This repository contains the code and pre-trained models for our paper SimCSE: Simple Contrastive Learning of Sentence Embeddings.
**************************** Updates ****************************
- 4/23: We released our training code.
- 4/20: We released our model checkpoints and evaluation code.
- 4/18: We released our paper. Check it out!
- Overview
- Pre-trained sentence embeddings
- Requirements
- Evaluation
- Training
- Bugs or Questions?
- Citation
- SimCSE Elsewhere
We propose a simple contrastive learning framework that works with both unlabeled and labeled data. Unsupervised SimCSE simply takes an input sentence and predicts itself in a contrastive learning framework, with only standard dropout used as noise. Our supervised SimCSE incorporates annotated pairs from NLI datasets into contrastive learning by using entailment
pairs as positives and contradiction
pairs as hard negatives. The following figure is an illustration of our models.
Our pre-trained models are now publicly available with HuggingFace's Transformers. Models and their performance are presented as follows:
Model | Avg. STS |
---|---|
unsup-simcse-bert-base-uncased | 74.54 |
unsup-simcse-bert-large-uncased | 76.05 |
unsup-simcse-roberta-base | 76.50 |
unsup-simcse-roberta-large | 77.47 |
sup-simcse-bert-base-uncased | 81.57 |
sup-simcse-bert-large-uncased | 82.21 |
sup-simcse-roberta-base | 82.52 |
sup-simcse-roberta-large | 83.76 |
Naming rules: unsup
and sup
represent "unsupervised" (trained on Wikipedia corpus) and "supervised" (trained on NLI datasets) respectively.
You can easily import our model in an out-of-the-box way with HuggingFace's API:
import torch
from scipy.spatial.distance import cosine
from transformers import AutoModel, AutoTokenizer
# Import our models. The package will take care of downloading the models automatically
tokenizer = AutoTokenizer.from_pretrained("princeton-nlp/sup-simcse-bert-base-uncased")
model = AutoModel.from_pretrained("princeton-nlp/sup-simcse-bert-base-uncased")
# Tokenize input texts
texts = [
"There's a kid on a skateboard.",
"A kid is skateboarding.",
"A kid is inside the house."
]
inputs = tokenizer(texts, padding=True, truncation=True, return_tensors="pt")
# Get the embeddings
with torch.no_grad():
embeddings = model(**inputs, output_hidden_states=True, return_dict=True).pooler_output
# Calculate cosine similarities
# Cosine similarities are in [-1, 1]. Higher means more similar
cosine_sim_0_1 = 1 - cosine(embeddings[0], embeddings[1])
cosine_sim_0_2 = 1 - cosine(embeddings[0], embeddings[2])
print("Cosine similarity between \"%s\" and \"%s\" is: %.3f" % (texts[0], texts[1], cosine_sim_0_1))
print("Cosine similarity between \"%s\" and \"%s\" is: %.3f" % (texts[0], texts[2], cosine_sim_0_2))
If you encounter any problem when directly loading the models by HuggingFace's API, you can also download the models manually from the above table and use model = AutoModel.from_pretrained({PATH TO THE DOWNLOAD MODEL})
.
If you only want to use our models in an out-of-the-box way, just installing the latest version of torch
, transformers
and scipy
is enough. If you want to use our training or evaluation code, see the requirement section below.
First, install PyTorch by following the instructions from the official website. To faithfully reproduce our results, please use the correct 1.7.1
version corresponding to your platforms/CUDA versions. PyTorch version higher than 1.7.1
should also work. For example, if you use Linux and CUDA11 (how to check CUDA version), install PyTorch by the following command,
pip install torch==1.7.1+cu110 -f https://download.pytorch.org/whl/torch_stable.html
If you instead use CUDA <11
or CPU, install PyTorch by the following command,
pip install torch==1.7.1
Then run the following script to install the remaining dependencies,
pip install -r requirements.txt
Our evaluation code for sentence embeddings is based on a modified version of SentEval. It evaluates sentence embeddings on semantic textual similarity (STS) tasks and downstream transfer tasks. For STS tasks, our evaluation takes the "all" setting, and report Spearman's correlation. See our paper (Appendix B) for evaluation details.
Before evaluation, please download the evaluation datasets by running
cd SentEval/data/downstream/
bash download_dataset.sh
Then come back to the root directory, you can evaluate any transformers
-based pre-trained models using our evaluation code. For example,
python evaluation.py \
--model_name_or_path princeton-nlp/sup-simcse-bert-base-uncased \
--pooler cls \
--task_set sts \
--mode test
which is expected to output the results in a tabular format:
------ test ------
+-------+-------+-------+-------+-------+--------------+-----------------+-------+
| STS12 | STS13 | STS14 | STS15 | STS16 | STSBenchmark | SICKRelatedness | Avg. |
+-------+-------+-------+-------+-------+--------------+-----------------+-------+
| 75.30 | 84.67 | 80.19 | 85.40 | 80.82 | 84.26 | 80.39 | 81.58 |
+-------+-------+-------+-------+-------+--------------+-----------------+-------+
Arguments for the evaluation script are as follows,
--model_name_or_path
: The name or path of atransformers
-based pre-trained checkpoint. You can directly use the models in the above table, e.g.,princeton-nlp/sup-simcse-bert-base-uncased
.--pooler
: Pooling method. Now we supportcls
(default): Use the representation of[CLS]
token. A linear+activation layer is applied after the representation (it's in the standard BERT implementation). If you use SimCSE, you should use this option.cls_before_pooler
: Use the representation of[CLS]
token without the extra linear+activation.avg
: Average embeddings of the last layer. If you use checkpoints of SBERT/SRoBERTa (paper), you should use this option.avg_top2
: Average embeddings of the last two layers.avg_first_last
: Average embeddings of the first and last layers. If you use vanilla BERT or RoBERTa, this works the best.
--mode
: Evaluation modetest
(default): The default test mode. To faithfully reproduce our results, you should use this option.dev
: Report the development set results. Note that in STS tasks, onlySTS-B
andSICK-R
have development sets, so we only report their numbers. It also takes a fast mode for transfer tasks, so the running time is much shorter than thetest
mode (though numbers are slightly lower).fasttest
: It is the same astest
, but with a fast mode so the running time is much shorter, but the reported numbers may be lower (only for transfer tasks).
--task_set
: What set of tasks to evaluate on (if set, it will override--tasks
)sts
(default): Evaluate on STS tasks, includingSTS 12~16
,STS-B
andSICK-R
. This is the most commonly-used set of tasks to evaluate the quality of sentence embeddings.transfer
: Evaluate on transfer tasks.full
: Evaluate on both STS and transfer tasks.na
: Manually set tasks by--tasks
.
--tasks
: Specify which dataset(s) to evaluate on. Will be overridden if--task_set
is notna
. See the code for a full list of tasks.
For unsupervised SimCSE, we sample 1 million sentences from English Wikipedia; for supervised SimCSE, we use the SNLI and MNLI datasets. You can run data/download_wiki.sh
and data/download_nli.sh
to download the two datasets.
We provide example training scripts for both unsupervised and supervised SimCSE. In run_unsup_example.sh
, we provide a single-GPU (or CPU) example for the unsupervised version, and in run_sup_example.sh
we give a multiple-GPU example for the supervised version. Both scripts call train.py
for training. We explain the arguments in following:
--train_file
: Training file path. We support "txt" files (one line for one sentence) and "csv" files (2-column: pair data with no hard negative; 3-column: pair data with one corresponding hard negative instance). You can use our provided Wikipedia or NLI data, or you can use your own data with the same format.--model_name_or_path
: Pre-trained checkpoints to start with. For now we support BERT-based models (bert-base-uncased
,bert-large-uncased
, etc.) and RoBERTa-based models (RoBERTa-base
,RoBERTa-large
, etc.).--temp
: Temperature for the contrastive loss.--pooler_type
: Pooling method. It's the same as the--pooler_type
in the evaluation part.--hard_negative_weight
: If using hard negatives (i.e., there are 3 columns in the training file), this is the logarithm of the weight. For example, if the weight is 1, then this argument should be set as 0 (default value).--do_mlm
: Whether to use the MLM auxiliary objective. If True:--mlm_weight
: Weight for the MLM objective.--mlm_probability
: Masking rate for the MLM objective.
All the other arguments are standard Huggingface's transformers
training arguments. Some of the often-used arguments are: --output_dir
, --learning_rate
, --per_device_train_batch_size
. In our example scripts, we also set to evaluate the model on the STS-B development set (need to download the dataset following the evaluation section) and save the best checkpoint.
REPRODUCTION: For results in the paper, we use Nvidia 3090 GPUs with CUDA 11. Using different types of devices or different versions of CUDA/other softwares may lead to slightly different performance.
IMPORTANT: Our saved checkpoints are slightly different from Huggingface's pre-trained checkpoints. Run python simcse_to_huggingface.py --path {PATH_TO_CHECKPOINT_FOLDER}
to convert it. After that, you can evaluate it by our evaluation code or directly use it out of the box.
If you have any questions related to the code or the paper, feel free to email Tianyu ([email protected]
) and Xingcheng ([email protected]
). If you encounter any problems when using the code, or want to report a bug, you can open an issue. Please try to specify the problem with details so we can help you better and quicker!
Please cite our paper if you use SimCSE in your work:
@article{gao2021simcse,
title={{SimCSE}: Simple Contrastive Learning of Sentence Embeddings},
author={Gao, Tianyu and Yao, Xingcheng and Chen, Danqi},
journal={arXiv preprint arXiv:2104.08821},
year={2021}
}
We thank the community's efforts for extending SimCSE!
- Jianlin Su has provided a Chinese version of SimCSE.
- AK391 has provided a Gradio Web Demo of SimCSE.
- Nils Reimers has implemented a
sentence-transformers
-based training code for SimCSE.