Model | Description | Dataset | Download |
---|---|---|---|
transformer_lm.gbw.adaptive_huge |
Adaptive Inputs (Baevski and Auli, 2018) 1026M params |
Google Billion Words | download (.tar.bz2) |
transformer_lm.wiki103.adaptive |
Adaptive Inputs (Baevski and Auli, 2018) 247M params |
WikiText-103 | download (.tar.bz2) |
transformer_lm.wmt19.en |
English LM (Ng et al., 2019) |
WMT News Crawl | download (.tar.gz) |
transformer_lm.wmt19.de |
German LM (Ng et al., 2019) |
WMT News Crawl | download (.tar.gz) |
transformer_lm.wmt19.ru |
Russian LM (Ng et al., 2019) |
WMT News Crawl | download (.tar.gz) |
We require a few additional Python dependencies for preprocessing:
pip install fastBPE sacremoses
To sample from a language model using PyTorch Hub:
import torch
# List available models
torch.hub.list('pytorch/fairseq') # [..., 'transformer_lm.wmt19.en', ...]
# Load an English LM trained on WMT'19 News Crawl data
en_lm = torch.hub.load('pytorch/fairseq', 'transformer_lm.wmt19.en', tokenizer='moses', bpe='fastbpe')
en_lm.eval() # disable dropout
# Move model to GPU
en_lm.cuda()
# Sample from the language model
en_lm.sample('Barack Obama', beam=1, sampling=True, sampling_topk=10, temperature=0.8)
# "Barack Obama is coming to Sydney and New Zealand (...)"
# Compute perplexity for a sequence
en_lm.score('Barack Obama is coming to Sydney and New Zealand')['positional_scores'].mean().neg().exp()
# tensor(15.1474)
# The same interface can be used with custom models as well
from fairseq.models.transformer_lm import TransformerLanguageModel
custom_lm = TransformerLanguageModel.from_pretrained('/path/to/model/dir', 'checkpoint100.pt', tokenizer='moses', bpe='fastbpe')
custom_lm.sample('Barack Obama', beam=5)
# "Barack Obama (...)"
First download and prepare the WikiText-103 dataset:
cd examples/language_model/
bash prepare-wikitext-103.sh
cd ../..
Next preprocess/binarize the data:
TEXT=examples/language_model/wikitext-103
fairseq-preprocess \
--only-source \
--trainpref $TEXT/wiki.train.tokens \
--validpref $TEXT/wiki.valid.tokens \
--testpref $TEXT/wiki.test.tokens \
--destdir data-bin/wikitext-103 \
--workers 20
Next we'll train a basic transformer language model on wikitext-103. For more advanced usage, see the adaptive inputs README.
To train a basic LM (assumes 2 GPUs):
$ fairseq-train --task language_modeling \
data-bin/wikitext-103 \
--save-dir checkpoints/transformer_wikitext-103 \
--arch transformer_lm --share-decoder-input-output-embed \
--dropout 0.1 \
--optimizer adam --adam-betas '(0.9, 0.98)' --weight-decay 0.01 --clip-norm 0.0 \
--lr 0.0005 --lr-scheduler inverse_sqrt --warmup-updates 4000 --warmup-init-lr 1e-07 \
--tokens-per-sample 512 --sample-break-mode none \
--max-tokens 2048 --update-freq 16 \
--fp16 \
--max-update 50000
If you run out of memory, try reducing --max-tokens
(max number of tokens per
batch) or --tokens-per-sample
(max sequence length). You can also adjust
--update-freq
to accumulate gradients and simulate training on a different
number of GPUs.
fairseq-eval-lm data-bin/wikitext-103 \
--path checkpoints/transformer_wiki103/checkpoint_best.pt \
--max-sentences 2 \
--tokens-per-sample 512 \
--context-window 400
# | Evaluated 245569 tokens in 56.1s (4379.02 tokens/s)
# | Loss: 3.4164, Perplexity: 30.46
Note: The --context-window
option controls how much context is provided to
each token when computing perplexity. When the window size is 0, the dataset is
chunked into segments of length 512 and perplexity is computed over each segment
normally. However, this results in worse (higher) perplexity since tokens that
appear earlier in each segment have less conditioning. When the maximum window
size is used (511 in this case), then we compute perplexity for each token
fully conditioned on 511 tokens of context. This slows down evaluation
significantly, since we must run a separate forward pass for every token in the
dataset, but results in better (lower) perplexity.
Please see the convolutional LM README for instructions on training convolutional language models.