Text classification examples

GLUE tasks

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 any of the models on the hub and can also be used for a dataset hosted on our hub or your own data in a csv or a JSON file (the script might need some tweaks in that case, refer to the comments inside for help).

GLUE is made up of a total of 9 different tasks. Here is how to run the script on one of them:

export TASK_NAME=mrpc

python run_glue.py \
  --model_name_or_path bert-base-cased \
  --task_name $TASK_NAME \
  --do_train \
  --do_eval \
  --max_seq_length 128 \
  --per_device_train_batch_size 32 \
  --learning_rate 2e-5 \
  --num_train_epochs 3 \
  --output_dir /tmp/$TASK_NAME/

where task name can be one of cola, sst2, mrpc, stsb, qqp, mnli, qnli, rte, wnli.

We get the following results on the dev set of the benchmark with the previous commands (with an exception for MRPC and WNLI which are tiny and where we used 5 epochs instead of 3). Trainings are seeded so you should obtain the same results with PyTorch 1.6.0 (and close results with different versions), training times are given for information (a single Titan RTX was used):

Task	Metric	Result	Training time
CoLA	Matthews corr	56.53	3:17
SST-2	Accuracy	92.32	26:06
MRPC	F1/Accuracy	88.85/84.07	2:21
STS-B	Pearson/Spearman corr.	88.64/88.48	2:13
QQP	Accuracy/F1	90.71/87.49	2:22:26
MNLI	Matched acc./Mismatched acc.	83.91/84.10	2:35:23
QNLI	Accuracy	90.66	40:57
RTE	Accuracy	65.70	57
WNLI	Accuracy	56.34	24

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 website.

The following example fine-tunes BERT on the imdb dataset hosted on our hub:

python run_glue.py \
  --model_name_or_path bert-base-cased \
  --dataset_name imdb  \
  --do_train \
  --do_predict \
  --max_seq_length 128 \
  --per_device_train_batch_size 32 \
  --learning_rate 2e-5 \
  --num_train_epochs 3 \
  --output_dir /tmp/imdb/

If your model classification head dimensions do not fit the number of labels in the dataset, you can specify --ignore_mismatched_sizes to adapt it.

Text classification

As an alternative, we can use the script run_classification.py to fine-tune models on a single/multi-label classification task.

The following example fine-tunes BERT on the en subset of amazon_reviews_multi dataset. We can specify the metric, the label column and aso choose which text columns to use jointly for classification.

dataset="amazon_reviews_multi"
subset="en"
python run_classification.py \
    --model_name_or_path  bert-base-uncased \
    --dataset_name ${dataset} \
    --dataset_config_name ${subset} \
    --shuffle_train_dataset \
    --metric_name accuracy \
    --text_column_name "review_title,review_body,product_category" \
    --text_column_delimiter "\n" \
    --label_column_name stars \
    --do_train \
    --do_eval \
    --max_seq_length 512 \
    --per_device_train_batch_size 32 \
    --learning_rate 2e-5 \
    --num_train_epochs 1 \
    --output_dir /tmp/${dataset}_${subset}/

Training for 1 epoch results in acc of around 0.5958 for review_body only and 0.659 for title+body+category.

The following is a multi-label classification example. It fine-tunes BERT on the reuters21578 dataset hosted on our hub:

dataset="reuters21578"
subset="ModApte"
python run_classification.py \
    --model_name_or_path bert-base-uncased \
    --dataset_name ${dataset} \
    --dataset_config_name ${subset} \
    --shuffle_train_dataset \
    --remove_splits "unused" \
    --metric_name f1 \
    --text_column_name text \
    --label_column_name topics \
    --do_train \
    --do_eval \
    --max_seq_length 512 \
    --per_device_train_batch_size 32 \
    --learning_rate 2e-5 \
    --num_train_epochs 15 \
    --output_dir /tmp/${dataset}_${subset}/

It results in a Micro F1 score of around 0.82 without any text and label filtering. Note that you have to explictly remove the "unused" split from the dataset, since it is not used for classification.

Mixed precision training

If you have a GPU with mixed precision capabilities (architecture Pascal or more recent), you can use mixed precision training with PyTorch 1.6.0 or latest, or by installing the Apex library for previous versions. Just add the flag --fp16 to your command launching one of the scripts mentioned above!

Using mixed precision training usually results in 2x-speedup for training with the same final results:

Task	Metric	Result	Training time	Result (FP16)	Training time (FP16)
CoLA	Matthews corr	56.53	3:17	56.78	1:41
SST-2	Accuracy	92.32	26:06	91.74	13:11
MRPC	F1/Accuracy	88.85/84.07	2:21	88.12/83.58	1:10
STS-B	Pearson/Spearman corr.	88.64/88.48	2:13	88.71/88.55	1:08
QQP	Accuracy/F1	90.71/87.49	2:22:26	90.67/87.43	1:11:54
MNLI	Matched acc./Mismatched acc.	83.91/84.10	2:35:23	84.04/84.06	1:17:06
QNLI	Accuracy	90.66	40:57	90.96	20:16
RTE	Accuracy	65.70	57	65.34	29
WNLI	Accuracy	56.34	24	56.34	12

PyTorch version, no Trainer

Based on the script run_glue_no_trainer.py.

Like run_glue.py, this script allows you to fine-tune any of the models on the hub on a text classification task, either a GLUE task or your own data in a csv or a JSON file. The main difference is that this script exposes the bare training loop, to allow you to quickly experiment and add any customization you would like.

It offers less options than the script with Trainer (for instance you can easily change the options for the optimizer or the dataloaders directly in the script) but still run in a distributed setup, on TPU and supports mixed precision by the mean of the 🤗 Accelerate library. You can use the script normally after installing it:

pip install git+https://github.com/huggingface/accelerate

then

export TASK_NAME=mrpc

python run_glue_no_trainer.py \
  --model_name_or_path bert-base-cased \
  --task_name $TASK_NAME \
  --max_length 128 \
  --per_device_train_batch_size 32 \
  --learning_rate 2e-5 \
  --num_train_epochs 3 \
  --output_dir /tmp/$TASK_NAME/

You can then use your usual launchers to run in it in a distributed environment, but the easiest way is to run

accelerate config

and reply to the questions asked. Then

accelerate test

that will check everything is ready for training. Finally, you can launch training with

export TASK_NAME=mrpc

accelerate launch run_glue_no_trainer.py \
  --model_name_or_path bert-base-cased \
  --task_name $TASK_NAME \
  --max_length 128 \
  --per_device_train_batch_size 32 \
  --learning_rate 2e-5 \
  --num_train_epochs 3 \
  --output_dir /tmp/$TASK_NAME/

This command is the same and will work for:

a CPU-only setup
a setup with one GPU
a distributed training with several GPUs (single or multi node)
a training on TPUs

Note that this library is in alpha release so your feedback is more than welcome if you encounter any problem using it.

XNLI

Based on the script run_xnli.py.

XNLI is a 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.

python run_xnli.py \
  --model_name_or_path bert-base-multilingual-cased \
  --language de \
  --train_language en \
  --do_train \
  --do_eval \
  --per_device_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