Skip to content

v2.5.0 - 2D Matryoshka & Adaptive Layer models, CrossEncoder (re)ranking
Compare
Choose a tag to compare
@tomaarsen tomaarsen released this 29 Feb 14:07
· 259 commits to master since this release
v2.5.0
c4b32c2

This release brings two new loss functions, a new way to (re)rank with CrossEncoder models, and more fixes

Install this version with

pip install sentence-transformers==2.5.0

2D Matryoshka & Adaptive Layer models (#2506)

Embedding models are often encoder models with numerous layers, such as 12 (e.g. all-mpnet-base-v2) or 6 (e.g. all-MiniLM-L6-v2). To get embeddings, every single one of these layers must be traversed. 2D Matryoshka Sentence Embeddings (2DMSE) revisits this concept by proposing an approach to train embedding models that will perform well when only using a selection of all layers. This results in faster inference speeds at relatively low performance costs.

For example, using Sentence Transformers, you can train an Adaptive Layer model that can be sped up by 2x at a 15% reduction in performance, or 5x on GPU & 10x on CPU for a 20% reduction in performance. The 2DMSE paper highlights scenarios where this is superior to using a smaller model.

Training

Training with Adaptive Layer support is quite elementary: rather than applying some loss function on only the last layer, we also apply that same loss function on the pooled embeddings from previous layers. Additionally, we employ a KL-divergence loss that aims to make the embeddings of the non-last layers match that of the last layer. This can be seen as a fascinating approach of knowledge distillation, but with the last layer as the teacher model and the prior layers as the student models.

For example, with the 12-layer microsoft/mpnet-base, it will now be trained such that the model produces meaningful embeddings after each of the 12 layers.

from sentence_transformers import SentenceTransformer
from sentence_transformers.losses import CoSENTLoss, AdaptiveLayerLoss

model = SentenceTransformer("microsoft/mpnet-base")

base_loss = CoSENTLoss(model=model)
loss = AdaptiveLayerLoss(model=model, loss=base_loss)

Additionally, this can be combined with the MatryoshkaLoss such that the resulting model can be reduced both in the number of layers, but also in the size of the output dimensions. See also the Matryoshka Embeddings for more information on reducing output dimensions. In Sentence Transformers, the combination of these two losses is called Matryoshka2dLoss, and a shorthand is provided for simpler training.

from sentence_transformers import SentenceTransformer
from sentence_transformers.losses import CoSENTLoss, Matryoshka2dLoss

model = SentenceTransformer("microsoft/mpnet-base")

base_loss = CoSENTLoss(model=model)
loss = Matryoshka2dLoss(model=model, loss=base_loss, matryoshka_dims=[768, 512, 256, 128, 64])
Performance Results

Results

Let's look at the performance that we may be able to expect from an Adaptive Layer embedding model versus a regular embedding model. For this experiment, I have trained two models:

Both of these models were trained on the AllNLI dataset, which is a concatenation of the SNLI and MultiNLI datasets. I have evaluated these models on the STSBenchmark test set using multiple different embedding dimensions. The results are plotted in the following figure:

adaptive_layer_results

The first figure shows that the Adaptive Layer model stays much more performant when reducing the number of layers in the model. This is also clearly shown in the second figure, which displays that 80% of the performance is preserved when the number of layers is reduced all the way to 1.

Lastly, the third figure shows the expected speedup ratio for GPU & CPU devices in my tests. As you can see, removing half of the layers results in roughly a 2x speedup, at a cost of ~15% performance on STSB (~86 -> ~75 Spearman correlation). When removing even more layers, the performance benefit gets larger for CPUs, and between 5x and 10x speedups are very feasible with a 20% loss in performance.

Inference

Inference

After a model has been trained using the Adaptive Layer loss, you can then truncate the model layers to your desired layer count. Note that this requires doing a bit of surgery on the model itself, and each model is structured a bit differently, so the steps are slightly different depending on the model.

First of all, we will load the model & access the underlying transformers model like so:

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("tomaarsen/mpnet-base-nli-adaptive-layer")

# We can access the underlying model with `model[0].auto_model`
print(model[0].auto_model)
MPNetModel(
  (embeddings): MPNetEmbeddings(
    (word_embeddings): Embedding(30527, 768, padding_idx=1)
    (position_embeddings): Embedding(514, 768, padding_idx=1)
    (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
    (dropout): Dropout(p=0.1, inplace=False)
  )
  (encoder): MPNetEncoder(
    (layer): ModuleList(
      (0-11): 12 x MPNetLayer(
        (attention): MPNetAttention(
          (attn): MPNetSelfAttention(
            (q): Linear(in_features=768, out_features=768, bias=True)
            (k): Linear(in_features=768, out_features=768, bias=True)
            (v): Linear(in_features=768, out_features=768, bias=True)
            (o): Linear(in_features=768, out_features=768, bias=True)
            (dropout): Dropout(p=0.1, inplace=False)
          )
          (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
          (dropout): Dropout(p=0.1, inplace=False)
        )
        (intermediate): MPNetIntermediate(
          (dense): Linear(in_features=768, out_features=3072, bias=True)
          (intermediate_act_fn): GELUActivation()
        )
        (output): MPNetOutput(
          (dense): Linear(in_features=3072, out_features=768, bias=True)
          (LayerNorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
          (dropout): Dropout(p=0.1, inplace=False)
        )
      )
    )
    (relative_attention_bias): Embedding(32, 12)
  )
  (pooler): MPNetPooler(
    (dense): Linear(in_features=768, out_features=768, bias=True)
    (activation): Tanh()
  )
)

This output will differ depending on the model. We will look for the repeated layers in the encoder. For this MPNet model, this is stored under model[0].auto_model.encoder.layer. Then we can slice the model to only keep the first few layers to speed up the model:

new_num_layers = 3
model[0].auto_model.encoder.layer = model[0].auto_model.encoder.layer[:new_num_layers]

Then we can run inference with it using SentenceTransformers.encode.

from sentence_transformers import SentenceTransformer
from sentence_transformers.util import cos_sim

model = SentenceTransformer("tomaarsen/mpnet-base-nli-adaptive-layer")
new_num_layers = 3
model[0].auto_model.encoder.layer = model[0].auto_model.encoder.layer[:new_num_layers]

embeddings = model.encode(
    [
        "The weather is so nice!",
        "It's so sunny outside!",
        "He drove to the stadium.",
    ]
)
# Similarity of the first sentence with the other two
similarities = cos_sim(embeddings[0], embeddings[1:])
# => tensor([[0.7761, 0.1655]])
# compared to tensor([[ 0.7547, -0.0162]]) for the full model

As you can see, the similarity between the related sentences is much higher than the unrelated sentence, despite only using 3 layers. Feel free to copy this script locally, modify the new_num_layers, and observe the difference in similarities.

Extra information:

Example training scripts:

CrossEncoder (re)rank (#2514)

CrossEncoder models are often even better than biencoder (SentenceTransformer) models, as the model can compare two texts using the attention mechanism, unlike biencoders. However, they are more computationally expensive as well. They are commonly used for reranking the top retrieval results of a biencoder model. As of this release, that should now be more convenient!

We now support a rank method, which allows you to rank a bunch of documents given a query:

from sentence_transformers.cross_encoder import CrossEncoder

# Pre-trained cross encoder
model = CrossEncoder("cross-encoder/stsb-distilroberta-base")

# We want to compute the similarity between the query sentence
query = "A man is eating pasta."

# With all sentences in the corpus
corpus = [
    "A man is eating food.",
    "A man is eating a piece of bread.",
    "The girl is carrying a baby.",
    "A man is riding a horse.",
    "A woman is playing violin.",
    "Two men pushed carts through the woods.",
    "A man is riding a white horse on an enclosed ground.",
    "A monkey is playing drums.",
    "A cheetah is running behind its prey.",
]

# We rank all sentences in the corpus for the query
ranks = model.rank(query, corpus)

# Print the scores
print("Query:", query)
for rank in ranks:
    print(f"{rank['score']:.2f}\t{corpus[rank['corpus_id']]}")
0.67    A man is eating food.
0.34    A man is eating a piece of bread.
0.08    A man is riding a horse.
0.07    A man is riding a white horse on an enclosed ground.
0.01    The girl is carrying a baby.
0.01    Two men pushed carts through the woods.
0.01    A monkey is playing drums.
0.01    A woman is playing violin.

Extra information:

All changes

New Contributors

I especially want to thank @SeanLee97 and @fkdosilovic for their valuable contributions in this release.

Full Changelog: v2.4.0...v2.5.0

Contributors

fkdosilovic, fxmarty, and 4 other contributors
Assets 2
Loading