feedback Transformer in Pytorch. They improve on Transformer-XL by having each token have access to the representations of all previous layers through time. This is achieved by aggregating the outputs of all layers into a shared memory, which each token across layers can attend to at each time step.
The main drawback is longer training time, due to its non-parallel nature. But I thought I'd build it to further exploration and research into this line of work.
I also took the liberty to add some various enhancements, including pre-normalization, GLU gated feedforwards, as well as simplified T5 relative positional embeddings.
Install
$ pip install feedback-transformer-pytorch
Usage
import torch
from feedback_transformer_pytorch import FeedbackTransformer
model = FeedbackTransformer(
num_tokens = 20000, # number of tokens
dim = 512, # dimension
depth = 6, # depth
seq_len = 2, # the sequence length of each segment or window
mem_len = 256, # length of the memory buffer
dim_head = 64, # dimension of each head
heads = 8, # number of heads
attn_dropout = 0.1, # attention dropout
ff_dropout = 0.1 # feedforward dropout
).cuda()
x = torch.randint(0, 20000, (2, 64)).cuda()
model(x) # (2, 64, 20000)
If you would like to have fine control over the memory (when to detach, etc), you can do it with some extra keyword arguments on .forward
import torch
from feedback_transformer_pytorch import FeedbackTransformer
model = FeedbackTransformer(
num_tokens = 20000,
dim = 512,
depth = 6,
seq_len = 32,
mem_len = 256
).cuda()
x1 = torch.randint(0, 20000, (2, 32)).cuda()
x2 = torch.randint(0, 20000, (2, 32)).cuda()
x3 = torch.randint(0, 20000, (2, 32)).cuda()
out1, mem1 = model(x1, return_memory = True)
out2, mem2 = model(x2, memory = mem1, return_memory = True)
out3, mem3 = model(x3, memory = mem2, return_memory = True) # (2, 32, 20000)
Citations
@misc{fan2021addressing,
title = {Addressing Some Limitations of Transformers with Feedback Memory},
author = {Angela Fan and Thibaut Lavril and Edouard Grave and Armand Joulin and Sainbayar Sukhbaatar},
year = {2021},
eprint = {2002.09402},
archivePrefix = {arXiv},
primaryClass = {cs.LG}
}
12 Oct 18, 2022
82 Oct 13, 2022
184 Dec 18, 2022
573 Jan 04, 2023
0 Feb 25, 2022
37 Dec 27, 2022
3.2k Jan 08, 2023
313 Oct 18, 2022
96 Nov 09, 2022
3.3k Jan 05, 2023
6 Jan 25, 2022
342 Jan 03, 2023
51 Dec 27, 2022
8 Oct 26, 2022
22.9k Jan 09, 2023
139 Jan 01, 2023
6 Nov 25, 2022
59 Dec 09, 2022
4 Aug 01, 2022
0 Aug 10, 2021