Hugging Face Search models, datasets, users...

Models Datasets Spaces Posts Docs Pricing Log In Sign Up

Back to blog

Leveraging Pre-trained Language

Model Checkpoints for Encoder-
Decoder Models
Published November 9, 2020

Update on GitHub

patrickvonplaten
Patrick von Platen

Open in Colab

Transformer-based encoder-decoder models were proposed in Vaswani et al. (2017) and have
recently experienced a surge of interest, e.g. Lewis et al. (2019), Raffel et al. (2019), Zhang et
al. (2020), Zaheer et al. (2020), Yan et al. (2020).

Similar to BERT and GPT2, massive pre-trained encoder-decoder models have shown to
significantly boost performance on a variety of sequence-to-sequence tasks Lewis et al. (2019),
Raffel et al. (2019). However, due to the enormous computational cost attached to pre-training
encoder-decoder models, the development of such models is mainly limited to large companies
and institutes.

In Leveraging Pre-trained Checkpoints for Sequence Generation Tasks (2020), Sascha Rothe,
Shashi Narayan and Aliaksei Severyn initialize encoder-decoder model with pre-trained
encoder and/or decoder-only checkpoints (e.g. BERT, GPT2) to skip the costly pre-training. The
authors show that such warm-started encoder-decoder models yield competitive results to
large pre-trained encoder-decoder models, such as T5, and Pegasus on multiple sequence-to-
sequence tasks at a fraction of the training cost.

In this notebook, we will explain in detail how encoder-decoder models can be warm-started,
give practical tips based on Rothe et al. (2020), and finally go over a complete code example
showing how to warm-start encoder-decoder models with 🤗Transformers.
This notebook is divided into 4 parts:

Introduction - Short summary of pre-trained language models in NLP and the need for
warm-starting encoder-decoder models.

Warm-starting encoder-decoder models (Theory) - Illustrative explanation on how

encoder-decoder models are warm-started?

Warm-starting encoder-decoder models (Analysis) - Summary of Leveraging Pre-

trained Checkpoints for Sequence Generation Tasks (2020) - What model combinations are
effective to warm-start encoder-decoder models; How does it differ from task to task?

Warm-starting encoder-decoder models with 🤗Transformers (Practice) - Complete

code example showcasing in-detail how to use the EncoderDecoderModel framework to warm-
start transformer-based encoder-decoder models.

It is highly recommended (probably even necessary) to have read this blog post about
transformer-based encoder-decoder models.
Let's start by giving some back-ground on warm-starting encoder-decoder models.

Introduction

Recently, pre-trained language models 1 have revolutionized the field of natural language
processing (NLP).

The first pre-trained language models were based on recurrent neural networks (RNN) as
proposed Dai et al. (2015). Dai et. al showed that pre-training an RNN-based model on
unlabelled data and subsequently fine-tuning 2 it on a specific task yields better results than
training a randomly initialized model directly on such a task. However, it was only in 2018,
when pre-trained language models become widely accepted in NLP. ELMO by Peters et al. and
ULMFit by Howard et al. were the first pre-trained language model to significantly improve the
state-of-the-art on an array of natural language understanding (NLU) tasks. Just a couple of
months later, OpenAI and Google published transformer-based pre-trained language models,
called GPT by Radford et al. and BERT by Devlin et al. respectively. The improved efficiency of
transformer-based language models over RNNs allowed GPT2 and BERT to be pre-trained on
massive amounts of unlabeled text data. Once pre-trained, BERT and GPT were shown to
require very little fine-tuning to shatter state-of-art results on more than a dozen NLU tasks 3 .

The capability of pre-trained language models to effectively transfer task-agnostic knowledge

to task-specific knowledge turned out to be a great catalyst for NLU. Whereas engineers and
researchers previously had to train a language model from scratch, now publicly available
checkpoints of large pre-trained language models can be fine-tuned at a fraction of the cost and
time. This can save millions in industry and allows for faster prototyping and better
benchmarks in research.

Pre-trained language models have established a new level of performance on NLU tasks and
more and more research has been built upon leveraging such pre-trained language models for
improved NLU systems. However, standalone BERT and GPT models have been less successful
for sequence-to-sequence tasks, e.g. text-summarization, machine translation, sentence-
rephrasing, etc.

Sequence-to-sequence tasks are defined as a mapping from an input sequence X1:n to an ​

output sequence Y 1:m of a-priori unknown output length m. Hence, a sequence-to-sequence

​

model should define the conditional probability distribution of the output sequence Y 1:m ​

conditioned on the input sequence X1:n : ​

pθmodel (Y 1:m ∣X1:n ).

​
​ ​ ​

Without loss of generality, an input word sequence of n words is hereby represented by the
vector sequnece X1:n ​ = x1 , … , xn and an output sequence of m words as Y 1:m =
​ ​ ​

y1 , … , ym .
​ ​

Let's see how BERT and GPT2 would be fit to model sequence-to-sequence tasks.

BERT

BERT is an encoder-only model, which maps an input sequence X1:n to a contextualized ​

encoded sequence X1:n : ​

fθBERT : X1:n → X1:n .

​
​ ​ ​

BERT's contextualized encoded sequence X1:n can then further be processed by a classification
​

layer for NLU classification tasks, such as sentiment analysis, natural language inference, etc. To
do so, the classification layer, i.e. typically a pooling layer followed by a feed-forward layer, is
added as a final layer on top of BERT to map the contextualized encoded sequence X1:n to a ​

class c :

fθp,c : X1:n → c.
​ ​

It has been shown that adding a pooling- and classification layer, defined as θp,c , on top of a
​

pre-trained BERT model θBERT and subsequently fine-tuning the complete model
​

{θp,c , θBERT } can yield state-of-the-art performances on a variety of NLU tasks, cf. to BERT by
​ ​

Devlin et al..

Let's visualize BERT.

The BERT model is shown in grey. The model stacks multiple BERT blocks, each of which is
composed of bi-directional self-attention layers (shown in the lower part of the red box) and
two feed-forward layers (short in the upper part of the red box).

Each BERT block makes use of bi-directional self-attention to process an input sequence
x′ 1 , … , x′ n (shown in light grey) to a more "refined" contextualized output sequence
​ ​

x′′ 1 , … , x′′ n (shown in slightly darker grey) 4 . The contextualized output sequence of the
​ ​

final BERT block, i.e. X1:n , can then be mapped to a single output class c by adding a task-
​

specific classification layer (shown in orange) as explained above.

Encoder-only models can only map an input sequence to an output sequence of a priori known
output length. In conclusion, the output dimension does not depend on the input sequence,
which makes it disadvantageous and impractical to use encoder-only models for sequence-to-
sequence tasks.

As for all encoder-only models, BERT's architecture corresponds exactly to the architecture of
the encoder part of transformer-based encoder-decoder models as shown in the "Encoder"
section in the Encoder-Decoder notebook.

GPT2

GPT2 is a decoder-only model, which makes use of uni-directional (i.e. "causal") self-attention
to define a mapping from an input sequence Y 0:m−1 1 to a "next-word" logit vector sequence
​

L1:m : ​

fθGPT2 : Y 0:m−1 → L1:m .

​
​ ​ ​

By processing the logit vectors L1:m with the softmax operation, the model can define the
​

probability distribution of the word sequence Y 1:m . To be exact, the probability distribution of
​

the word sequence Y 1:m can be factorized into m − 1 conditional "next word" distributions:
​
 m
pθGPT2 (Y 1:m ) = ∏ pθGPT2 (yi ∣Y 0:i−1 ).
​
​ ​ ​

​
​ ​ ​

i=1

pθGPT2 (yi ∣Y 0:i−1 ) hereby presents the probability distribution of the next word yi given all
​
​ ​ ​ ​

previous words y0 , … , yi−1 3 and is defined as the softmax operation applied on the logit
​ ​

vector li . To summarize, the following equations hold true.

​

pθgpt2 (yi ∣Y 0:i−1 ) = Softmax(li ) = Softmax(fθGPT2 (Y 0:i−1 )).

​
​ ​ ​ ​

​
​ ​

For more detail, please refer to the decoder section of the encoder-decoder blog post.

Let's visualize GPT2 now as well.

Analogous to BERT, GPT2 is composed of a stack of GPT2 blocks. In contrast to BERT block,
GPT2 block makes use of uni-directional self-attention to process some input vectors
y′ 0 , … , y′ m−1 (shown in light blue on the bottom right) to an output vector sequence
​ ​

y′′ 0 , … , y′′ m−1 (shown in darker blue on the top right). In addition to the GPT2 block stack,
​ ​

the model also has a linear layer, called LM Head, which maps the output vectors of the final
GPT2 block to the logit vectors l1 , … , lm . As mentioned earlier, a logit vector li can then be​ ​ ​

5
used to sample of new input vector yi . ​

GPT2 is mainly used for open-domain text generation. First, an input prompt Y 0:i−1 is fed to ​

the model to yield the conditional distribution pθgpt2 (y∣Y 0:i−1 ). Then the next word yi is ​
​ ​ ​

sampled from the distribution (represented by the grey arrows in the graph above) and
consequently append to the input. In an auto-regressive fashion the word yi+1 can then be ​

sampled from pθgpt2 (y∣Y 0:i ) and so on. ​

​ ​

GPT2 is therefore well-suited for language generation, but less so for conditional generation. By
setting the input prompt Y 0:i−1 equal to the sequence input X1:n , GPT2 can very well be used
​ ​

for conditional generation. However, the model architecture has a fundamental drawback
compared to the encoder-decoder architecture as explained in Raffel et al. (2019) on page 17.
In short, uni-directional self-attention forces the model's representation of the sequence input
X1:n to be unnecessarily limited since xi cannot depend on xi+1 , ∀i ∈ {1, … , n}.
​ ​ ​

Encoder-Decoder

Because encoder-only models require to know the output length a priori, they seem unfit for
sequence-to-sequence tasks. Decoder-only models can function well for sequence-to-sequence
tasks, but also have certain architectural limitations as explained above.

The current predominant approach to tackle sequence-to-sequence tasks are transformer-based

encoder-decoder models - often also called seq2seq transformer models. Encoder-decoder
models were introduced in Vaswani et al. (2017) and since then have been shown to perform
better on sequence-to-sequence tasks than stand-alone language models (i.e. decoder-only
models), e.g. Raffel et al. (2020). In essence, an encoder-decoder model is the combination of a
stand-alone encoder, such as BERT, and a stand-alone decoder model, such as GPT2. For more
details on the exact architecture of transformer-based encoder-decoder models, please refer to
this blog post.

Now, we know that freely available checkpoints of large pre-trained stand-alone encoder and
decoder models, such as BERT and GPT, can boost performance and reduce training cost for
many NLU tasks, We also know that encoder-decoder models are essentially the combination of
stand-alone encoder and decoder models. This naturally brings up the question of how one can
leverage stand-alone model checkpoints for encoder-decoder models and which model
combinations are most performant on certain sequence-to-sequence tasks.

In 2020, Sascha Rothe, Shashi Narayan, and Aliaksei Severyn investigated exactly this question
in their paper Leveraging Pre-trained Checkpoints for Sequence Generation Tasks. The
paper offers a great analysis of different encoder-decoder model combinations and fine-tuning
techniques, which we will study in more detail later.

Composing an encoder-decoder model of pre-trained stand-alone model checkpoints is defined

as warm-starting the encoder-decoder model. The following sections show how warm-starting
an encoder-decoder model works in theory, how one can put the theory into practice with 🤗
Transformers, and also gives practical tips for better performance.

1
A pre-trained language model is defined as a neural network:

that has been trained on unlabeled text data, i.e. in a task-agnostic, unsupervised fashion,
and

that processes a sequence of input words into a context-dependent embedding. E.g. the
continuous bag-of-words and skip-gram model from Mikolov et al. (2013) is not
considered a pre-trained language model because the embeddings are context-agnostic.

2
Fine-tuning is defined as the task-specific training of a model that has been initialized with the
weights of a pre-trained language model.

3
The input vector y0 corresponds hereby to the BOS embedding vector required to predict
​

the very first output word y1 . ​

4
Without loss of generalitiy, we exclude the normalization layers to not clutter the equations
and illustrations.

5
For more detail on why uni-directional self-attention is used for "decoder-only" models, such
as GPT2, and how sampling works exactly, please refer to the decoder section of the encoder-
decoder blog post.

Warm-starting encoder-decoder models (Theory)

Having read the introduction, we are now familiar with encoder-only- and decoder-only models.
We have noticed that the encoder-decoder model architecture is essentially a composition of a
stand-alone encoder model and a stand-alone decoder model, which led us to the question of
how one can warm-start encoder-decoder models from stand-alone model checkpoints.

There are multiple possibilities to warm-start an encoder-decoder model. One can

1. initialize both the encoder and decoder part from an encoder-only model checkpoint, e.g.
BERT,

2. initialize the encoder part from an encoder-only model checkpoint, e.g. BERT, and the
decoder part from and a decoder-only checkpoint, e.g. GPT2,

3. initialize only the encoder part with an encoder-only model checkpoint, or

4. initialize only the decoder part with a decoder-only model checkpoint.

In the following, we will put the focus on possibilities 1. and 2. Possibilities 3. and 4. are trivial
after having understood the first two.

Recap Encoder-Decoder Model

First, let's do a quick recap of the encoder-decoder architecture.

The encoder (shown in green) is a stack of encoder blocks. Each encoder block is composed of a
bi-directional self-attention layer, and two feed-forward layers 1 . The decoder (shown in
orange) is a stack of decoder blocks, followed by a dense layer, called LM Head. Each decoder
block is composed of a uni-directional self-attention layer, a cross-attention layer, and two feed-
forward layers.

The encoder maps the input sequence X1:n to a contextualized encoded sequence X1:n in the
​ ​

exact same way BERT does. The decoder then maps the contextualized encoded sequence
X1:n and a target sequence Y 0:m−1 to the logit vectors L1:m . Analogous to GPT2, the logits
​ ​ ​

are then used to define the distribution of the target sequence Y 1:m conditioned on the input ​

sequence X1:n by means of a softmax operation.

​

To put it into mathematical terms, first, the conditional distribution is factorized into m − 1
conditional distributions of the next word yi by Bayes' rule. ​

m
pθenc, dec (Y 1:m ∣X1:n ) = pθdec (Y 1:m ∣X1:n ) = ∏ pθdec (yi ∣Y 0:i−1 , X1:n ), with X1:n = fθenc (X1:n ).
​
​ ​ ​

​
​ ​ ​ ​

​
​ ​ ​ ​ ​

​
​ ​

i=1

Each "next-word" conditional distributions is thereby defined by the softmax of the logit vector
as follows.

pθdec (yi ∣Y 0:i−1 , X1:n ) = Softmax(li ).

​
​ ​ ​ ​
For more detail, please refer to the Encoder-Decoder notebook.

Warm-staring Encoder-Decoder with BERT

Let's now illustrate how a pre-trained BERT model can be used to warm-start the encoder-
decoder model. BERT's pre-trained weight parameters are used to both initialize the encoder's
weight parameters as well as the decoder's weight parameters. To do so, BERT's architecture is
compared to the encoder's architecture and all layers of the encoder that also exist in BERT will
be initialized with the pre-trained weight parameters of the respective layers. All layers of the
encoder that do not exist in BERT will simply have their weight parameters be randomly
initialized.

Let's visualize.

We can see that the encoder architecture corresponds 1-to-1 to BERT's architecture. The
weight parameters of the bi-directional self-attention layer and the two feed-forward layers of
all encoder blocks are initialized with the weight parameters of the respective BERT blocks.
This is illustrated examplary for the second encoder block (red boxes at bottow) whose weight
self-attn,2 feed-forward,2 feed-forward,2
parameters θenc and θenc
​
are set to BERT's weight parameters θBERT
​ ​

self-attn,2
and θBERT , respectively at initialization.
​

Before fine-tuning, the encoder therefore behaves exactly like a pre-trained BERT model.
Assuming the input sequence x1 , … , xn (shown in green) passed to the encoder is equal to
​ ​

the input sequence xBERT

1 , … , xBERT
​

n (shown in grey) passed to BERT, this means that the

​

respective output vector sequences x1 , … , xn (shown in darker green) and

​ ​

xBERT
1 , … , xBERT
​

n (shown in darker grey) also have to be equal.

​

Next, let's illustrate how the decoder is warm-started.

The architecture of the decoder is different from BERT's architecture in three ways.

1. First, the decoder has to be conditioned on the contextualized encoded sequence X1:n by ​

means of cross-attention layers. Consequently, randomly initialized cross-attention layers

are added between the self-attention layer and the two feed-forward layers in each BERT
cross-attention, 2
block. This is represented exemplary for the second block by +θdec ​ and
illustrated by the newly added fully connected graph in red in the lower red box on the
right. This necessarily changes the behavior of each modified BERT block so that an input
vector, e.g. y′ 0 now yields a random output vector y′′ 0 (highlighted by the red border
​ ​

around the output vector y′′ 0 ).

​

2. Second, BERT's bi-directional self-attention layers have to be changed to uni-directional

self-attention layers to comply with auto-regressive generation. Because both the bi-
directional and the uni-directional self-attention layer are based on the same key, query
and value projection weights, the decoder's self-attention layer weights can be initialized
with BERT's self-attention layer weights. E.g. the query, key and value weight parameters
of the decoder's uni-directional self-attention layer are initialized with those of BERT's bi-
self-attn,2
directional self-attention layer θBERT ​ =
self-attn,2 self-attn,2 self-attn,2 self-attn,2
{WBERT,k , WBERT,v , WBERT,q } → θdec
​ ​ ​ = ​

self-attn,2 self-attn,2 self-attn,2

{Wdec,k , Wdec,v
​ ​, Wdec,q }. However, in uni-directional self-attention each
​

token only attends to all previous tokens, so that the decoder's self-attention layers yield
different output vectors than BERT's self-attention layers even though they share the same
weights. Compare e.g., the decoder's causally connected graph in the right box versus
BERT's fully connected graph in the left box.

3. Third, the decoder outputs a sequence of logit vectors L1:m in order to define the
​

conditional probability distribution pθdec (Y 1:n ∣X). As a result, a LM Head layer is added
​
​ ​

on top of the last decoder block. The weight parameters of the LM Head layer usually
correspond to the weight parameters of the word embedding Wemb and thus are not ​

word-emb
randomly initialized. This is illustrated in the top by the initialization θBERT ​ →
lm-head
θdec .
​

To conclude, when warm-starting the decoder from a pre-trained BERT model only the cross-
attention layer weights are randomly initialized. All other weights including those of the self-
attention layer and LM Head are initialized with BERT's pre-trained weight parameters.
Having warm-stared the encoder-decoder model, the weights are then fine-tuned on a
sequence-to-sequence downstream task, such as summarization.

Warm-staring Encoder-Decoder with BERT and GPT2

Instead of warm-starting both the encoder and decoder with a BERT checkpoint, we can
instead leverage the BERT checkpoint for the encoder and a GPT2 checkpoint for the decoder.
At first glance, a decoder-only GPT2 checkpoint seems to be better-suited to warm-start the
decoder because it has already been trained on causal language modeling and uses uni-
directional self-attention layers.

Let's illustrate how a GPT2 checkpoint can be used to warm-start the decoder.

We can see that decoder is more similar to GPT2 than it is to BERT. The weight parameters of
decoder's LM Head can directly be initialized with GPT2's LM Head weight parameters, e.g.
lm-head lm-head
θGPT2 → θdec
​ . In addition, the blocks of the decoder and GPT2 both make use of uni-
​

directional self-attention so that the output vectors of the decoder's self-attention layer are
GPT2
equivalent to GPT2's output vectors assuming the input vectors are the same, e.g. y′ 0 =​

y′ 0 . In contrast to the BERT-initialized decoder, the GPT2-initialized decoder, therefore, keeps

​

the causal connected graph of the self-attention layer as can be seen in the red boxes on the
bottom.

Nevertheless, the GPT2-initialized decoder also has to condition the decoder on X1:n . ​

Analoguos to the BERT-initialized decoder, randomly initialized weight parameters for the
cross-attention layer are therefore added to each decoder block. This is illustrated e.g. for the
cross-attention, 2
second encoder block by +θdec .
​

Even though GPT2 resembles the decoder part of an encoder-decoder model more than BERT,
a GPT2-initialized decoder will also yield random logit vectors L1:m without fine-tuning due to
​

randomly initialized cross-attention layers in every decoder block. It would be interesting to

investigate whether a GPT2-initialized decoder yields better results or can be fine-tuned more
efficiently.

Encoder-Decoder Weight Sharing

In Raffel et al. (2020), the authors show that a randomly-initialized encoder-decoder model
that shares the encoder's weights with the decoder, and therefore reduces the memory
footprint by half, performs only slightly worse than its "non-shared" version. Sharing the
encoder's weights with the decoder means that all layers of the decoder that are found at the
same position in the encoder share the same weight parameters, i.e. the same node in the
network's computation graph.
E.g. the query, key, and value projection matrices of the self-attention layer in the third
self-attn,3 self-attn,3 self-attn,3
encoder block, defined as WEnc,k , WEnc,v
​ , WEnc,q
​ ​ are identical to the respective
query, key, and value projections matrices of the self-attention layer in the third decoder block
2
:

Wself-attn,3
k = Wself-attn,3
​

enc,k ≡ Wself-attn,3
dec,k , ​ ​

self-attn,3
Wself-attn,3
q = Wself-attn,3
​

enc,q ≡ Wdec,q ​
,
​

Wself-attn,3
v = Wself-attn,3
​

enc,v ≡ Wself-attn,3
dec,v , ​ ​

self-attn,3
As a result, the key projection weights Wk , Wself-attn,3
​

v , Wself-attn,3
q ​ are updated ​

twice for each backward propagation pass - once when the gradient is backpropagated through
the third decoder block and once when the gradient is backprapageted thourgh the third
encoder block.

In the same way, we can warm-start an encoder-decoder model by sharing the encoder weights
with the decoder. Being able to share the weights between the encoder and decoder requires
the decoder architecture (excluding the cross-attention weights) to be identical to the encoder
architecture. Therefore, encoder-decoder weight sharing is only relevant if the encoder-decoder
model is warm-started from a single encoder-only pre-trained checkpoint.

Great! That was the theory about warm-starting encoder-decoder models. Let's now look at
some results.

1
Without loss of generality, we exclude the normalization layers to not clutter the equations
and illustrations. 2 For more detail on how self-attention layers function, please refer to this
section of the transformer-based encoder-decoder model blog post for the encoder-part (and
this section for the decoder part respectively).

Warm-starting encoder-decoder models (Analysis)

In this section, we will summarize the findings on warm-starting encoder-decoder models as

presented in Leveraging Pre-trained Checkpoints for Sequence Generation Tasks by Sascha
Rothe, Shashi Narayan, and Aliaksei Severyn. The authors compared the performance of
warm-started encoder-decoder models to randomly initialized encoder-decoder models on
multiple sequence-to-sequence tasks, notably summarization, translation, sentence splitting, and
sentence fusion.

To be more precise, the publicly available pre-trained checkpoints of BERT, RoBERTa, and
GPT2 were leveraged in different variations to warm-start an encoder-decoder model. E.g. a
BERT-initialised encoder was paired with a BERT-initialized decoder yielding a BERT2BERT
model or a RoBERTa-initialized encoder was paired with a GPT2-initialized decoder to yield a
RoBERTa2GPT2 model. Additionally, the effect of sharing the encoder and decoder weights (as
explained in the previous section) was investigated for RoBERTa, i.e. RoBERTaShare, and for
BERT, i.e. BERTShare. Randomly or partly randomly initialized encoder-decoder models were
used as a baseline, such as a fully randomly initialized encoder-decoder model, coined
Rnd2Rnd or a BERT-initialized decoder paired with a randomly initialized encoder, defined as
Rnd2BERT.
The following table shows a complete list of all investigated model variants including the
number of randomly initialized weights, i.e. "random", and the number of weights initialized
from the respective pre-trained checkpoints, i.e. "leveraged". All models are based on a 12-
layer architecture with 768-dim hidden size embeddings, corresponding to the bert-base-cased,
bert-base-uncased, roberta-base, and gpt2 checkpoints in the 🤗Transformers model hub.
Model random leveraged total

Rnd2Rnd 221M 0 221M

Rnd2BERT 112M 109M 221M

BERT2Rnd 112M 109M 221M

Rnd2GPT2 114M 125M 238M

BERT2BERT 26M 195M 221M

BERTShare 26M 109M 135M

RoBERTaShare 26M 126M 152M

BERT2GPT2 26M 234M 260M

RoBERTa2GPT2 26M 250M 276M

The model Rnd2Rnd, which is based on the BERT2BERT architecture, contains 221M weight
parameters - all of which are randomly initialized. The other two "BERT-based" baselines
Rnd2BERT and BERT2Rnd have roughly half of their weights, i.e. 112M parameters, randomly
initialized. The other 109M weight parameters are leveraged from the pre-trained bert-base-
uncased checkpoint for the encoder- or decoder part respectively. The models BERT2BERT,
BERT2GPT2, and RoBERTa2GPT2 have all of their encoder weight parameters leveraged (from
bert-base-uncased, roberta-base respectively) and most of the decoder weight parameter
weights as well (from gpt2, bert-base-uncased respectively). 26M decoder weight parameters,
which correspond to the 12 cross-attention layers, are thereby randomly initialized.
RoBERTa2GPT2 and BERT2GPT2 are compared to the Rnd2GPT2 baseline. Also, it should be
noted that the shared model variants BERTShare and RoBERTaShare have significantly fewer
parameters because all encoder weight parameters are shared with the respective decoder
weight parameters.

Experiments

The above models were trained and evaluated on four sequence-to-sequence tasks of
increasing complexity: sentence-level fusion, sentence-level splitting, translation, and
abstractive summarization. The following table shows which datasets were used for each task.

Seq2Seq Task Datasets Paper 🤗datasets

Sentence Fusion DiscoFuse Geva et al. (2019) link

Sentence Splitting WikiSplit Botha et al. (2018) -

Translation WMT14 EN => DE Bojar et al. (2014) link

WMT14 DE => EN Bojar et al. (2014) link

Abstractive Summarizaion CNN/Dailymail Hermann et al. (2015) link

BBC XSum Narayan et al. (2018a) link

Gigaword Napoles et al. (2012) link

Depending on the task, a slightly different training regime was used. E.g. according to the size
of the dataset and the specific task, the number of training steps ranges from 200K to 500K, the
batch size is set to either 128 or 256, the input length ranges from 128 to 512 and the output
length varies between 32 to 128. It shall be emphasized however that within each task, all
models were trained and evaluated using the same hyperparameters to ensure a fair
comparison. For more information on the task-specific hyperparameter settings, the reader is
advised to see the Experiments section in the paper.

We will now give a condensed overview of the results for each task.

Sentence Fusion and -Splitting (DiscoFuse, WikiSplit)

Sentence Fusion is the task of combining multiple sentences into a single coherent sentence.
E.g. the two sentences:

As a run-blocker, Zeitler moves relatively well. Zeitler too often struggles at the point of contact in
space.

should be connected with a fitting linking word, such as:

As a run-blocker, Zeitler moves relatively well. However, he too often struggles at the point of
contact in space.

As can be seen the linking word "however" provides a coherent transition from the first
sentence to the second one. A model that is capable of generating such a linking word has
arguably learned to infer that the two sentences above contrast to each other.

The inverse task is called Sentence splitting and consists of splitting a single complex sentence
into multiple simpler ones that together retain the same meaning. Sentence splitting is
considered as an important task in text simplification, cf. to Botha et al. (2018).

As an example, the sentence:

Street Rod is the first in a series of two games released for the PC and Commodore 64 in 1989

can be simplified into

Street Rod is the first in a series of two games . It was released for the PC and Commodore 64 in
1989

It can be seen that the long sentence tries to convey two important pieces of information. One
is that the game was the first of two games being released for the PC, and the second being the
year in which it was released. Sentence splitting, therefore, requires the model to understand
which part of the sentence should be divided into two sentences, making the task more difficult
than sentence fusion.

A common metric to evaluate the performance of models on sentence fusion resp. -splitting
tasks is SARI (Wu et al. (2016), which is broadly based on the F1-score of label and model
output.

Let's see how the models perform on sentence fusion and -splitting.

Model 100% DiscoFuse (SARI) 10% DiscoFuse (SARI) 100% WikiSplit (SARI)

Rnd2Rnd 86.9 81.5 61.7

Rnd2BERT 87.6 82.1 61.8

BERT2Rnd 89.3 86.1 63.1

Rnd2GPT2 86.5 81.4 61.3

BERT2BERT 89.3 86.1 63.2

BERTShare 89.2 86.0 63.5

RoBERTaShare 89.7 86.0 63.4

BERT2GPT2 88.4 84.1 62.4

RoBERTa2GPT2 89.9 87.1 63.2

--- --- --- ---

RoBERTaShare (large) 90.3 87.7 63.8

The first two columns show the performance of the encoder-decoder models on the DiscoFuse
evaluation data. The first column states the results of encoder-decoder models trained on all
(100%) of the training data, while the second column shows the results of the models trained
only on 10% of the training data. We observe that warm-started models perform significantly
better than the randomly initialized baseline models Rnd2Rnd, Rnd2Bert, and Rnd2GPT2. A
warm-started RoBERTa2GPT2 model trained only on 10% of the training data is on par with an
Rnd2Rnd model trained on 100% of the training data. Interestingly, the Bert2Rnd baseline
performs equally well as a fully warm-started Bert2Bert model, which indicates that warm-
starting the encoder-part is more effective than warm-starting the decoder-part. The best
results are obtained by RoBERTa2GPT2, followed by RobertaShare. Sharing encoder and
decoder weight parameters does seem to slightly increase the model's performance.

On the more difficult sentence splitting task, a similar pattern emerges. Warm-started encoder-
decoder models significantly outperform encoder-decoder models whose encoder is randomly
initialized and encoder-decoder models with shared weight parameters yield better results than
those with uncoupled weight parameters. On sentence splitting the BertShare models yields the
best performance closely followed by RobertaShare.

In addition to the 12-layer model variants, the authors also trained and evaluated a 24-layer
RobertaShare (large) model which outperforms all 12-layer models significantly.

Machine Translation (WMT14)

Next, the authors evaluated warm-started encoder-decoder models on the probably most
common benchmark in machine translation (MT) - the En → De and De → En WMT14 dataset.
In this notebook, we present the results on the newstest2014 eval dataset. Because the
benchmark requires the model to understand both an English and a German vocabulary the
BERT-initialized encoder-decoder models were warm-started from the multilingual pre-trained
checkpoint bert-base-multilingual-cased. Because there is no publicly available multilingual
RoBERTa checkpoint, RoBERTa-initialized encoder-decoder models were excluded for MT.
GPT2-initialized models were initialized from the gpt2 pre-trained checkpoint as in the
previous experiment. The translation results are reported using the BLUE-4 score metric 1 .

Model En → De (BLEU-4) De → En (BLEU-4)

Rnd2Rnd 26.0 29.1

Rnd2BERT 27.2 30.4

BERT2Rnd 30.1 32.7

Rnd2GPT2 19.6 23.2

BERT2BERT 30.1 32.7

BERTShare 29.6 32.6

BERT2GPT2 23.2 31.4

--- --- ---

BERT2Rnd (large, custom) 31.7 34.2

BERTShare (large, custom) 30.5 33.8

Again, we observe a significant performance boost by warm-starting the encoder-part, with

BERT2Rnd and BERT2BERT yielding the best results on both the En → De and De → En tasks.
GPT2 initialized models perform significantly worse even than the Rnd2Rnd baseline on En →
De. Taking into consideration that the gpt2 checkpoint was trained only on English text, it is not
very surprising that BERT2GPT2 and Rnd2GPT2 models have difficulties generating German
translations. This hypothesis is supported by the competitive results (e.g. 31.4 vs. 32.7) of
BERT2GPT2 on the De → En task for which GPT2's vocabulary fits the English output format.
Contrary to the results obtained on sentence fusion and sentence splitting, sharing encoder and
decoder weight parameters does not yield a performance boost in MT. Possible reasons for this
as stated by the authors include

the encoder-decoder model capacity is an important factor in MT, and

the encoder and decoder have to deal with different grammar and vocabulary

Since the bert-base-multilingual-cased checkpoint was trained on more than 100 languages, its
vocabulary is probably undesirably large for En → De and De → En MT. Thus, the authors pre-
trained a large BERT encoder-only checkpoint on the English and German subset of the
Wikipedia dump and subsequently used it to warm-start a BERT2Rnd and BERTShare encoder-
decoder model. Thanks to the improved vocabulary, another significant performance boost is
observed, with BERT2Rnd (large, custom) significantly outperforming all other models.

Summarization (CNN/Dailymail, BBC XSum, Gigaword)

Finally, the encoder-decoder models were evaluated on the arguably most challenging
sequence-to-sequence task - summarization. The authors picked three summarization datasets
with different characteristics for evaluation: Gigaword (headline generation), BBC XSum
(extreme summarization), and CNN/Dailymayl (abstractive summarization).

The Gigaword dataset contains sentence-level abstractive summarizations, requiring the model
to learn sentence-level understanding, abstraction, and eventually paraphrasing. A typical data
sample in Gigaword, such as

"*venezuelan president hugo chavez said thursday he has ordered a probe into a suspected
coup plot allegedly involving active and retired military officers .*",

would have a corresponding headline as its label, e.g.:

"chavez orders probe into suspected coup plot".

The BBC XSum dataset consists of much longer article-like text inputs with the labels being
mostly single sentence summarizations. This dataset requires the model not only to learn
document-level inference but also a high level of abstractive paraphrasing. Some data samples
of the BBC XSUM datasets are shown here.
For the CNN/Dailmail dataset, documents, which are of similar length than those in the BBC
XSum dataset, have to be summarized to bullet-point story highlights. The labels therefore
often consist of multiple sentences. Besides document-level understanding, the CNN/Dailymail
dataset requires models to be good at copying the most salient information. Some examples
can be viewed here.

The models are evaluated using the Rouge metric, whereas the Rouge-2 scores are shown
below.

Alright, let's take a look at the results.

Model CNN/Dailymail (Rouge-2) BBC XSum (Rouge-2) Gigaword (Rouge-2)

Rnd2Rnd 14.00 10.23 18.71

Rnd2BERT 15.55 11.52 18.91

BERT2Rnd 17.76 15.83 19.26

Rnd2GPT2 8.81 8.77 18.39

BERT2BERT 17.84 15.24 19.68

BERTShare 18.10 16.12 19.81

RoBERTaShare 18.95 17.50 19.70

BERT2GPT2 4.96 8.37 18.23

RoBERTa2GPT2 14.72 5.20 19.21

--- --- --- ---

RoBERTaShare (large) 18.91 18.79 19.78

We observe again that warm-starting the encoder-part gives a significant improvement over
models with randomly-initialized encoders, which is especially visible for document-level
abstraction tasks, i.e. CNN/Dailymail and BBC XSum. This shows that tasks requiring a high
level of abstraction benefit more from a pre-trained encoder part than those requiring only
sentence-level abstraction. Except for Gigaword GPT2-based encoder-decoder models seem to
be unfit for summarization.

Furthermore, the shared encoder-decoder models are the best performing models for
summarization. RoBERTaShare and BERTShare are the best performing models on all datasets
whereas the margin is especially significant on the BBC XSum dataset on which RoBERTaShare
(large) outperforms BERT2BERT and BERT2Rnd by ca. 3 Rouge-2 points and Rnd2Rnd by more
than 8 Rouge-2 points. As brought forward by the authors, "this is probably because the BBC
summary sentences follow a distribution that is similar to that of the sentences in the document,
whereas this is not necessarily the case for the Gigaword headlines and the CNN/DailyMail
bullet-point highlights". Intuitively this means that in BBC XSum, the input sentences processed
by the encoder are very similar in structure to the single sentence summary processed by the
decoder, i.e. same length, similar choice of words, similar syntax.

Conclusion

Alright, let's draw a conclusion and try to derive some practical tips.

We have observed on all tasks that a warm-started encoder-part gives a significant

performance boost compared to encoder-decoder models having a randomly initialized
encoder. On the other hand, warm-starting the decoder seems to be less important, with
 BERT2BERT being on par with BERT2Rnd on most tasks. An intuitive reason would be that
since a BERT- or RoBERTa-initialized encoder part has none of its weight parameters
randomly initialized, the encoder can fully exploit the acquired knowledge of BERT's or
RoBERTa's pre-trained checkpoints, respectively. In contrast, the warm-started decoder
always has parts of its weight parameters randomly initialized which possibly makes it
much harder to effectively leverage the knowledge acquired by the checkpoint used to
initialize the decoder.

Next, we noticed that it is often beneficial to share encoder and decoder weights,
especially if the target distribution is similar to the input distribution (e.g. BBC XSum).
However, for datasets whose target data distribution differs more significantly from the
input data distribution and for which model capacity 2 is known to play an important role,
e.g. WMT14, encoder-decoder weight sharing seems to be disadvantageous.

Finally, we have seen that it is very important that the vocabulary of the pre-trained
"stand-alone" checkpoints fit the vocabulary required to solve the sequence-to-sequence
task. E.g. a warm-started BERT2GPT2 encoder-decoder will perform poorly on En → De
MT because GPT2 was pre-trained on English whereas the target language is German. The
overall poor performance of the BERT2GPT2, Rnd2GPT2, and RoBERTa2GPT2 compared
to BERT2BERT, BERTShared, and RoBERTaShared suggests that it is more effective to have
a shared vocabulary. Also, it shows that initializing the decoder part with a pre-trained
GPT2 checkpoint is not more effective than initializing it with a pre-trained BERT
checkpoint besides GPT2 being more similar to the decoder in its architecture.

For each of the above tasks, the most performant models were ported to 🤗Transformers and
can be accessed here:

RoBERTaShared (large) - Wikisplit: google/roberta2roberta_L-24_wikisplit.

RoBERTaShared (large) - Discofuse: google/roberta2roberta_L-24_discofuse.

BERT2BERT (large) - WMT en → de: google/bert2bert_L-24_wmt_en_de.

BERT2BERT (large) - WMT de → en: google/bert2bert_L-24_wmt_de_en.

RoBERTaShared (large) - CNN/Dailymail: google/roberta2roberta_L-24_cnn_daily_mail.

RoBERTaShared (large) - BBC XSum: google/roberta2roberta_L-24_bbc.

RoBERTaShared (large) - Gigaword: google/roberta2roberta_L-24_gigaword.

1
To retrieve BLEU-4 scores, a script from the Tensorflow Official Transformer implementation
https://github.com/tensorflow/models/tree master/official/nlp/transformer was used. Note
that, differently from the tensor2tensor/utils/ get_ende_bleu.sh used by Vaswani et al. (2017),
this script does not split noun compounds, but utf-8 quotes were normalized to ascii quotes
after having noted that the pre-processed training set contains only ascii quotes.

2
Model capacity is an informal definition of how good the model is at modeling complex
patterns. It is also sometimes defined as the ability of a model to learn from more and more
data. Model capacity is broadly measured by the number of trainable parameters - the more
parameters, the higher the model capacity.

Warm-starting encoder-decoder
models with Transformers
(Practice)
🤗
We have explained the theory of warm-starting encoder-decoder models, analyzed empirical
results on multiple datasets, and have derived practical conclusions. Let's now walk through a
complete code example showcasing how a BERT2BERT model can be warm-started and
consequently fine-tuned on the CNN/Dailymail summarization task. We will be leveraging the
🤗datasets and 🤗Transformers libraries.
In addition, the following list provides a condensed version of this and other notebooks on
warm-starting other combinations of encoder-decoder models.

for BERT2BERT on CNN/Dailymail (a condensed version of this notebook), click here.

for RoBERTaShare on BBC XSum, click here.

for BERT2Rnd on WMT14 En → De, click here.

for RoBERTa2GPT2 on DiscoFuse, click here.

Note: This notebook only uses a few training, validation, and test data samples for
demonstration purposes. To fine-tune an encoder-decoder model on the full training data, the
user should change the training and data preprocessing parameters accordingly as highlighted
by the comments.

Data Preprocessing

In this section, we show how the data can be pre-processed for training. More importantly, we
try to give the reader some insight into the process of deciding how to preprocess the data.

We will need datasets and transformers to be installed.

!pip install datasets==1.0.2

!pip install transformers==4.2.1

Let's start by downloading the CNN/Dailymail dataset.

import datasets

train_data = datasets.load_dataset("cnn_dailymail", "3.0.0", split="train")

Alright, let's get a first impression of the dataset. Alternatively, the dataset can also be
visualized using the awesome datasets viewer online.

train_data.info.description

Our input is called article and our labels are called highlights. Let's now print out the first
example of the training data to get a feeling for the data.

import pandas as pd

from IPython.display import display, HTML

from datasets import ClassLabel

df = pd.DataFrame(train_data[:1])
del df["id"]

for column, typ in train_data.features.items():

if isinstance(typ, ClassLabel):

df[column] = df[column].transform(lambda i: typ.names[i])

display(HTML(df.to_html()))
 OUTPUT:
-------
Article:
"""It's official: U.S. President Barack Obama wants lawmakers to weigh in on whether to u
Summary:
"""Syrian official: Obama climbed to the top of the tree, "doesn't know how to get down"\

The input data seems to consist of short news articles. Interestingly, the labels appear to be
bullet-point-like summaries. At this point, one should probably take a look at a couple of other
examples to get a better feeling for the data.

One should also notice here that the text is case-sensitive. This means that we have to be careful
if we want to use case-insensitive models. As CNN/Dailymail is a summarization dataset, the
model will be evaluated using the ROUGE metric. Checking the description of ROUGE in 🤗
datasets, cf. here, we can see that the metric is case-insensitive, meaning that upper case letters
will be normalized to lower case letters during evaluation. Thus, we can safely leverage uncased
checkpoints, such as bert-base-uncased.

Cool! Next, let's get a sense of the length of input data and labels.

As models compute length in token-length, we will make use of the bert-base-uncased tokenizer
to compute the article and summary length.

First, we load the tokenizer.

from transformers import BertTokenizerFast

tokenizer = BertTokenizerFast.from_pretrained("bert-base-uncased")

Next, we make use of .map() to compute the length of the article and its summary. Since we
know that the maximum length that bert-base-uncased can process amounts to 512, we are also
interested in the percentage of input samples being longer than the maximum length. Similarly,
we compute the percentage of summaries that are longer than 64, and 128 respectively.

We can define the .map() function as follows.

# map article and summary len to dict as well as if sample is longer than 512 tokens
def map_to_length(x):

x["article_len"] = len(tokenizer(x["article"]).input_ids)
x["article_longer_512"] = int(x["article_len"] > 512)

x["summary_len"] = len(tokenizer(x["highlights"]).input_ids)

x["summary_longer_64"] = int(x["summary_len"] > 64)

x["summary_longer_128"] = int(x["summary_len"] > 128)

return x

It should be sufficient to look at the first 10000 samples. We can speed up the mapping by using
multiple processes with num_proc=4.

sample_size = 10000

data_stats = train_data.select(range(sample_size)).map(map_to_length, num_proc=4)

Having computed the length for the first 10000 samples, we should now average them
together. For this, we can make use of the .map() function with batched=True and batch_size=-1
to have access to all 10000 samples within the .map() function.
 def compute_and_print_stats(x):
if len(x["article_len"]) == sample_size:
print(
"Article Mean: {}, %-Articles > 512:{}, Summary Mean:{}, %-Summary > 64:{}, %-Sum
sum(x["article_len"]) / sample_size,
sum(x["article_longer_512"]) / sample_size,
sum(x["summary_len"]) / sample_size,

sum(x["summary_longer_64"]) / sample_size,
sum(x["summary_longer_128"]) / sample_size,
)
)

output = data_stats.map(
compute_and_print_stats,
batched=True,
batch_size=-1,
)

OUTPUT:
-------

Article Mean: 847.6216, %-Articles > 512:0.7355, Summary Mean:57.7742, %-Summary > 64

We can see that on average an article contains 848 tokens with ca. 3/4 of the articles being
longer than the model's max_length 512. The summary is on average 57 tokens long. Over 30%
of our 10000-sample summaries are longer than 64 tokens, but none are longer than 128
tokens.

bert-base-cased is limited to 512 tokens, which means we would have to cut possibly important
information from the article. Because most of the important information is often found at the
beginning of articles and because we want to be computationally efficient, we decide to stick to
bert-base-cased with a max_length of 512 in this notebook. This choice is not optimal but has
shown to yield good results on CNN/Dailymail. Alternatively, one could leverage long-range
sequence models, such as Longformer to be used as the encoder.

Regarding the summary length, we can see that a length of 128 already includes all of the
summary labels. 128 is easily within the limits of bert-base-cased, so we decide to limit the
generation to 128.

Again, we will make use of the .map() function - this time to transform each training batch into
a batch of model inputs.

"article" and "highlights" are tokenized and prepared as the Encoder's "input_ids" and

Decoder's "decoder_input_ids" respectively.

"labels" are shifted automatically to the left for language modeling training.

Lastly, it is very important to remember to ignore the loss of the padded labels. In 🤗
Transformers this can be done by setting the label to -100. Great, let's write down our mapping
function then.

encoder_max_length=512
decoder_max_length=128

def process_data_to_model_inputs(batch):
# tokenize the inputs and labels
inputs = tokenizer(batch["article"], padding="max_length", truncation=True, max_length=
outputs = tokenizer(batch["highlights"], padding="max_length", truncation=True, max_len
 batch["input_ids"] = inputs.input_ids
batch["attention_mask"] = inputs.attention_mask
batch["labels"] = outputs.input_ids.copy()

# because BERT automatically shifts the labels, the labels correspond exactly to `decod
# We have to make sure that the PAD token is ignored
batch["labels"] = [[-100 if token == tokenizer.pad_token_id else token for token in lab

return batch

In this notebook, we train and evaluate the model just on a few training examples for
demonstration and set the batch_size to 4 to prevent out-of-memory issues.

The following line reduces the training data to only the first 32 examples. The cell can be
commented out or not run for a full training run. Good results were obtained with a batch_size
of 16.

train_data = train_data.select(range(32))

Alright, let's prepare the training data.

# batch_size = 16
batch_size=4

train_data = train_data.map(
process_data_to_model_inputs,

batched=True,

batch_size=batch_size,

remove_columns=["article", "highlights", "id"]

Taking a look at the processed training dataset we can see that the column names article,
highlights, and id have been replaced by the arguments expected by the EncoderDecoderModel.

train_data

OUTPUT:
-------
Dataset(features: {'attention_mask': Sequence(feature=Value(dtype='int64', id=None), leng

So far, the data was manipulated using Python's List format. Let's convert the data to PyTorch
Tensors to be trained on GPU.

train_data.set_format(

type="torch", columns=["input_ids", "attention_mask", "labels"],

Awesome, the data processing of the training data is finished. Analogous, we can do the same
for the validation data.

First, we load 10% of the validation dataset:

val_data = datasets.load_dataset("cnn_dailymail", "3.0.0", split="validation[:10%]")