# Overview: Efficient Fine-Tuning Methods

Large pre-trained Transformer-based language models (LMs) have become the foundation of NLP in recent years.
While the most prevalent method of using these LMs for transfer learning involves costly *full fine-tuning* of all model parameters, a series of *efficient* and *lightweight* alternatives have been established in recent time.
Instead of updating all parameters of the pre-trained LM towards a downstream target task, these methods commonly introduce a small amount of new parameters and only update these while keeping the pre-trained model weights fixed.

```{admonition} Why use Efficient Fine-Tuning?
Efficient fine-tuning methods offer multiple benefits over full fine-tuning of LMs:

- They are **parameter-efficient**, i.e. they only update a very small subset (often under 1%) of a model's parameters.
- They often are **modular**, i.e. the updated parameters can be extracted and shared independently of the base model parameters.
- They are easy to share and easy to deploy due to their **small file sizes**, e.g. having only ~3MB per task instead of ~440MB for sharing a full model.
- They **speed up training**, i.e. efficient fine-tuning often needs less time for training compared fully fine-tuning LMs.
- They are **composable**, e.g. multiple adapters trained on different tasks can be stacked, fused or mixed to leverage their combined knowledge.
- They often provide **on-par performance** with full fine-tuning.
```

More specifically, let the parameters of a LM be composed of a set of pre-trained parameters $\Theta$ (frozen) and a set of (newly introduced) parameters $\Phi$.
Then, efficient fine-tuning methods optimize only $\Phi$ according to a loss function $L$ on a dataset $D$:

$$
\Phi^* \leftarrow \arg \min_{\Phi} L(D; \{\Theta, \Phi\})
$$

Efficient fine-tuning might insert parameters $\Phi$ at different locations of a Transformer-based LM.
One early and successful method, (bottleneck) adapters, introduces bottleneck feed-forward layers in each layer of a Transformer model.
While these adapters have laid the foundation of the adapter-transformers library, multiple alternative methods have been introduced and integrated since.
In the following, we present all methods currently integrated into this library (see [here](https://github.com/adapter-hub/adapter-transformers#implemented-methods) for a tabular overview).

**Configuration:** All presented methods can be added, trained, saved and shared using the same set of model class functions (see [class documentation](transformers.ModelAdaptersMixin)).
Each method is specified and configured using a specific configuration class, all of which derive from the common [`AdapterConfigBase`](transformers.AdapterConfigBase) class.
E.g., adding one of the methods presented below to an existing model instance follows this scheme:
```python
config = ... # config class deriving from AdapterConfigBase
model.add_adapter("name", config=config)
```

```{eval-rst}
.. important::
    In literature, different terms are used to refer to efficient fine-tuning methods.
    The term "adapter" is usually only applied to bottleneck adapter modules.
    However, most efficient fine-tuning methods follow the same general idea of inserting a small set of new parameters and by this "adapting" the pre-trained LM to a new task.
    In adapter-transformers, the term "adapter" thus may refer to any efficient fine-tuning method if not specified otherwise.
```

## Bottleneck Adapters

_Configuration class_: [`AdapterConfig`](transformers.AdapterConfig)

Bottleneck adapters introduce bottleneck feed-forward layers in each layer of a Transformer model.
Generally, these adapter layers consist of a down-projection matrix $W_{down}$ that projects the layer hidden states into a lower dimension $d_{bottleneck}$, a non-linearity $f$, an up-projection $W_{up}$ that projects back into the original hidden layer dimension and a residual connection $r$:

$$
h \leftarrow W_{up} \cdot f(W_{down} \cdot h) + r
$$

Depending on the concrete adapter configuration, these layers can be introduced at different locations within a Transformer block. Further, residual connections, layer norms, activation functions and bottleneck sizes etc. can be configured.

The most important configuration hyperparameter to be highlighted here is the bottleneck dimension $d_{bottleneck}$.
In adapter-transformers, this bottleneck dimension is specified indirectly via the `reduction_factor` attribute of a configuration.
This `reduction_factor` defines the ratio between a model's layer hidden dimension and the bottleneck dimension, i.e.:

$$
\text{reduction_factor} = \frac{d_{hidden}}{d_{bottleneck}}
$$

A visualization of further configuration options related to the adapter structure is given in the figure below. For more details, refer to the documentation of [`AdapterConfig`](transformers.AdapterConfig).


```{eval-rst}
.. figure:: img/architecture.png
    :width: 350
    :align: center
    :alt: Adapter architectures

    Visualization of possible adapter configurations with corresponding dictionary keys.
```

adapter-transformers comes with pre-defined configurations for some bottleneck adapter architectures proposed in literature:

- [`HoulsbyConfig`](transformers.HoulsbyConfig) as proposed by [Houlsby et al. (2019)](https://arxiv.org/pdf/1902.00751.pdf) places adapter layers after both the multi-head attention and feed-forward block in each Transformer layer.
- [`PfeifferConfig`](transformers.PfeifferConfig) as proposed by [Pfeiffer et al. (2020)](https://arxiv.org/pdf/2005.00052.pdf) places an adapter layer only after the feed-forward block in each Transformer layer.
- [`ParallelConfig`](transformers.ParallelConfig) as proposed by [He et al. (2021)](https://arxiv.org/pdf/2110.04366.pdf) places adapter layers in parallel to the original Transformer layers.

_Example_:
```python
from transformers.adapters import AdapterConfig

config = AdapterConfig(mh_adapter=True, output_adapter=True, reduction_factor=16, non_linearity="relu")
model.add_adapter("bottleneck_adapter", config=config)
```

_Papers:_

* [Parameter-Efficient Transfer Learning for NLP](https://arxiv.org/pdf/1902.00751.pdf) (Houlsby et al., 2019)
* [Simple, Scalable Adaptation for Neural Machine Translation](https://arxiv.org/pdf/1909.08478.pdf) (Bapna and Firat, 2019)
* [AdapterFusion: Non-Destructive Task Composition for Transfer Learning](https://aclanthology.org/2021.eacl-main.39.pdf) (Pfeiffer et al., 2021)
* [AdapterHub: A Framework for Adapting Transformers](https://arxiv.org/pdf/2007.07779.pdf) (Pfeiffer et al., 2020)

## Language Adapters - Invertible Adapters

_Configuration class_: [`PfeifferInvConfig`](transformers.PfeifferInvConfig), [`HoulsbyInvConfig`](transformers.HoulsbyInvConfig)

The MAD-X setup ([Pfeiffer et al., 2020](https://arxiv.org/pdf/2005.00052.pdf)) proposes language adapters to learn language-specific transformations.
After being trained on a language modeling task, a language adapter can be stacked before a task adapter for training on a downstream task.
To perform zero-shot cross-lingual transfer, one language adapter can simply be replaced by another.

In terms of architecture, language adapters are largely similar to regular bottleneck adapters, except for an additional _invertible adapter_ layer after the LM embedding layer.
Embedding outputs are passed through this invertible adapter in the forward direction before entering the first Transformer layer and in the inverse direction after leaving the last Transformer layer.
Invertible adapter architectures are further detailed in [Pfeiffer et al. (2020)](https://arxiv.org/pdf/2005.00052.pdf) and can be configured via the `inv_adapter` attribute of the `AdapterConfig` class.

_Example_:
```python
from transformers.adapters import PfeifferInvConfig

config = PfeifferInvConfig()
model.add_adapter("lang_adapter", config=config)
```

_Papers:_
- [MAD-X: An Adapter-based Framework for Multi-task Cross-lingual Transfer](https://arxiv.org/pdf/2005.00052.pdf) (Pfeiffer et al., 2020)

```{eval-rst}
.. note::
    V1.x of adapter-transformers made a distinction between task adapters (without invertible adapters) and language adapters (with invertible adapters) with the help of the ``AdapterType`` enumeration.
    This distinction was dropped with v2.x.
```

## Prefix Tuning

_Configuration class_: [`PrefixTuningConfig`](transformers.PrefixTuningConfig)

Prefix Tuning ([Li and Liang, 2021](https://aclanthology.org/2021.acl-long.353.pdf)) introduces new parameters in the multi-head attention blocks in each Transformer layer.
More, specifically, it prepends trainable prefix vectors $P^K$ and $P^V$ to the keys and values of the attention head input, each of a configurable prefix length $l$ (`prefix_length` attribute):

$$
head_i = \text{Attention}(Q W_i^Q, [P_i^K, K W_i^K], [P_i^V, V W_i^V])
$$

Following the original authors, the prefix vectors in $P^K$ and $P^V$ are note optimized directly, but reparameterized via a bottleneck MLP.
This behavior is controlled via the `flat` attribute of the configuration.
Using `PrefixTuningConfig(flat=True)` will create prefix tuning vectors that are optimized without reparameterization.

_Example_:
```python
from transformers.adapters import PrefixTuningConfig

config = PrefixTuningConfig(flat=False, prefix_length=30)
model.add_adapter("prefix_tuning", config=config)
```

As reparameterization using the bottleneck MLP is not necessary for performing inference on an already trained Prefix Tuning module, adapter-transformers includes a function to "eject" a reparameterized Prefix Tuning into a flat one:
```python
model.eject_prefix_tuning("prefix_tuning")
```
This will only retain the necessary parameters and reduces the size of the trained Prefix Tuning.

_Papers:_
- [Prefix-Tuning: Optimizing Continuous Prompts for Generation](https://arxiv.org/pdf/2101.00190.pdf) (Li and Liang, 2021)

## LoRA

_Configuration class_: [`LoRAConfig`](transformers.LoRAConfig)

Low-Rank Adaptation (LoRA) is an efficient fine-tuning technique proposed by [Hu et al. (2021)](https://arxiv.org/pdf/2106.09685.pdf).
LoRA injects trainable low-rank decomposition matrices into the layers of a pre-trained model.
For any model layer expressed as a matrix multiplication of the form $h = W_0 x$, it therefore performs a reparameterization, such that:

$$
h = W_0 x + \frac{\alpha}{r} B A x
$$

Here, $A \in \mathbb{R}^{r\times k}$ and $B \in \mathbb{R}^{d\times r}$ are the decomposition matrices and $r$, the low-dimensional rank of the decomposition, is the most important hyperparameter.

While, in principle, this reparameterization can be applied to any weights matrix in a model, the original paper only adapts the attention weights of the Transformer self-attention sub-layer with LoRA.
`adapter-transformers` additionally allows injecting LoRA into the dense feed-forward layers in the intermediate and output components of a Transformer block.
You can configure the locations where LoRA weights should be injected using the attributes in the [`LoRAConfig`](transformers.LoRAConfig) class.

_Example_:
```python
from transformers.adapters import LoRAConfig

config = LoRAConfig(r=8, alpha=16)
model.add_adapter("lora_adapter", config=config)
```

In the design of LoRA, Hu et al. (2021) also pay special attention to keeping the inference latency overhead compared to full fine-tuning at a minumum.
To accomplish this, the LoRA reparameterization can be merged with the original pre-trained weights of a model for inference.
Thus, the adapted weights are directly used in every forward pass without passing activations through an additional module.
In `adapter-transformers`, this can be realized using the built-in `merge_lora()` method:
```python
model.merge_lora("lora_adapter")
```

To continue training on this LoRA adapter or to deactivate it entirely, the merged weights first have to be reset again:
```python
model.reset_lora("lora_adapter")
```

_Papers:_
- [LoRA: Low-Rank Adaptation of Large Language Models](https://arxiv.org/pdf/2106.09685.pdf) (Hu et al., 2021)

## Compacter

_Configuration class_: [`CompacterConfig`](transformers.CompacterConfig), [`CompacterPlusPlusConfig`](transformers.CompacterPlusPlusConfig)

The Compacter architecture proposed by [Mahabadi et al., 2021](https://arxiv.org/pdf/2106.04647.pdf)
is similar to the bottleneck adapter architecture. It only exchanges the linear down- and 
up-projection with a PHM layer. Unlike the linear layer, the PHM layer constructs its weight matrix from two smaller matrices, which reduces the number of parameters.
 These matrices can be factorized and shared between all adapter layers. You can exchange the down- and up-projection layers from any of the bottleneck adapters described in the previous section
for a PHM layer by specifying `use_phm=True` in the config.

The PHM layer has the following additional properties: `phm_dim`, `shared_phm_rule`, `factorized_phm_rule`, `learn_phm`, 
`factorized_phm_W`, `shared_W_phm`, `phm_c_init`, `phm_init_range`, `hypercomplex_nonlinearity`

For more information check out the [`AdapterConfig`](transformers.AdapterConfig) class.

To add a Compacter to your model you can use the predefined configs:
```python
from transformers.adapters import CompacterConfig

config = CompacterConfig()
model.add_adapter("dummy", config=config)
```
_Papers:_
- [COMPACTER: Efficient Low-Rank Hypercomplex Adapter Layers](https://arxiv.org/pdf/2106.04647.pdf) (Mahabadi, Henderson and Ruder, 2021)

## Combinations - Mix-and-Match Adapters

_Configuration class_: [`ConfigUnion`](transformers.ConfigUnion)

While different efficient fine-tuning methods and configurations have often been proposed as standalone, it might be beneficial to combine them for joint training.
To make this process easier, adapter-transformers provides the possibility to group multiple configuration instances together using the `ConfigUnion` class.

For example, this could be used to define different reduction factors for the adapter modules placed after the multi-head attention and the feed-forward blocks:

```python
from transformers.adapters import AdapterConfig, ConfigUnion

config = ConfigUnion(
    AdapterConfig(mh_adapter=True, output_adapter=False, reduction_factor=16, non_linearity="relu"),
    AdapterConfig(mh_adapter=False, output_adapter=True, reduction_factor=2, non_linearity="relu"),
)
model.add_adapter("union_adapter", config=config)
```

[He et al. (2021)](https://arxiv.org/pdf/2110.04366.pdf) study various variants and combinations of efficient fine-tuning methods.
Among others, they propose _Mix-and-Match Adapters_ as a combination of Prefix Tuning and parallel bottleneck adapters.
This configuration is supported by adapter-transformers out-of-the-box:

```python
from transformers.adapters import MAMConfig

config = MAMConfig()
model.add_adapter("mam_adapter", config=config)
```

and is identical to using the following `ConfigUnion`:

```python
from transformers.adapters import ConfigUnion, ParallelConfig, PrefixTuningConfig

config = ConfigUnion(
    PrefixTuningConfig(bottleneck_size=800),
    ParallelConfig(),
)
model.add_adapter("mam_adapter", config=config)
```

_Papers:_
- [Towards a Unified View of Parameter-Efficient Transfer Learning](https://arxiv.org/pdf/2110.04366.pdf) (He et al., 2021)