This repository contains the code and the final checkpoint of the Shortformer model. This file explains how to run our experiments on the WikiText-103 dataset. Read the full ACL 2021 paper here.
The Shortformer is a combination of two methods:
- Staged Training: We first train the model on short input subsequences and then train it on longer ones. This improves both train speed and evaluation perplexity.
- Position-Infused Attention (PIA) + Caching: We cache previously computed subsequence representations and attend to them using Position-Infused Attention. Position-Infused Attention modifies the model so that position embeddings are not added to the word embeddings at the bottom of the network, but instead, they are added to the keys and queries in the attention sublayer (but not to the values). We show that PIA + caching vastly speeds up generation and also improves perplexity.
To use PIA in your transformer model, simply call the attention function as such:
self-attn(keys=x+position_embeddings,
queries=x+position_embeddings,
values=x)
Instead of how we usually call the attention function in the vanilla transformer:
self-attn(keys=x,
queries=x,
values=x)
Staged training requires no modification to the original code. To see how we implemented the Position-Infused Attention and caching, click here. Implementing PIA and caching is very easy, and we've provided detailed comments in the code to explain what how we did it.
If you use this code or results from our paper, please cite:
@inproceedings{shortformer,
title = "Shortformer: Better Language Modeling using Shorter Inputs",
author = "Press, Ofir and
Smith, Noah A. and
Lewis, Mike",
booktitle = "Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers)",
month = aug,
year = "2021",
address = "Online",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2021.acl-long.427",
pages = "5493--5505",
}
This repository is a fork of the Fairseq repository and so has the same requirements.
Once you've installed the dependencies, you can install this repository by running:
pip install --editable .To download and preprocess the data, run:
cd examples/language_model/
bash prepare-wikitext-103.sh
cd ../..
TEXT=examples/language_model/wikitext-103
python preprocess.py \
--only-source \
--trainpref $TEXT/wiki.train.tokens \
--validpref $TEXT/wiki.valid.tokens \
--testpref $TEXT/wiki.test.tokens \
--destdir data-bin/wikitext-103 \
--workers 20Our Shortformer model takes the baseline and adds caching, Position-Infused Attention, and Staged Training.
To train the first stage:
python train.py --task language_modeling data-bin/wikitext-103 --save-dir checkpoints128e100/ --arch transformer_lm_wiki103 --max-update 140100 --max-lr 1.0 --t-mult 2 --lr-period-updates 270000 --lr-scheduler cosine --lr-shrink 0.75 --warmup-updates 16000 --warmup-init-lr 1e-07 --min-lr 1e-09 --optimizer nag --lr 0.0001 --clip-norm 0.1 --criterion adaptive_loss --max-tokens 9216 --update-freq 1 --seed 1 --fp16 --sample-break-mode none --skip-invalid-size-inputs-valid-test --ddp-backend=no_c10d --tokens-per-sample 128 --max-tokens-valid 128 --tokens-from-prev 128 --curriculum 1000 --required-batch-size-multiple 1 --save-interval 100If your GPUs don't have enough memory to execute that command, you can set --update-freq to 2 and --max-tokens to 4608, or set --update-freq to 3 and --max-tokens to 3072 for running the model with even lower memory constraints. This chunks the batch into 2 or 3 different parts and computes each part seperately (instead of in parallel), so it uses less memory but runs slower.
After that, to train the model with the second (and final) stage:
python train.py --task language_modeling data-bin/wikitext-103 --save-dir shortformer/ --restore-file checkpoints128e100/checkpoint100.pt --arch transformer_lm_wiki103 --max-update 286000 --max-lr 1.0 --t-mult 2 --lr-period-updates 270000 --lr-scheduler cosine --lr-shrink 0.75 --warmup-updates 16000 --warmup-init-lr 1e-07 --min-lr 1e-09 --optimizer nag --lr 0.0001 --clip-norm 0.1 --criterion adaptive_loss --max-tokens 9216 --update-freq 1 --seed 1 --fp16 --sample-break-mode none --skip-invalid-size-inputs-valid-test --ddp-backend=no_c10d --tokens-per-sample 512 --max-tokens-valid 512 --tokens-from-prev 512 --curriculum 1000 --required-batch-size-multiple 1 --no-epoch-checkpointsAgain, you can use the update-freq/max-tokens method from above if you run out of memory.
If you'd like to download the Shortformer instead of training it, it is available here.
Rename that file to checkpoint_best.pt if you'd like to follow the directions below.
For nonoverlapping evaluation of the validation set, run:
fairseq-eval-lm data-bin/wikitext-103 --path shortformer/checkpoint_best.pt --sample-break-mode none --gen-subset valid --max-sentences 1For token-by-token generation of the validation set, run:
fairseq-eval-lm data-bin/wikitext-103 --path shortformer/checkpoint_best.pt --sample-break-mode none --gen-subset valid --max-sentences 1 --sliding-inf 1 --context-window 511 --max-tokens 512(Note that --context-window is a fairseq command and doesn't have the exact meaning that the term "context window" has in our paper.)
Staged training improves the perplexity of the model and makes training faster, so there's no reason not to use it, but if you would like to train the Shortformer without it, the command is
python train.py --task language_modeling data-bin/wikitext-103 --save-dir shortformer-no-st/ --arch transformer_lm_wiki103 --max-update 286000 --max-lr 1.0 --t-mult 2 --lr-period-updates 270000 --lr-scheduler cosine --lr-shrink 0.75 --warmup-updates 16000 --warmup-init-lr 1e-07 --min-lr 1e-09 --optimizer nag --lr 0.0001 --clip-norm 0.1 --criterion adaptive_loss --max-tokens 9216 --update-freq 1 --seed 1 --fp16 --sample-break-mode none --skip-invalid-size-inputs-valid-test --ddp-backend=no_c10d --tokens-per-sample 512 --max-tokens-valid 512 --tokens-from-prev 512 --curriculum 1000 --required-batch-size-multiple 1 --no-epoch-checkpointsFor inference, use the same commands as the ones for the Shortformer (above).
Our Shortformer model is fast to train and for token-by-token generation, but if speed is not an issue, we can achieve slightly better performance by just applying Staged Training to the Baevski & Auli baseline LM. This model is very slow but achieves the best perplexity.
To train the first stage, download the unmodified fairseq reporsitory and then run:
python train.py --task language_modeling data-bin/wikitext-103 --save-dir checkpoints-st-128e50/ --arch transformer_lm_wiki103 --max-update 70050 --max-lr 1.0 --t-mult 2 --lr-period-updates 270000 --lr-scheduler cosine --lr-shrink 0.75 --warmup-updates 16000 --warmup-init-lr 1e-07 --min-lr 1e-09 --optimizer nag --lr 0.0001 --clip-norm 0.1 --criterion adaptive_loss --max-tokens 9216 --update-freq 1 --seed 1 --fp16 --sample-break-mode none --skip-invalid-size-inputs-valid-test --ddp-backend=no_c10d --tokens-per-sample 128 --required-batch-size-multiple 1 --save-interval 50We used fairseq commit 23d8502bdde88a3e58e0910e2ee49834f8478b39, so you should also use that, since in future versions the names of some of the arguments have changed.
After that, to train the model with the second (and final) stage:
python train.py --task language_modeling data-bin/wikitext-103 --save-dir st/ --restore-file checkpoints-st-128e50/checkpoint50.pt --arch transformer_lm_wiki103 --max-update 286000 --max-lr 1.0 --t-mult 2 --lr-period-updates 270000 --lr-scheduler cosine --lr-shrink 0.75 --warmup-updates 16000 --warmup-init-lr 1e-07 --min-lr 1e-09 --optimizer nag --lr 0.0001 --clip-norm 0.1 --criterion adaptive_loss --max-tokens 3072 --update-freq 3 --seed 1 --fp16 --sample-break-mode none --skip-invalid-size-inputs-valid-test --ddp-backend=no_c10d --tokens-per-sample 3072 --no-epoch-checkpointsFor nonoverlapping evaluation of the validation set, run:
fairseq-eval-lm data-bin/wikitext-103 --path st/checkpoint_best.pt --sample-break-mode none --gen-subset valid --max-sentences 1For sliding window evaluation of the validation set, with a stride of 2,560, run:
fairseq-eval-lm data-bin/wikitext-103 --path st/checkpoint_best.pt --sample-break-mode none --gen-subset valid --max-sentences 1 --context-window 2560To train the baseline, download the unmodified fairseq repository and then run:
python train.py --task language_modeling data-bin/wikitext-103 --save-dir baseline/ --arch transformer_lm_wiki103 --max-update 286000 --max-lr 1.0 --t-mult 2 --lr-period-updates 270000 --lr-scheduler cosine --lr-shrink 0.75 --warmup-updates 16000 --warmup-init-lr 1e-07 --min-lr 1e-09 --optimizer nag --lr 0.0001 --clip-norm 0.1 --criterion adaptive_loss --max-tokens 3072 --update-freq 3 --seed 1 --fp16 --sample-break-mode none --skip-invalid-size-inputs-valid-test --ddp-backend=no_c10d --tokens-per-sample 3072 --no-epoch-checkpointsAgain, we used fairseq commit 23d8502bdde88a3e58e0910e2ee49834f8478b39, so you should also use that, since in future versions the names of some of the arguments have changed.
Use the same commands as in the 'Baseline with Staged Training' inference subsection.