xlnet

Introduction

XLNet is a new unsupervised language representation learning method based on a novel generalized permutation language modeling objective. Additionally, XLNet employs Transformer-XL as the backbone model, exhibiting excellent performance for language tasks involving long context. Overall, XLNet achieves state-of-the-art (SOTA) results on various downstream language tasks including question answering, natural language inference, sentiment analysis, and document ranking.

For a detailed description of technical details and experimental results, please refer to our paper:

​ XLNet: Generalized Autoregressive Pretraining for Language Understanding

​ Zhilin Yang*, Zihang Dai*, Yiming Yang, Jaime Carbonell, Ruslan Salakhutdinov, Quoc V. Le

​ (*: equal contribution)

​ Preprint 2019

Release Notes

• July 16, 2019: XLNet-Base.
• June 19, 2019: initial release with XLNet-Large and code.

Results

As of June 19, 2019, XLNet outperforms BERT on 20 tasks and achieves state-of-the-art results on 18 tasks. Below are some comparison between XLNet-Large and BERT-Large, which have similar model sizes:

Results on Reading Comprehension

Model RACE accuracy SQuAD1.1 EM SQuAD2.0 EM BERT-Large 72.0 84.1 78.98 XLNet-Base 80.18 XLNet-Large 81.75 88.95 86.12

We use SQuAD dev results in the table to exclude other factors such as using additional training data or other data augmentation techniques. See SQuAD leaderboard for test numbers.

Results on Text Classification

Model IMDB Yelp-2 Yelp-5 DBpedia Amazon-2 Amazon-5 BERT-Large 4.51 1.89 29.32 0.64 2.63 34.17 XLNet-Large 3.79 1.55 27.80 0.62 2.40 32.26

The above numbers are error rates.

Results on GLUE

Model MNLI QNLI QQP RTE SST-2 MRPC CoLA STS-B BERT-Large 86.6 92.3 91.3 70.4 93.2 88.0 60.6 90.0 XLNet-Base 86.8 91.7 91.4 74.0 94.7 88.2 60.2 89.5 XLNet-Large 89.8 93.9 91.8 83.8 95.6 89.2 63.6 91.8

We use single-task dev results in the table to exclude other factors such as multi-task learning or using ensembles.

Pre-trained models

Released Models

As of July 16, 2019, the following models have been made available:

• XLNet-Large, Cased: 24-layer, 1024-hidden, 16-heads
• XLNet-Base, Cased: 12-layer, 768-hidden, 12-heads. This model is trained on full data (different from the one in the paper).

We only release cased models for now because on the tasks we consider, we found: (1) for the base setting, cased and uncased models have similar performance; (2) for the large setting, cased models are a bit better in some tasks.

Each .zip file contains three items:

• A TensorFlow checkpoint (xlnet_model.ckpt) containing the pre-trained weights (which is actually 3 files).
• A Sentence Piece model (spiece.model) used for (de)tokenization.
• A config file (xlnet_config.json) which specifies the hyperparameters of the model.

Future Release Plan

We also plan to continuously release more pretrained models under different settings, including:

• A pretrained model that is finetuned on Wikipedia. This can be used for tasks with Wikipedia text such as SQuAD and HotpotQA.
• Pretrained models with other hyperparameter configurations, targeting specific downstream tasks.
• Pretrained models that benefit from new techniques.

Subscribing to XLNet on Google Groups

To receive notifications about updates, announcements and new releases, we recommend subscribing to the XLNet on Google Groups.

Fine-tuning with XLNet

As of June 19, 2019, this code base has been tested with TensorFlow 1.13.1 under Python2.

Memory Issue during Finetuning

• Most of the SOTA results in our paper were produced on TPUs, which generally have more RAM than common GPUs. As a result, it is currently very difficult (costly) to re-produce most of the XLNet-Large SOTA results in the paper using GPUs with 12GB - 16GB of RAM, because a 16GB GPU is only able to hold a single sequence with length 512 for XLNet-Large. Therefore, a large number (ranging from 32 to 128, equal to batch_size) of GPUs are required to reproduce many results in the paper.
• We are experimenting with gradient accumulation to potentially relieve the memory burden, which could be included in a near-future update.
• Alternative methods of finetuning XLNet on constrained hardware have been presented in renatoviolin's repo, which obtained 86.24 F1 on SQuAD2.0 with a 8GB memory GPU.

Given the memory issue mentioned above, using the default finetuning scripts (run_classifier.py and run_squad.py), we benchmarked the maximum batch size on a single 16GB GPU with TensorFlow 1.13.1:

System Seq Length Max Batch Size XLNet-Base 64 120 ... 128 56 ... 256 24 ... 512 8 XLNet-Large 64 16 ... 128 8 ... 256 2 ... 512 1

In most cases, it is possible to reduce the batch size train_batch_size or the maximum sequence length max_seq_length to fit in given hardware. The decrease in performance depends on the task and the available resources.

Text Classification/Regression

The code used to perform classification/regression finetuning is in run_classifier.py. It also contains examples for standard one-document classification, one-document regression, and document pair classification. Here, we provide two concrete examples of how run_classifier.py can be used.

From here on, we assume XLNet-Large and XLNet-base has been downloaded to $LARGE_DIR and $BASE_DIR respectively.

(1) STS-B: sentence pair relevance regression (with GPUs)

• Download the GLUE data by running this script and unpack it to some directory $GLUE_DIR.

• Perform multi-GPU (4 V100 GPUs) finetuning with XLNet-Large by running

…

• Evaluate the finetuning results with a single GPU by

  CUDA_VISIBLE_DEVICES=0 python run_classifier.py \
    --do_train=False \
    --do_eval=True \
    --task_name=sts-b \
    --data_dir=${GLUE_DIR}/STS-B \
    --output_dir=proc_data/sts-b \
    --model_dir=exp/sts-b \
    --uncased=False \
    --spiece_model_file=${LARGE_DIR}/spiece.model \
    --model_config_path=${LARGE_DIR}/xlnet_config.json \
    --max_seq_length=128 \
    --eval_batch_size=8 \
    --num_hosts=1 \
    --num_core_per_host=1 \
    --eval_all_ckpt=True \
    --is_regression=True
  
  # Expected performance: "eval_pearsonr 0.916+ "
  

Notes:

• In the context of GPU training, num_core_per_host denotes the number of GPUs to use.
• In the multi-GPU setting, train_batch_size refers to the per-GPU batch size.
• eval_all_ckpt allows one to evaluate all saved checkpoints (save frequency is controlled by save_steps) after training finishes and choose the best model based on dev performance.
• data_dir and output_dir refer to the directories of the "raw data" and "preprocessed tfrecords" respectively, while model_dir is the working directory for saving checkpoints and tensorflow events. model_dir should be set as a separate folder to init_checkpoint.
• To try out XLNet-base, one can simply set --train_batch_size=32 and --num_core_per_host=1, along with according changes in init_checkpoint and model_config_path.
• For GPUs with smaller RAM, please proportionally decrease the train_batch_size and increase num_core_per_host to use the same training setting.
• Important: we separate the training and evaluation into "two phases", as using multi GPUs to perform evaluation is tricky (one has to correctly separate the data across GPUs). To ensure correctness, we only support single-GPU evaluation for now.

(2) IMDB: movie review sentiment classification (with TPU V3-8)

• Download and unpack the IMDB dataset by running

  wget http://ai.stanford.edu/~amaas/data/sentiment/aclImdb_v1.tar.gz
  tar zxvf aclImdb_v1.tar.gz
  

• Launch a Google cloud TPU V3-8 instance (see the Google Cloud TPU tutorial for how to set up Cloud TPUs).

• Set up your Google storage bucket path $GS_ROOT and move the IMDB dataset and pretrained checkpoint into your Google storage.

• Perform TPU finetuning with XLNet-Large by running

…

Notes:

• To obtain the SOTA on the IMDB dataset, using sequence length 512 is necessary. Therefore, we show how this can be done with a TPU V3-8.
• Alternatively, one can use a sequence length smaller than 512, a smaller batch size, or switch to XLNet-base to train on GPUs. But performance drop is expected.
• Notice that the data_dir and spiece_model_file both use a local path rather than a Google Storage path. The reason is that data preprocessing is actually performed locally. Hence, using local paths leads to a faster preprocessing speed.

SQuAD2.0

The code for the SQuAD dataset is included in run_squad.py.

To run the code:

(1) Download the SQuAD2.0 dataset into $SQUAD_DIR by:

mkdir -p ${SQUAD_DIR} && cd ${SQUAD_DIR}
wget https://rajpurkar.github.io/SQuAD-explorer/dataset/train-v2.0.json
wget https://rajpurkar.github.io/SQuAD-explorer/dataset/dev-v2.0.json

(2) Perform data preprocessing using the script scripts/prepro_squad.sh.

• This will take quite some time in order to accurately map character positions (raw data) to sentence piece positions (used for training).

• For faster parallel preprocessing, please refer to the flags --num_proc and --proc_id in run_squad.py.

(3) Perform training and evaluation.

For the best performance, XLNet-Large uses sequence length 512 and batch size 48 for training.

• As a result, reproducing the best result with GPUs is quite difficult.

• For training with one TPU v3-8, one can simply run the script scripts/tpu_squad_large.sh after both the TPU and Google storage have been setup.

• run_squad.py will automatically perform threshold searching on the dev set of squad and output the score. With scripts/tpu_squad_large.sh, the expected F1 score should be around 88.6 (median of our multiple runs).

Alternatively, one can use XLNet-Base with GPUs (e.g. three V100). One set of reasonable hyper-parameters can be found in the script scripts/gpu_squad_base.sh.

RACE reading comprehension

The code for the reading comprehension task RACE is included in run_race.py.

• Notably, the average length of the passages in RACE is over 300 tokens (not peices), which is significantly longer than other popular reading comprehension datasets such as SQuAD.
• Also, many questions can be very difficult and requires complex reasoning for machines to solve (see one example here).

To run the code:

(1) Download the RACE dataset from the [official website](https: