Examples

In this section a few examples are put together. All of these examples work for several models, making use of the very similar API between the different models.

Important To run the latest versions of the examples, you have to install from source and install some specific requirements for the examples. Execute the following steps in a new virtual environment:

git clone https://github.com/huggingface/transformers
cd transformers
pip install .
pip install -r ./examples/requirements.txt

Section	Description
TensorFlow 2.0 models on GLUE	Examples running BERT TensorFlow 2.0 model on the GLUE tasks.
Language Model training	Fine-tuning (or training from scratch) the library models for language modeling on a text dataset. Causal language modeling for GPT/GPT-2, masked language modeling for BERT/RoBERTa.
Language Generation	Conditional text generation using the auto-regressive models of the library: GPT, GPT-2, Transformer-XL and XLNet.
GLUE	Examples running BERT/XLM/XLNet/RoBERTa on the 9 GLUE tasks. Examples feature distributed training as well as half-precision.
SQuAD	Using BERT/RoBERTa/XLNet/XLM for question answering, examples with distributed training.
Multiple Choice	Examples running BERT/XLNet/RoBERTa on the SWAG/RACE/ARC tasks.
Named Entity Recognition	Using BERT for Named Entity Recognition (NER) on the CoNLL 2003 dataset, examples with distributed training.
XNLI	Examples running BERT/XLM on the XNLI benchmark.
Adversarial evaluation of model performances	Testing a model with adversarial evaluation of natural language inference on the Heuristic Analysis for NLI Systems (HANS) dataset (McCoy et al., 2019.)

TensorFlow 2.0 Bert models on GLUE

Based on the script run_tf_glue.py.

Fine-tuning the library TensorFlow 2.0 Bert model for sequence classification on the MRPC task of the GLUE benchmark: General Language Understanding Evaluation.

This script has an option for mixed precision (Automatic Mixed Precision / AMP) to run models on Tensor Cores (NVIDIA Volta/Turing GPUs) and future hardware and an option for XLA, which uses the XLA compiler to reduce model runtime. Options are toggled using USE_XLA or USE_AMP variables in the script. These options and the below benchmark are provided by @tlkh.

Quick benchmarks from the script (no other modifications):

GPU	Mode	Time (2nd epoch)	Val Acc (3 runs)
Titan V	FP32	41s	0.8438/0.8281/0.8333
Titan V	AMP	26s	0.8281/0.8568/0.8411
V100	FP32	35s	0.8646/0.8359/0.8464
V100	AMP	22s	0.8646/0.8385/0.8411
1080 Ti	FP32	55s	-

Mixed precision (AMP) reduces the training time considerably for the same hardware and hyper-parameters (same batch size was used).

Language model training

Based on the script run_language_modeling.py.

Fine-tuning (or training from scratch) the library models for language modeling on a text dataset for GPT, GPT-2, BERT and RoBERTa (DistilBERT to be added soon). GPT and GPT-2 are fine-tuned using a causal language modeling (CLM) loss while BERT and RoBERTa are fine-tuned using a masked language modeling (MLM) loss.

Before running the following example, you should get a file that contains text on which the language model will be trained or fine-tuned. A good example of such text is the WikiText-2 dataset.

We will refer to two different files: $TRAIN_FILE, which contains text for training, and $TEST_FILE, which contains text that will be used for evaluation.

GPT-2/GPT and causal language modeling

The following example fine-tunes GPT-2 on WikiText-2. We’re using the raw WikiText-2 (no tokens were replaced before the tokenization). The loss here is that of causal language modeling.

export TRAIN_FILE=/path/to/dataset/wiki.train.raw
export TEST_FILE=/path/to/dataset/wiki.test.raw

python run_language_modeling.py \
    --output_dir=output \
    --model_type=gpt2 \
    --model_name_or_path=gpt2 \
    --do_train \
    --train_data_file=$TRAIN_FILE \
    --do_eval \
    --eval_data_file=$TEST_FILE

This takes about half an hour to train on a single K80 GPU and about one minute for the evaluation to run. It reaches a score of ~20 perplexity once fine-tuned on the dataset.

RoBERTa/BERT and masked language modeling

The following example fine-tunes RoBERTa on WikiText-2. Here too, we’re using the raw WikiText-2. The loss is different as BERT/RoBERTa have a bidirectional mechanism; we’re therefore using the same loss that was used during their pre-training: masked language modeling.

In accordance to the RoBERTa paper, we use dynamic masking rather than static masking. The model may, therefore, converge slightly slower (over-fitting takes more epochs).

We use the --mlm flag so that the script may change its loss function.

export TRAIN_FILE=/path/to/dataset/wiki.train.raw
export TEST_FILE=/path/to/dataset/wiki.test.raw

python run_language_modeling.py \
    --output_dir=output \
    --model_type=roberta \
    --model_name_or_path=roberta-base \
    --do_train \
    --train_data_file=$TRAIN_FILE \
    --do_eval \
    --eval_data_file=$TEST_FILE \
    --mlm

Language generation

Based on the script run_generation.py.

Conditional text generation using the auto-regressive models of the library: GPT, GPT-2, Transformer-XL, XLNet, CTRL. A similar script is used for our official demo Write With Transfomer, where you can try out the different models available in the library.

Example usage:

python run_generation.py \
    --model_type=gpt2 \
    --model_name_or_path=gpt2

GLUE

Based on the script run_glue.py.

Fine-tuning the library models for sequence classification on the GLUE benchmark: General Language Understanding Evaluation. This script can fine-tune the following models: BERT, XLM, XLNet and RoBERTa.

GLUE is made up of a total of 9 different tasks. We get the following results on the dev set of the benchmark with an uncased BERT base model (the checkpoint bert-base-uncased). All experiments ran single V100 GPUs with a total train batch sizes between 16 and 64. Some of these tasks have a small dataset and training can lead to high variance in the results between different runs. We report the median on 5 runs (with different seeds) for each of the metrics.

Task	Metric	Result
CoLA	Matthew's corr	49.23
SST-2	Accuracy	91.97
MRPC	F1/Accuracy	89.47/85.29
STS-B	Person/Spearman corr.	83.95/83.70
QQP	Accuracy/F1	88.40/84.31
MNLI	Matched acc./Mismatched acc.	80.61/81.08
QNLI	Accuracy	87.46
RTE	Accuracy	61.73
WNLI	Accuracy	45.07

Some of these results are significantly different from the ones reported on the test set of GLUE benchmark on the website. For QQP and WNLI, please refer to FAQ #12 on the webite.

Before running any one of these GLUE tasks you should download the GLUE data by running this script and unpack it to some directory $GLUE_DIR.

export GLUE_DIR=/path/to/glue
export TASK_NAME=MRPC

python run_glue.py \
  --model_type bert \
  --model_name_or_path bert-base-cased \
  --task_name $TASK_NAME \
  --do_train \
  --do_eval \
  --do_lower_case \
  --data_dir $GLUE_DIR/$TASK_NAME \
  --max_seq_length 128 \
  --per_gpu_train_batch_size 32 \
  --learning_rate 2e-5 \
  --num_train_epochs 3.0 \
  --output_dir /tmp/$TASK_NAME/

where task name can be one of CoLA, SST-2, MRPC, STS-B, QQP, MNLI, QNLI, RTE, WNLI.

The dev set results will be present within the text file eval_results.txt in the specified output_dir. In case of MNLI, since there are two separate dev sets (matched and mismatched), there will be a separate output folder called /tmp/MNLI-MM/ in addition to /tmp/MNLI/.

The code has not been tested with half-precision training with apex on any GLUE task apart from MRPC, MNLI, CoLA, SST-2. The following section provides details on how to run half-precision training with MRPC. With that being said, there shouldn’t be any issues in running half-precision training with the remaining GLUE tasks as well, since the data processor for each task inherits from the base class DataProcessor.

MRPC

Fine-tuning example

The following examples fine-tune BERT on the Microsoft Research Paraphrase Corpus (MRPC) corpus and runs in less than 10 minutes on a single K-80 and in 27 seconds (!) on single tesla V100 16GB with apex installed.

Before running any one of these GLUE tasks you should download the GLUE data by running this script and unpack it to some directory $GLUE_DIR.

export GLUE_DIR=/path/to/glue

python run_glue.py \
  --model_type bert \
  --model_name_or_path bert-base-cased \
  --task_name MRPC \
  --do_train \
  --do_eval \
  --do_lower_case \
  --data_dir $GLUE_DIR/MRPC/ \
  --max_seq_length 128 \
  --per_gpu_train_batch_size 32 \
  --learning_rate 2e-5 \
  --num_train_epochs 3.0 \
  --output_dir /tmp/mrpc_output/

Our test ran on a few seeds with the original implementation hyper- parameters gave evaluation results between 84% and 88%.

Using Apex and mixed-precision

Using Apex and 16 bit precision, the fine-tuning on MRPC only takes 27 seconds. First install apex, then run the following example:

export GLUE_DIR=/path/to/glue

python run_glue.py \
  --model_type bert \
  --model_name_or_path bert-base-cased \
  --task_name MRPC \
  --do_train \
  --do_eval \
  --do_lower_case \
  --data_dir $GLUE_DIR/MRPC/ \
  --max_seq_length 128 \
  --per_gpu_train_batch_size 32 \
  --learning_rate 2e-5 \
  --num_train_epochs 3.0 \
  --output_dir /tmp/mrpc_output/ \
  --fp16

Distributed training

Here is an example using distributed training on 8 V100 GPUs. The model used is the BERT whole-word-masking and it reaches F1 > 92 on MRPC.

export GLUE_DIR=/path/to/glue

python -m torch.distributed.launch \
    --nproc_per_node 8 run_glue.py \
    --model_type bert \
    --model_name_or_path bert-base-cased \
    --task_name MRPC \
    --do_train \
    --do_eval \
    --do_lower_case \
    --data_dir $GLUE_DIR/MRPC/ \
    --max_seq_length 128 \
    --per_gpu_train_batch_size 8 \
    --learning_rate 2e-5 \
    --num_train_epochs 3.0 \
    --output_dir /tmp/mrpc_output/

Training with these hyper-parameters gave us the following results:

acc = 0.8823529411764706
acc_and_f1 = 0.901702786377709
eval_loss = 0.3418912578906332
f1 = 0.9210526315789473
global_step = 174
loss = 0.07231863956341798

MNLI

The following example uses the BERT-large, uncased, whole-word-masking model and fine-tunes it on the MNLI task.

export GLUE_DIR=/path/to/glue

python -m torch.distributed.launch \
    --nproc_per_node 8 run_glue.py \
    --model_type bert \
    --model_name_or_path bert-base-cased \
    --task_name mnli \
    --do_train \
    --do_eval \
    --do_lower_case \
    --data_dir $GLUE_DIR/MNLI/ \
    --max_seq_length 128 \
    --per_gpu_train_batch_size 8 \
    --learning_rate 2e-5 \
    --num_train_epochs 3.0 \
    --output_dir output_dir \

The results are the following:

***** Eval results *****
  acc = 0.8679706601466992
  eval_loss = 0.4911287787382479
  global_step = 18408
  loss = 0.04755385363816904

***** Eval results *****
  acc = 0.8747965825874695
  eval_loss = 0.45516540421714036
  global_step = 18408
  loss = 0.04755385363816904

Multiple Choice

Based on the script run_multiple_choice.py.

Fine-tuning on SWAG

Download swag data

#training on 4 tesla V100(16GB) GPUS
export SWAG_DIR=/path/to/swag_data_dir
python ./examples/run_multiple_choice.py \
--model_type roberta \
--task_name swag \
--model_name_or_path roberta-base \
--do_train \
--do_eval \
--do_lower_case \
--data_dir $SWAG_DIR \
--learning_rate 5e-5 \
--num_train_epochs 3 \
--max_seq_length 80 \
--output_dir models_bert/swag_base \
--per_gpu_eval_batch_size=16 \
--per_gpu_train_batch_size=16 \
--gradient_accumulation_steps 2 \
--overwrite_output

Training with the defined hyper-parameters yields the following results:

***** Eval results *****
eval_acc = 0.8338998300509847
eval_loss = 0.44457291918821606

SQuAD

Based on the script run_squad.py.

Fine-tuning BERT on SQuAD1.0

This example code fine-tunes BERT on the SQuAD1.0 dataset. It runs in 24 min (with BERT-base) or 68 min (with BERT-large) on a single tesla V100 16GB. The data for SQuAD can be downloaded with the following links and should be saved in a $SQUAD_DIR directory.

• train-v1.1.json

• dev-v1.1.json

• evaluate-v1.1.py

And for SQuAD2.0, you need to download:

• train-v2.0.json

• dev-v2.0.json

• evaluate-v2.0.py

export SQUAD_DIR=/path/to/SQUAD

python run_squad.py \
  --model_type bert \
  --model_name_or_path bert-base-uncased \
  --do_train \
  --do_eval \
  --do_lower_case \
  --train_file $SQUAD_DIR/train-v1.1.json \
  --predict_file $SQUAD_DIR/dev-v1.1.json \
  --per_gpu_train_batch_size 12 \
  --learning_rate 3e-5 \
  --num_train_epochs 2.0 \
  --max_seq_length 384 \
  --doc_stride 128 \
  --output_dir /tmp/debug_squad/

Training with the previously defined hyper-parameters yields the following results:

f1 = 88.52
exact_match = 81.22

Distributed training

Here is an example using distributed training on 8 V100 GPUs and Bert Whole Word Masking uncased model to reach a F1 > 93 on SQuAD1.1:

python -m torch.distributed.launch --nproc_per_node=8 ./examples/run_squad.py \
    --model_type bert \
    --model_name_or_path bert-large-uncased-whole-word-masking \
    --do_train \
    --do_eval \
    --do_lower_case \
    --train_file $SQUAD_DIR/train-v1.1.json \
    --predict_file $SQUAD_DIR/dev-v1.1.json \
    --learning_rate 3e-5 \
    --num_train_epochs 2 \
    --max_seq_length 384 \
    --doc_stride 128 \
    --output_dir ./examples/models/wwm_uncased_finetuned_squad/ \
    --per_gpu_eval_batch_size=3   \
    --per_gpu_train_batch_size=3   \

Training with the previously defined hyper-parameters yields the following results:

f1 = 93.15
exact_match = 86.91

This fine-tuned model is available as a checkpoint under the reference bert-large-uncased-whole-word-masking-finetuned-squad.

Fine-tuning XLNet on SQuAD

This example code fine-tunes XLNet on both SQuAD1.0 and SQuAD2.0 dataset. See above to download the data for SQuAD .

Command for SQuAD1.0:

export SQUAD_DIR=/path/to/SQUAD

python run_squad.py \
    --model_type xlnet \
    --model_name_or_path xlnet-large-cased \
    --do_train \
    --do_eval \
    --do_lower_case \
    --train_file $SQUAD_DIR/train-v1.1.json \
    --predict_file $SQUAD_DIR/dev-v1.1.json \
    --learning_rate 3e-5 \
    --num_train_epochs 2 \
    --max_seq_length 384 \
    --doc_stride 128 \
    --output_dir ./wwm_cased_finetuned_squad/ \
    --per_gpu_eval_batch_size=4  \
    --per_gpu_train_batch_size=4   \
    --save_steps 5000

Command for SQuAD2.0:

export SQUAD_DIR=/path/to/SQUAD

python run_squad.py \
    --model_type xlnet \
    --model_name_or_path xlnet-large-cased \
    --do_train \
    --do_eval \
    --version_2_with_negative \
    --train_file $SQUAD_DIR/train-v2.0.json \
    --predict_file $SQUAD_DIR/dev-v2.0.json \
    --learning_rate 3e-5 \
    --num_train_epochs 4 \
    --max_seq_length 384 \
    --doc_stride 128 \
    --output_dir ./wwm_cased_finetuned_squad/ \
    --per_gpu_eval_batch_size=2  \
    --per_gpu_train_batch_size=2   \
    --save_steps 5000

Larger batch size may improve the performance while costing more memory.

Results for SQuAD1.0 with the previously defined hyper-parameters:

{
"exact": 85.45884578997162,
"f1": 92.5974600601065,
"total": 10570,
"HasAns_exact": 85.45884578997162,
"HasAns_f1": 92.59746006010651,
"HasAns_total": 10570
}

Results for SQuAD2.0 with the previously defined hyper-parameters:

{
"exact": 80.4177545691906,
"f1": 84.07154997729623,
"total": 11873,
"HasAns_exact": 76.73751686909581,
"HasAns_f1": 84.05558584352873,
"HasAns_total": 5928,
"NoAns_exact": 84.0874684608915,
"NoAns_f1": 84.0874684608915,
"NoAns_total": 5945
}

XNLI

Based on the script run_xnli.py.

XNLI is crowd-sourced dataset based on MultiNLI. It is an evaluation benchmark for cross-lingual text representations. Pairs of text are labeled with textual entailment annotations for 15 different languages (including both high-resource language such as English and low-resource languages such as Swahili).

Fine-tuning on XNLI

This example code fine-tunes mBERT (multi-lingual BERT) on the XNLI dataset. It runs in 106 mins on a single tesla V100 16GB. The data for XNLI can be downloaded with the following links and should be both saved (and un-zipped) in a $XNLI_DIR directory.

• XNLI 1.0

• XNLI-MT 1.0

export XNLI_DIR=/path/to/XNLI

python run_xnli.py \
  --model_type bert \
  --model_name_or_path bert-base-multilingual-cased \
  --language de \
  --train_language en \
  --do_train \
  --do_eval \
  --data_dir $XNLI_DIR \
  --per_gpu_train_batch_size 32 \
  --learning_rate 5e-5 \
  --num_train_epochs 2.0 \
  --max_seq_length 128 \
  --output_dir /tmp/debug_xnli/ \
  --save_steps -1

Training with the previously defined hyper-parameters yields the following results on the test set:

acc = 0.7093812375249501

MM-IMDb

Based on the script run_mmimdb.py.

MM-IMDb is a Multimodal dataset with around 26,000 movies including images, plots and other metadata.

Training on MM-IMDb

python run_mmimdb.py \
    --data_dir /path/to/mmimdb/dataset/ \
    --model_type bert \
    --model_name_or_path bert-base-uncased \
    --output_dir /path/to/save/dir/ \
    --do_train \
    --do_eval \
    --max_seq_len 512 \
    --gradient_accumulation_steps 20 \
    --num_image_embeds 3 \
    --num_train_epochs 100 \
    --patience 5

Adversarial evaluation of model performances

Here is an example on evaluating a model using adversarial evaluation of natural language inference with the Heuristic Analysis for NLI Systems (HANS) dataset McCoy et al., 2019. The example was gracefully provided by Nafise Sadat Moosavi.

The HANS dataset can be downloaded from this location.

This is an example of using test_hans.py:

export HANS_DIR=path-to-hans
export MODEL_TYPE=type-of-the-model-e.g.-bert-roberta-xlnet-etc
export MODEL_PATH=path-to-the-model-directory-that-is-trained-on-NLI-e.g.-by-using-run_glue.py

python examples/hans/test_hans.py \
        --task_name hans \
        --model_type $MODEL_TYPE \
        --do_eval \
        --do_lower_case \
        --data_dir $HANS_DIR \
        --model_name_or_path $MODEL_PATH \
        --max_seq_length 128 \
        --output_dir $MODEL_PATH \

This will create the hans_predictions.txt file in MODEL_PATH, which can then be evaluated using hans/evaluate_heur_output.py from the HANS dataset.

The results of the BERT-base model that is trained on MNLI using batch size 8 and the random seed 42 on the HANS dataset is as follows:

Heuristic entailed results:
lexical_overlap: 0.9702
subsequence: 0.9942
constituent: 0.9962

Heuristic non-entailed results:
lexical_overlap: 0.199
subsequence: 0.0396
constituent: 0.118