Adalayers

Frozen transformer models adaptive layers selection for downstream tasks efficient solving.

---

Best results

Note: results are highly dependent on base model domain. Clearly, some task-specific model could have been used to achieve even higher scores. Though, such comparison would not be fair.

As you see, currently just one frozen base model with proposed algorithm achieves great results

Dataset	Base model	Score
IMDB	RoBERTa-large	96.1% acc (SOTA level)
CoLA	RoBERTa-large	83.6% acc (SOTA level)
CoNLL	RoBERTa-large	89.4% f1

Paper

Using Internal Representations of Frozen Transformer Models to Solve Downstream Tasks

Problem

Let's consider the case when you already have a functioning SOTA-level large transformer model and you need to solve a different task on the same data. Aka speech recognition + emotion recognition, text summarization + NER, etc.

One possible solution is to use a second model. However, deploying a second transformer model requires lots of resources. Combining two tasks into a single model training is not always feasible and may deteriorate main task metrics.

Solution

Let's reuse transformer hidden states! It is well-known that different layers of the transformer model extract features of different levels. Therefore, if we effectively combine hidden features, we could achieve good results.

Moreover, in such a case, the base model stays intact, and as we reuse its computations, the proposed algorithm is highly computationally efficient.

The general idea is presented in the image

Code for F can be found in adalayers/models/ada_layers_base.py. All models are implemented with Huggingface interfaces.

Launch

Training is omegaconf+hydra configurable. Configs can be found in configs.

The environment is poetry-controlled. You can set it up by calling poetry install

You can launch simple training by calling adalayers/training/run.py like

python adalayers/training/run.py --config-name adalayers_imdb.yaml