Tutorial
Yen-Chieh (David) Liao | Sohini Timbadia | Stefan Müller
University of Birmingham & University College DublinSource:
vignettes/tutorial.Rmd
tutorial.Rmd
The Overview
Flair NLP is an open-source library for Natural Language Processing (NLP) developed by Zalando Research. Known for its state-of-the-art solutions, such as contextual string embeddings for NLP tasks like Named Entity Recognition (NER), Part-of-Speech tagging (POS), and more, it has garnered the attention of the NLP community for its ease of use and powerful functionalities.
In addition, Flair NLP offers pre-trained models for various languages and tasks, and is compatible with fine-tuned transformers hosted on Hugging Face.
Sentence and Token
Sentence and Token are fundamental classes.
Sentence
A Sentence in Flair is an object that contains a sequence of Token objects, and it can be annotated with labels, such as named entities, part-of-speech tags, and more. It also can store embeddings for the sentence as a whole and different kinds of linguistic annotations.
Here’s a simple example of how you create a Sentence:
# Creating a Sentence object
library(flaiR)
string <- "What I see in UCD today, what I have seen of UCD in its impact on my own life and the life of Ireland."
Sentence <- flair_data()$Sentence
sentence <- Sentence(string)
Sentence[26]
means that there are a total of 26 tokens
in the sentence.
print(sentence)
#> Sentence[26]: "What I see in UCD today, what I have seen of UCD in its impact on my own life and the life of Ireland."
Token
When you use Flair to handle text data,1 Sentence
and Token
objects often play central roles in many use
cases. When you create a Sentence object, it automatically tokenizes the
text, removing the need to create the Token object manually.
Unlike R, which indexes from 1, Python indexes from 0. Therefore,
when using a for loop, I use seq_along(sentence) - 1
. The
output should be something like:
# The Sentence object has automatically created and contains multiple Token objects
# We can iterate through the Sentence object to view each Token
for (i in seq_along(sentence)-1) {
print(sentence[[i]])
}
#> Token[0]: "What"
#> Token[1]: "I"
#> Token[2]: "see"
#> Token[3]: "in"
#> Token[4]: "UCD"
#> Token[5]: "today"
#> Token[6]: ","
#> Token[7]: "what"
#> Token[8]: "I"
#> Token[9]: "have"
#> Token[10]: "seen"
#> Token[11]: "of"
#> Token[12]: "UCD"
#> Token[13]: "in"
#> Token[14]: "its"
#> Token[15]: "impact"
#> Token[16]: "on"
#> Token[17]: "my"
#> Token[18]: "own"
#> Token[19]: "life"
#> Token[20]: "and"
#> Token[21]: "the"
#> Token[22]: "life"
#> Token[23]: "of"
#> Token[24]: "Ireland"
#> Token[25]: "."
Or you can directly use $tokens
method to print all
tokens.
print(sentence$tokens)
#> [[1]]
#> Token[0]: "What"
#>
#> [[2]]
#> Token[1]: "I"
#>
#> [[3]]
#> Token[2]: "see"
#>
#> [[4]]
#> Token[3]: "in"
#>
#> [[5]]
#> Token[4]: "UCD"
#>
#> [[6]]
#> Token[5]: "today"
#>
#> [[7]]
#> Token[6]: ","
#>
#> [[8]]
#> Token[7]: "what"
#>
#> [[9]]
#> Token[8]: "I"
#>
#> [[10]]
#> Token[9]: "have"
#>
#> [[11]]
#> Token[10]: "seen"
#>
#> [[12]]
#> Token[11]: "of"
#>
#> [[13]]
#> Token[12]: "UCD"
#>
#> [[14]]
#> Token[13]: "in"
#>
#> [[15]]
#> Token[14]: "its"
#>
#> [[16]]
#> Token[15]: "impact"
#>
#> [[17]]
#> Token[16]: "on"
#>
#> [[18]]
#> Token[17]: "my"
#>
#> [[19]]
#> Token[18]: "own"
#>
#> [[20]]
#> Token[19]: "life"
#>
#> [[21]]
#> Token[20]: "and"
#>
#> [[22]]
#> Token[21]: "the"
#>
#> [[23]]
#> Token[22]: "life"
#>
#> [[24]]
#> Token[23]: "of"
#>
#> [[25]]
#> Token[24]: "Ireland"
#>
#> [[26]]
#> Token[25]: "."
Retrieve the Token
To comprehend the string representation format of the Sentence
object, tagging at least one token is adequate. Python’s
get_token(n)
method allows us to retrieve the Token object
for a particular token. Additionally, we can use
[]
to index a specific token.
# method in Python
sentence$get_token(5)
#> Token[4]: "UCD"
# indexing in R
sentence[6]
#> Token[6]: ","
Each word (and punctuation) in the text is treated as an individual Token object. These Token objects store text information and other possible linguistic information (such as part-of-speech tags or named entity tags) and embeddings (if you used a model to generate them).
While you do not need to create Token objects manually, understanding how to manage them is useful in situations where you might want to fine-tune the tokenization process. For example, you can control the exactness of tokenization by manually creating Token objects from a Sentence object.
This makes Flair very flexible when handling text data since the automatic tokenization feature can be used for rapid development, while also allowing users to fine-tune their tokenization.
Annotate POS tag and NER tag
The add_label(label_type, value)
method can be employed
to assign a label to the token. In Universal POS tags, if
sentence[10]
is ‘see’, ‘seen’ might be tagged as
VERB
, indicating it is a past participle form of a
verb.
sentence[10]$add_label('manual-pos', 'VERB')
print(sentence[10])
#> Token[10]: "seen" → VERB (1.0000)
We can also add a NER (Named Entity Recognition) tag to
sentence[4]
, “UCD”, identifying it as a university in
Dublin.
sentence[4]$add_label('ner', 'ORG')
print(sentence[4])
#> Token[4]: "UCD" → ORG (1.0000)
If we print the sentence object, Sentence[50]
provides
information for 50 tokens → [‘in’/ORG, ‘seen’/VERB], thus displaying two
tagging pieces of information.
print(sentence)
#> Sentence[26]: "What I see in UCD today, what I have seen of UCD in its impact on my own life and the life of Ireland." → ["UCD"/ORG, "seen"/VERB]
Corpus
The Corpus object in Flair is a fundamental data structure that represents a dataset containing text samples, usually comprising of a training set, a development set (or validation set), and a test set. It’s designed to work smoothly with Flair’s models for tasks like named entity recognition, text classification, and more.
Attributes:
-
train
: A list of sentences (List[Sentence]) that form the training dataset. -
dev
(or development): A list of sentences (List[Sentence]) that form the development (or validation) dataset. -
test
: A list of sentences (List[Sentence]) that form the test dataset.
Important Methods:
-
downsample
: This method allows you to downsample (reduce) the number of sentences in the train, dev, and test splits. -
obtain_statistics
: This method gives a quick overview of the statistics of the corpus, including the number of sentences and the distribution of labels. -
make_vocab_dictionary
: Used to create a vocabulary dictionary from the corpus.
library(flaiR)
Corpus <- flair_data()$Corpus
Sentence <- flair_data()$Sentence
# Create some example sentences
train <- list(Sentence('This is a training example.'))
dev <- list(Sentence('This is a validation example.'))
test <- list(Sentence('This is a test example.'))
# Create a corpus using the custom data splits
corp <- Corpus(train = train, dev = dev, test = test)
$obtain_statistics()
method of the Corpus object in the
Flair library provides an overview of the dataset statistics. The method
returns a Python
dictionary with details about the training, validation
(development), and test datasets that make up the corpus. In R, you can
use the jsonlite package to format JSON.
library(jsonlite)
data <- fromJSON(corp$obtain_statistics())
formatted_str <- toJSON(data, pretty=TRUE)
print(formatted_str)
#> {
#> "TRAIN": {
#> "dataset": ["TRAIN"],
#> "total_number_of_documents": [1],
#> "number_of_documents_per_class": {},
#> "number_of_tokens_per_tag": {},
#> "number_of_tokens": {
#> "total": [6],
#> "min": [6],
#> "max": [6],
#> "avg": [6]
#> }
#> },
#> "TEST": {
#> "dataset": ["TEST"],
#> "total_number_of_documents": [1],
#> "number_of_documents_per_class": {},
#> "number_of_tokens_per_tag": {},
#> "number_of_tokens": {
#> "total": [6],
#> "min": [6],
#> "max": [6],
#> "avg": [6]
#> }
#> },
#> "DEV": {
#> "dataset": ["DEV"],
#> "total_number_of_documents": [1],
#> "number_of_documents_per_class": {},
#> "number_of_tokens_per_tag": {},
#> "number_of_tokens": {
#> "total": [6],
#> "min": [6],
#> "max": [6],
#> "avg": [6]
#> }
#> }
#> }
In R
Below, we use data from the article The Temporal Focus of Campaign Communication by Stefan Muller, published in the Journal of Politics in 2020, as an example.
First, we vectorize the cc_muller$text
using the
Sentence function to transform it into a list object. Then, we reformat
cc_muller$class_pro_retro
as a factor. It’s essential to
note that R handles numerical values differently than Python. In R,
numerical values are represented with a floating point, so it’s
advisable to convert them into factors or strings. Lastly, we employ the
map function from the purrr package to assign labels to each sentence
corpus using the $add_label
method.
library(purrr)
#>
#> Attaching package: 'purrr'
#> The following object is masked from 'package:jsonlite':
#>
#> flatten
data(cc_muller)
# The `Sentence` object tokenizes text
text <- lapply( cc_muller$text, Sentence)
# split sentence object to train and test.
labels <- as.factor(cc_muller$class_pro_retro)
# `$add_label` method assigns the corresponding coded type to each Sentence corpus.
text <- map2(text, labels, ~ .x$add_label("classification", .y), .progress = TRUE)
To perform a train-test split using base R, we can follow these steps:
set.seed(2046)
sample <- sample(c(TRUE, FALSE), length(text), replace=TRUE, prob=c(0.8, 0.2))
train <- text[sample]
test <- text[!sample]
sprintf("Corpus object sizes - Train: %d | Test: %d", length(train), length(test))
#> [1] "Corpus object sizes - Train: 4710 | Test: 1148"
If you don’t provide a dev set, flaiR will not force you to carve out a portion of your test set to serve as a dev set. However, in some cases when only the train and test sets are provided without a dev set, flaiR might automatically take a fraction of the train set (e.g., 10%) to use as a dev set (#2259). This is to offer a mechanism for model selection and to prevent the model from overfitting on the train set.
In the “Corpus” function, there is a random selection of the “dev”
dataset. To ensure reproducibility, we need to set the seed in the flaiR
framework. We can accomplish this by calling the top-level module
“flair” from {flaiR} and using $set_seed(1964L)
to set the
seed.
flair <- import_flair()
flair$set_seed(1964L)
corp <- Corpus(train=train,
# dev=test,
test=test)
#> 2024-09-23 11:46:44,255 No dev split found. Using 10% (i.e. 471 samples) of the train split as dev data
sprintf("Corpus object sizes - Train: %d | Test: %d | Dev: %d",
length(corp$train),
length(corp$test),
length(corp$dev))
#> [1] "Corpus object sizes - Train: 4239 | Test: 1148 | Dev: 471"
In the later sections, there will be more similar processing using
the Corpus
. Following that, we will focus on advanced NLP
applications.
Sequence Taggings
Tag Entities in Text
Let’s run named entity recognition over the following example sentence: “I love Berlin and New York”. To do this, all you need to do is make a Sentence object for this text, load a pre-trained model and use it to predict tags for the object.
# attach flaiR in R
library(flaiR)
# make a sentence
Sentence <- flair_data()$Sentence
sentence <- Sentence('I love Berlin and New York.')
# load the NER tagger
Classifier <- flair_nn()$Classifier
tagger <- Classifier$load('ner')
#> 2024-09-23 11:46:45,580 SequenceTagger predicts: Dictionary with 20 tags: <unk>, O, S-ORG, S-MISC, B-PER, E-PER, S-LOC, B-ORG, E-ORG, I-PER, S-PER, B-MISC, I-MISC, E-MISC, I-ORG, B-LOC, E-LOC, I-LOC, <START>, <STOP>
# run NER over sentence
tagger$predict(sentence)
To print all annotations:
# print the sentence with all annotations
print(sentence)
#> Sentence[7]: "I love Berlin and New York." → ["Berlin"/LOC, "New York"/LOC]
Use a for loop to print out each POS tag. It’s important to note that
Python is indexed from 0. Therefore, in an R environment, we must use
seq_along(sentence$get_labels()) - 1
.
Tag Part-of-Speech in Text
We use flaiR/POS-english for POS tagging in the standard models on Hugging Face.
# attach flaiR in R
library(flaiR)
# make a sentence
Sentence <- flair_data()$Sentence
sentence <- Sentence('I love Berlin and New York.')
# load the NER tagger
Classifier <- flair_nn()$Classifier
tagger <- Classifier$load('pos')
#> 2024-09-23 11:46:46,849 SequenceTagger predicts: Dictionary with 53 tags: <unk>, O, UH, ,, VBD, PRP, VB, PRP$, NN, RB, ., DT, JJ, VBP, VBG, IN, CD, NNS, NNP, WRB, VBZ, WDT, CC, TO, MD, VBN, WP, :, RP, EX, JJR, FW, XX, HYPH, POS, RBR, JJS, PDT, NNPS, RBS, AFX, WP$, -LRB-, -RRB-, ``, '', LS, $, SYM, ADD
# run NER over sentence
tagger$predict(sentence)
To print all annotations:
# print the sentence with all annotations
print(sentence)
#> Sentence[7]: "I love Berlin and New York." → ["I"/PRP, "love"/VBP, "Berlin"/NNP, "and"/CC, "New"/NNP, "York"/NNP, "."/.]
Use a for loop to print out each POS tag.
for (i in seq_along(sentence$get_labels())) {
print(sentence$get_labels()[[i]])
}
#> 'Token[0]: "I"'/'PRP' (1.0)
#> 'Token[1]: "love"'/'VBP' (1.0)
#> 'Token[2]: "Berlin"'/'NNP' (0.9999)
#> 'Token[3]: "and"'/'CC' (1.0)
#> 'Token[4]: "New"'/'NNP' (1.0)
#> 'Token[5]: "York"'/'NNP' (1.0)
#> 'Token[6]: "."'/'.' (1.0)
Detect Sentiment
Let’s run sentiment analysis over the same sentence to determine whether it is POSITIVE or NEGATIVE.
You can do this with essentially the same code as above. Instead of loading the ‘ner’ model, you now load the ‘sentiment’ model:
# attach flaiR in R
library(flaiR)
# make a sentence
Sentence <- flair_data()$Sentence
sentence <- Sentence('I love Berlin and New York.')
# load the Classifier tagger from flair.nn module
Classifier <- flair_nn()$Classifier
tagger <- Classifier$load('sentiment')
# run sentiment analysis over sentence
tagger$predict(sentence)
# print the sentence with all annotations
print(sentence)
#> Sentence[7]: "I love Berlin and New York." → POSITIVE (0.9982)
Tagging Parts-of-Speech with Flair Models
You can load the pre-trained model "pos-fast"
. For more
pre-trained models, see https://flairnlp.github.io/docs/tutorial-basics/part-of-speech-tagging#-in-english.
texts <- c("UCD is one of the best universities in Ireland.",
"UCD has a good campus but is very far from my apartment in Dublin.",
"Essex is famous for social science research.",
"Essex is not in the Russell Group, but it is famous for political science research and in 1994 Group.",
"TCD is the oldest university in Ireland.",
"TCD is similar to Oxford.")
doc_ids <- c("doc1", "doc2", "doc3", "doc4", "doc5", "doc6")
tagger_pos <- load_tagger_pos("pos-fast")
#> 2024-09-23 11:46:50,325 SequenceTagger predicts: Dictionary with 53 tags: <unk>, O, UH, ,, VBD, PRP, VB, PRP$, NN, RB, ., DT, JJ, VBP, VBG, IN, CD, NNS, NNP, WRB, VBZ, WDT, CC, TO, MD, VBN, WP, :, RP, EX, JJR, FW, XX, HYPH, POS, RBR, JJS, PDT, NNPS, RBS, AFX, WP$, -LRB-, -RRB-, ``, '', LS, $, SYM, ADD
results <- get_pos(texts, doc_ids, tagger_pos)
head(results, n = 10)
#> doc_id token_id text_id token tag precision
#> <char> <num> <lgcl> <char> <char> <num>
#> 1: doc1 0 NA UCD NNP 0.9967
#> 2: doc1 1 NA is VBZ 1.0000
#> 3: doc1 2 NA one CD 0.9993
#> 4: doc1 3 NA of IN 1.0000
#> 5: doc1 4 NA the DT 1.0000
#> 6: doc1 5 NA best JJS 0.9988
#> 7: doc1 6 NA universities NNS 0.9997
#> 8: doc1 7 NA in IN 1.0000
#> 9: doc1 8 NA Ireland NNP 1.0000
#> 10: doc1 9 NA . . 0.9998
Tagging Entities with Flair Models
Load the pre-trained model ner
. For more pre-trained
models, see https://flairnlp.github.io/docs/tutorial-basics/tagging-entities.
tagger_ner <- load_tagger_ner("ner")
#> 2024-09-23 11:46:51,281 SequenceTagger predicts: Dictionary with 20 tags: <unk>, O, S-ORG, S-MISC, B-PER, E-PER, S-LOC, B-ORG, E-ORG, I-PER, S-PER, B-MISC, I-MISC, E-MISC, I-ORG, B-LOC, E-LOC, I-LOC, <START>, <STOP>
results <- get_entities(texts, doc_ids, tagger_ner)
head(results, n = 10)
#> doc_id entity tag
#> <char> <char> <char>
#> 1: doc1 UCD ORG
#> 2: doc1 Ireland LOC
#> 3: doc2 UCD ORG
#> 4: doc2 Dublin LOC
#> 5: doc3 Essex ORG
#> 6: doc4 Essex ORG
#> 7: doc4 Russell Group ORG
#> 8: doc5 TCD ORG
#> 9: doc5 Ireland LOC
#> 10: doc6 TCD ORG
Tagging Sentiment
Load the pre-trained model “sentiment
”. The pre-trained
models of “sentiment
”, “sentiment-fast
”, and
“de-offensive-language
” are currently available. For more
pre-trained models, see https://flairnlp.github.io/docs/tutorial-basics/tagging-sentiment.
tagger_sent <- load_tagger_sentiments("sentiment")
results <- get_sentiments(texts, doc_ids, tagger_sent)
head(results, n = 10)
#> doc_id sentiment score
#> <char> <char> <num>
#> 1: doc1 POSITIVE 0.9970598
#> 2: doc2 NEGATIVE 0.8472342
#> 3: doc3 POSITIVE 0.9928006
#> 4: doc4 POSITIVE 0.9901404
#> 5: doc5 POSITIVE 0.9952670
#> 6: doc6 POSITIVE 0.9291791
Flair Embedding
Flair is a very popular natural language processing library, providing a variety of embedding methods for text representation. Flair Embeddings is a word embedding framework developed by Zalando. It focuses on word-level representation and can capture contextual information of words, allowing the same word to have different embeddings in different contexts. Unlike traditional word embeddings (such as Word2Vec or GloVe), Flair can dynamically generate word embeddings based on context and has achieved excellent results in various NLP tasks. Below are some key points about Flair Embeddings:
Context-Aware
Flair is a dynamic word embedding technique that can understand the meaning of words based on context. In contrast, static word embeddings, such as Word2Vec or GloVe, provide a fixed embedding for each word without considering its context in a sentence.
Therefore, context-sensitive embedding techniques, such as Flair, can capture the meaning of words in specific sentences more accurately, thus enhancing the performance of language models in various tasks.
Example:
Consider the following two English sentences:
- “I am interested in the bank of the river.”
- “I need to go to the bank to withdraw money.”
Here, the word “bank” has two different meanings. In the first sentence, it refers to the edge or shore of a river. In the second sentence, it refers to a financial institution.
For static embeddings, the word “bank” might have an embedding that lies somewhere between these two meanings because it doesn’t consider context. But for dynamic embeddings like Flair, “bank” in the first sentence will have an embedding related to rivers, and in the second sentence, it will have an embedding related to finance.
FlairEmbeddings <- flair_embeddings()$FlairEmbeddings
Sentence <- flair_data()$Sentence
# Initialize Flair embeddings
flair_embedding_forward <- FlairEmbeddings('news-forward')
# Define the two sentences
sentence1 <- Sentence("I am interested in the bank of the river.")
sentence2 <- Sentence("I need to go to the bank to withdraw money.")
# Get the embeddings
flair_embedding_forward$embed(sentence1)
#> [[1]]
#> Sentence[10]: "I am interested in the bank of the river."
flair_embedding_forward$embed(sentence2)
#> [[1]]
#> Sentence[11]: "I need to go to the bank to withdraw money."
# Extract the embedding for "bank" from the sentences
bank_embedding_sentence1 = sentence1[5]$embedding # "bank" is the seventh word
bank_embedding_sentence2 = sentence2[6]$embedding # "bank" is the sixth word
Same word, similar vector representation, but essentially different. In this way, you can see how the dynamic embeddings for “bank” in the two sentences differ based on context. Although we printed the embeddings here, in reality, they would be high-dimensional vectors, so you might see a lot of numbers. If you want a more intuitive view of the differences, you could compute the cosine similarity or other metrics between the two embeddings.
This is just a simple demonstration. In practice, you can also
combine multiple embedding techniques, such as
WordEmbeddings
and FlairEmbeddings
, to get
richer word vectors.
library(lsa)
#> Loading required package: SnowballC
cosine(as.numeric( bank_embedding_sentence1$numpy()),
as.numeric( bank_embedding_sentence2$numpy()))
#> [,1]
#> [1,] 0.7329551
Character-Based
Flair uses a character-level language model, meaning it can generate embeddings for rare words or even misspelled words. This is an important feature because it allows the model to understand and process words that have never appeared in the training data. Flair uses a bidirectional LSTM (Long Short-Term Memory) network that operates at a character level. This allows it to feed individual characters into the LSTM instead of words.
Multilingual Support
Flair provides various pre-trained character-level language models, supporting contextual word embeddings for multiple languages. It allows you to easily combine different word embeddings (e.g., Flair Embeddings, Word2Vec, GloVe, etc.) to create powerful stacked embeddings.
Classic Wordembeddings
In Flair, the simplest form of embeddings that still contains semantic information about the word are called classic word embeddings. These embeddings are pre-trained and non-contextual.
Let’s retrieve a few word embeddings and use FastText embeddings with
the following code. To do so, we simply instantiate a WordEmbeddings
class by passing in the ID of the embedding of our choice. Then, we
simply wrap our text into a Sentence object, and call the
embed(sentence)
method on our WordEmbeddings class.
WordEmbeddings <- flair_embeddings()$WordEmbeddings
Sentence <- flair_data()$Sentence
embedding <- WordEmbeddings('crawl')
sentence <- Sentence("one two three one")
embedding$embed(sentence)
#> [[1]]
#> Sentence[4]: "one two three one"
for (i in seq_along(sentence$tokens)) {
print(head(sentence$tokens[[i]]$embedding), n =5)
}
#> tensor([-0.0535, -0.0368, -0.2851, -0.0381, -0.0486, 0.2383])
#> tensor([ 0.0282, -0.0786, -0.1236, 0.1756, -0.1199, 0.0964])
#> tensor([-0.0920, -0.0690, -0.1475, 0.2313, -0.0872, 0.0799])
#> tensor([-0.0535, -0.0368, -0.2851, -0.0381, -0.0486, 0.2383])
Flair supports a range of classic word embeddings, each offering unique features and application scopes. Below is an overview, detailing the ID required to load each embedding and its corresponding language.
Embedding Type | ID | Language |
---|---|---|
GloVe | glove | English |
Komninos | extvec | English |
English | ||
Turian (small) | turian | English |
FastText (crawl) | crawl | English |
FastText (news & Wikipedia) | ar | Arabic |
FastText (news & Wikipedia) | bg | Bulgarian |
FastText (news & Wikipedia) | ca | Catalan |
FastText (news & Wikipedia) | cz | Czech |
FastText (news & Wikipedia) | da | Danish |
FastText (news & Wikipedia) | de | German |
FastText (news & Wikipedia) | es | Spanish |
FastText (news & Wikipedia) | en | English |
FastText (news & Wikipedia) | eu | Basque |
FastText (news & Wikipedia) | fa | Persian |
FastText (news & Wikipedia) | fi | Finnish |
FastText (news & Wikipedia) | fr | French |
FastText (news & Wikipedia) | he | Hebrew |
FastText (news & Wikipedia) | hi | Hindi |
FastText (news & Wikipedia) | hr | Croatian |
FastText (news & Wikipedia) | id | Indonesian |
FastText (news & Wikipedia) | it | Italian |
FastText (news & Wikipedia) | ja | Japanese |
FastText (news & Wikipedia) | ko | Korean |
FastText (news & Wikipedia) | nl | Dutch |
FastText (news & Wikipedia) | no | Norwegian |
FastText (news & Wikipedia) | pl | Polish |
FastText (news & Wikipedia) | pt | Portuguese |
FastText (news & Wikipedia) | ro | Romanian |
FastText (news & Wikipedia) | ru | Russian |
FastText (news & Wikipedia) | si | Slovenian |
FastText (news & Wikipedia) | sk | Slovak |
FastText (news & Wikipedia) | sr | Serbian |
FastText (news & Wikipedia) | sv | Swedish |
FastText (news & Wikipedia) | tr | Turkish |
FastText (news & Wikipedia) | zh | Chinese |
Contexual Embeddings
The idea behind contextual string embeddings is that each word embedding should be defined by not only its syntactic-semantic meaning but also the context it appears in. What this means is that each word will have a different embedding for every context it appears in. Each pre-trained Flair model offers a forward version and a backward version. Let’s assume you are processing a language that, just like this text, uses the left-to-right script. The forward version takes into account the context that happens before the word – on the left-hand side. The backward version works in the opposite direction. It takes into account the context after the word – on the right-hand side of the word. If this is true, then two same words that appear at the beginning of two different sentences should have identical forward embeddings, because their context is null. Let’s test this out:
Because we are using a forward model, it only takes into account the context that occurs before a word. Additionally, since our word has no context on the left-hand side of its position in the sentence, the two embeddings are identical, and the code assumes they are identical, indeed output is True.
FlairEmbeddings <- flair_embeddings()$FlairEmbeddings
embedding <- FlairEmbeddings('news-forward')
s1 <- Sentence("nice shirt")
s2 <- Sentence("nice pants")
embedding$embed(s1)
#> [[1]]
#> Sentence[2]: "nice shirt"
embedding$embed(s2)
#> [[1]]
#> Sentence[2]: "nice pants"
cat(" s1 sentence:", paste(s1[0], sep = ""), "\n", "s2 sentence:", paste(s2[0], sep = ""))
#> s1 sentence: Token[0]: "nice"
#> s2 sentence: Token[0]: "nice"
We test whether the sum of the two 2048 embeddings of
nice
is equal to 2048. If it is true, it indicates that the
embedding results are consistent, which should theoretically be the
case.
length(s1[0]$embedding$numpy()) == sum(s1[0]$embedding$numpy() == s2[0]$embedding$numpy())
#> [1] TRUE
Now we separately add a few more words, very
and
pretty
, into two sentence objects.
s1 <- Sentence("very nice shirt")
s2 <- Sentence("pretty nice pants")
embedding$embed(s1)
#> [[1]]
#> Sentence[3]: "very nice shirt"
embedding$embed(s2)
#> [[1]]
#> Sentence[3]: "pretty nice pants"
The two sets of embeddings are not identical because the words are different, so it returns FALSE.
length(s1[0]$embedding$numpy()) == sum(s1[0]$embedding$numpy() == s2[0]$embedding$numpy())
#> [1] FALSE
The measure of similarity between two vectors in an inner product space is known as cosine similarity. The formula for calculating cosine similarity between two vectors, such as vectors A and B, is as follows:
library(lsa)
vector1 <- as.numeric(s1[0]$embedding$numpy())
vector2 <- as.numeric(s2[0]$embedding$numpy())
We can observe that the similarity between the two words is 0.55.
Extracting Embeddings from BERT
First, we utilize the
flair.embeddings.TransformerWordEmbeddings
function to
download BERT, and more transformer models can also be found on Flair NLP’s Hugging Face.
library(flaiR)
TransformerWordEmbeddings <- flair_embeddings()$TransformerWordEmbeddings("bert-base-uncased")
embedding <- TransformerWordEmbeddings$embed(sentence)
Next, we traverse each token in the sentence and print them.
# Iterate through each token in the sentence, printing them.
# Utilize reticulate::py_str(token) to view each token, given that the sentence is a Python object.
for (i in seq_along(sentence$tokens)) {
cat("Token: ", reticulate::py_str(sentence$tokens[[i]]), "\n")
# Access the embedding of the token, converting it to an R object,
# and print the first 10 elements of the vector.
token_embedding <- sentence$tokens[[i]]$embedding
print(head(token_embedding, 10))
}
#> Token: Token[0]: "one"
#> tensor([-0.0535, -0.0368, -0.2851, -0.0381, -0.0486, 0.2383, -0.1200, 0.2620,
#> -0.0575, 0.0228])
#> Token: Token[1]: "two"
#> tensor([ 0.0282, -0.0786, -0.1236, 0.1756, -0.1199, 0.0964, -0.1327, 0.4449,
#> -0.0264, -0.1168])
#> Token: Token[2]: "three"
#> tensor([-0.0920, -0.0690, -0.1475, 0.2313, -0.0872, 0.0799, -0.0901, 0.4403,
#> -0.0103, -0.1494])
#> Token: Token[3]: "one"
#> tensor([-0.0535, -0.0368, -0.2851, -0.0381, -0.0486, 0.2383, -0.1200, 0.2620,
#> -0.0575, 0.0228])
Training a Binary Classifier
In this section, we’ll train a sentiment analysis model that can categorize text as either positive or negative. This case study is adapted from pages 116 to 130 of Tadej Magajna’s book, ‘Natural Language Processing with Flair’. The process for training text classifiers in Flair mirrors the process followed for sequence labeling models. Specifically, the steps to train text classifiers are:
- Load a tagged corpus and compute the label dictionary map.
- Prepare the document embeddings.
- Initialize the
TextClassifier
class. - Train the model.
Loading a Tagged Corpus
Training text classification models requires a set of text documents (typically, sentences or paragraphs) where each document is associated with one or more classification labels. To train our sentiment analysis text classification model, we will be using the famous Internet Movie Database (IMDb) dataset, which contains 50,000 movie reviews from IMDB, where each review is labeled as either positive or negative. References to this dataset are already baked into Flair, so loading the dataset couldn’t be easier:
library(flaiR)
# load IMDB from flair_datasets module
Corpus <- flair_data()$Corpus
IMDB <- flair_datasets()$IMDB
# downsize to 0.05
corpus = IMDB()
#> 2024-09-23 11:47:01,384 Reading data from /Users/yenchiehliao/.flair/datasets/imdb_v4-rebalanced
#> 2024-09-23 11:47:01,384 Train: /Users/yenchiehliao/.flair/datasets/imdb_v4-rebalanced/train.txt
#> 2024-09-23 11:47:01,384 Dev: None
#> 2024-09-23 11:47:01,384 Test: None
#> 2024-09-23 11:47:01,915 No test split found. Using 10% (i.e. 5000 samples) of the train split as test data
#> 2024-09-23 11:47:01,926 No dev split found. Using 10% (i.e. 4500 samples) of the train split as dev data
#> 2024-09-23 11:47:01,927 Initialized corpus /Users/yenchiehliao/.flair/datasets/imdb_v4-rebalanced (label type name is 'sentiment')
corpus$downsample(0.05)
#> <flair.datasets.document_classification.IMDB object at 0x3284e3a10>
Print the sizes in the corpus object as follows - test: %d | train: %d | dev: %d”
test_size <- length(corpus$test)
train_size <- length(corpus$train)
dev_size <- length(corpus$dev)
output <- sprintf("Corpus object sizes - Test: %d | Train: %d | Dev: %d", test_size, train_size, dev_size)
print(output)
#> [1] "Corpus object sizes - Test: 250 | Train: 2025 | Dev: 225"
lbl_type = 'sentiment'
label_dict = corpus$make_label_dictionary(label_type=lbl_type)
#> 2024-09-23 11:47:02,021 Computing label dictionary. Progress:
#> 2024-09-23 11:47:04,851 Dictionary created for label 'sentiment' with 2 values: POSITIVE (seen 1014 times), NEGATIVE (seen 1011 times)
Loading the Embeddings
flaiR covers all the different types of document embeddings that we
can use. Here, we simply use DocumentPoolEmbeddings
. They
require no training prior to training the classification model
itself:
DocumentPoolEmbeddings <- flair_embeddings()$DocumentPoolEmbeddings
WordEmbeddings <- flair_embeddings()$WordEmbeddings
glove = WordEmbeddings('glove')
document_embeddings = DocumentPoolEmbeddings(glove)
Initializing the TextClassifier
# initiate TextClassifier
TextClassifier <- flair_models()$TextClassifier
classifier <- TextClassifier(document_embeddings,
label_dictionary = label_dict,
label_type = lbl_type)
$to
allows you to set the device to use CPU, GPU, or
specific MPS devices on Mac (such as mps:0, mps:1, mps:2).
classifier$to(flair_device("mps"))
TextClassifier(
(embeddings): DocumentPoolEmbeddings(
fine_tune_mode=none, pooling=mean
(embeddings): StackedEmbeddings(
(list_embedding_0): WordEmbeddings(
'glove'
(embedding): Embedding(400001, 100)
)
)
)
(decoder): Linear(in_features=100, out_features=3, bias=True)
(dropout): Dropout(p=0.0, inplace=False)
(locked_dropout): LockedDropout(p=0.0)
(word_dropout): WordDropout(p=0.0)
(loss_function): CrossEntropyLoss()
)
Training the Model
Training the text classifier model involves two simple steps:
- Defining the model trainer class by passing in the classifier model and the corpus
- Setting off the training process passing in the required training hyper-parameters.
It is worth noting that the ‘L’ in numbers like 32L and 5L is used in R to denote that the number is an integer. Without the ‘L’ suffix, numbers in R are treated as numeric, which are by default double-precision floating-point numbers. In contrast, Python determines the type based on the value of the number itself. Whole numbers (e.g., 5 or 32) are of type int, while numbers with decimal points (e.g., 5.0) are of type float. Floating-point numbers in both languages are representations of real numbers but can have some approximation due to the way they are stored in memory.
# initiate ModelTrainer
ModelTrainer <- flair_trainers()$ModelTrainer
# fit the model
trainer <- ModelTrainer(classifier, corpus)
# start to train
# note: the 'L' in 32L is used in R to denote that the number is an integer.
trainer$train('classifier',
learning_rate=0.1,
mini_batch_size=32L,
# specifies how embeddings are stored in RAM, ie."cpu", "cuda", "gpu", "mps".
# embeddings_storage_mode = "mps",
max_epochs=10L)
#> 2024-09-23 11:47:06,959 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:47:06,959 Model: "TextClassifier(
#> (embeddings): DocumentPoolEmbeddings(
#> fine_tune_mode=none, pooling=mean
#> (embeddings): StackedEmbeddings(
#> (list_embedding_0): WordEmbeddings(
#> 'glove'
#> (embedding): Embedding(400001, 100)
#> )
#> )
#> )
#> (decoder): Linear(in_features=100, out_features=2, bias=True)
#> (dropout): Dropout(p=0.0, inplace=False)
#> (locked_dropout): LockedDropout(p=0.0)
#> (word_dropout): WordDropout(p=0.0)
#> (loss_function): CrossEntropyLoss()
#> (weights): None
#> (weight_tensor) None
#> )"
#> 2024-09-23 11:47:06,959 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:47:06,959 Corpus: 2025 train + 225 dev + 250 test sentences
#> 2024-09-23 11:47:06,959 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:47:06,959 Train: 2025 sentences
#> 2024-09-23 11:47:06,959 (train_with_dev=False, train_with_test=False)
#> 2024-09-23 11:47:06,959 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:47:06,959 Training Params:
#> 2024-09-23 11:47:06,959 - learning_rate: "0.1"
#> 2024-09-23 11:47:06,959 - mini_batch_size: "32"
#> 2024-09-23 11:47:06,959 - max_epochs: "10"
#> 2024-09-23 11:47:06,959 - shuffle: "True"
#> 2024-09-23 11:47:06,959 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:47:06,959 Plugins:
#> 2024-09-23 11:47:06,959 - AnnealOnPlateau | patience: '3', anneal_factor: '0.5', min_learning_rate: '0.0001'
#> 2024-09-23 11:47:06,959 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:47:06,959 Final evaluation on model from best epoch (best-model.pt)
#> 2024-09-23 11:47:06,959 - metric: "('micro avg', 'f1-score')"
#> 2024-09-23 11:47:06,959 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:47:06,960 Computation:
#> 2024-09-23 11:47:06,960 - compute on device: cpu
#> 2024-09-23 11:47:06,960 - embedding storage: cpu
#> 2024-09-23 11:47:06,960 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:47:06,960 Model training base path: "classifier"
#> 2024-09-23 11:47:06,960 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:47:06,960 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:47:07,597 epoch 1 - iter 6/64 - loss 0.87478685 - time (sec): 0.64 - samples/sec: 301.48 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:08,248 epoch 1 - iter 12/64 - loss 0.86428784 - time (sec): 1.29 - samples/sec: 298.11 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:09,141 epoch 1 - iter 18/64 - loss 0.88217321 - time (sec): 2.18 - samples/sec: 264.11 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:09,826 epoch 1 - iter 24/64 - loss 0.89574670 - time (sec): 2.87 - samples/sec: 267.99 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:10,752 epoch 1 - iter 30/64 - loss 0.91061279 - time (sec): 3.79 - samples/sec: 253.12 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:11,436 epoch 1 - iter 36/64 - loss 0.89059635 - time (sec): 4.48 - samples/sec: 257.38 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:12,139 epoch 1 - iter 42/64 - loss 0.88880198 - time (sec): 5.18 - samples/sec: 259.51 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:12,804 epoch 1 - iter 48/64 - loss 0.88711355 - time (sec): 5.84 - samples/sec: 262.81 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:13,507 epoch 1 - iter 54/64 - loss 0.89178114 - time (sec): 6.55 - samples/sec: 263.91 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:14,402 epoch 1 - iter 60/64 - loss 0.89531235 - time (sec): 7.44 - samples/sec: 258.00 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:14,679 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:47:14,679 EPOCH 1 done: loss 0.8964 - lr: 0.100000
#> 2024-09-23 11:47:15,787 DEV : loss 0.9332443475723267 - f1-score (micro avg) 0.5467
#> 2024-09-23 11:47:16,135 - 0 epochs without improvement
#> 2024-09-23 11:47:16,135 saving best model
#> 2024-09-23 11:47:16,512 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:47:17,315 epoch 2 - iter 6/64 - loss 0.83218794 - time (sec): 0.80 - samples/sec: 239.40 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:17,959 epoch 2 - iter 12/64 - loss 0.85133497 - time (sec): 1.45 - samples/sec: 265.52 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:18,798 epoch 2 - iter 18/64 - loss 0.86734511 - time (sec): 2.29 - samples/sec: 252.03 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:19,489 epoch 2 - iter 24/64 - loss 0.86159388 - time (sec): 2.98 - samples/sec: 258.02 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:20,210 epoch 2 - iter 30/64 - loss 0.86639629 - time (sec): 3.70 - samples/sec: 259.66 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:20,849 epoch 2 - iter 36/64 - loss 0.87058405 - time (sec): 4.34 - samples/sec: 265.65 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:21,799 epoch 2 - iter 42/64 - loss 0.87439957 - time (sec): 5.29 - samples/sec: 254.25 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:22,483 epoch 2 - iter 48/64 - loss 0.86491772 - time (sec): 5.97 - samples/sec: 257.27 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:23,193 epoch 2 - iter 54/64 - loss 0.86861187 - time (sec): 6.68 - samples/sec: 258.68 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:23,914 epoch 2 - iter 60/64 - loss 0.87278313 - time (sec): 7.40 - samples/sec: 259.41 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:24,402 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:47:24,402 EPOCH 2 done: loss 0.8677 - lr: 0.100000
#> 2024-09-23 11:47:25,524 DEV : loss 1.107783555984497 - f1-score (micro avg) 0.4533
#> 2024-09-23 11:47:25,875 - 1 epochs without improvement
#> 2024-09-23 11:47:25,876 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:47:26,672 epoch 3 - iter 6/64 - loss 0.89582851 - time (sec): 0.80 - samples/sec: 241.25 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:27,401 epoch 3 - iter 12/64 - loss 0.88788915 - time (sec): 1.53 - samples/sec: 251.77 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:28,305 epoch 3 - iter 18/64 - loss 0.88291740 - time (sec): 2.43 - samples/sec: 237.14 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:28,999 epoch 3 - iter 24/64 - loss 0.87198854 - time (sec): 3.12 - samples/sec: 245.95 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:29,690 epoch 3 - iter 30/64 - loss 0.85297724 - time (sec): 3.81 - samples/sec: 251.69 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:30,393 epoch 3 - iter 36/64 - loss 0.86674236 - time (sec): 4.52 - samples/sec: 255.06 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:31,099 epoch 3 - iter 42/64 - loss 0.86457429 - time (sec): 5.22 - samples/sec: 257.32 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:31,767 epoch 3 - iter 48/64 - loss 0.86558150 - time (sec): 5.89 - samples/sec: 260.72 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:32,614 epoch 3 - iter 54/64 - loss 0.86672546 - time (sec): 6.74 - samples/sec: 256.46 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:33,294 epoch 3 - iter 60/64 - loss 0.87017401 - time (sec): 7.42 - samples/sec: 258.84 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:33,586 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:47:33,586 EPOCH 3 done: loss 0.8687 - lr: 0.100000
#> 2024-09-23 11:47:34,677 DEV : loss 1.0332790613174438 - f1-score (micro avg) 0.4533
#> 2024-09-23 11:47:35,024 - 2 epochs without improvement
#> 2024-09-23 11:47:35,024 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:47:35,951 epoch 4 - iter 6/64 - loss 0.88925039 - time (sec): 0.93 - samples/sec: 207.28 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:36,604 epoch 4 - iter 12/64 - loss 0.85552217 - time (sec): 1.58 - samples/sec: 243.13 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:37,309 epoch 4 - iter 18/64 - loss 0.86623563 - time (sec): 2.28 - samples/sec: 252.19 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:38,057 epoch 4 - iter 24/64 - loss 0.84697705 - time (sec): 3.03 - samples/sec: 253.29 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:38,947 epoch 4 - iter 30/64 - loss 0.83577118 - time (sec): 3.92 - samples/sec: 244.73 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:39,628 epoch 4 - iter 36/64 - loss 0.81822756 - time (sec): 4.60 - samples/sec: 250.24 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:40,319 epoch 4 - iter 42/64 - loss 0.83048539 - time (sec): 5.29 - samples/sec: 253.88 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:41,022 epoch 4 - iter 48/64 - loss 0.83219909 - time (sec): 6.00 - samples/sec: 256.11 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:41,898 epoch 4 - iter 54/64 - loss 0.83317216 - time (sec): 6.87 - samples/sec: 251.40 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:42,559 epoch 4 - iter 60/64 - loss 0.83902879 - time (sec): 7.53 - samples/sec: 254.84 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:43,020 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:47:43,021 EPOCH 4 done: loss 0.8383 - lr: 0.100000
#> 2024-09-23 11:47:43,900 DEV : loss 0.669340968132019 - f1-score (micro avg) 0.6
#> 2024-09-23 11:47:44,246 - 0 epochs without improvement
#> 2024-09-23 11:47:44,246 saving best model
#> 2024-09-23 11:47:44,605 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:47:45,672 epoch 5 - iter 6/64 - loss 0.91042657 - time (sec): 1.07 - samples/sec: 179.98 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:46,351 epoch 5 - iter 12/64 - loss 0.86260041 - time (sec): 1.75 - samples/sec: 219.91 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:47,047 epoch 5 - iter 18/64 - loss 0.83276293 - time (sec): 2.44 - samples/sec: 235.86 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:47,991 epoch 5 - iter 24/64 - loss 0.84274461 - time (sec): 3.39 - samples/sec: 226.81 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:48,706 epoch 5 - iter 30/64 - loss 0.81705600 - time (sec): 4.10 - samples/sec: 234.11 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:49,419 epoch 5 - iter 36/64 - loss 0.81764214 - time (sec): 4.81 - samples/sec: 239.28 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:50,113 epoch 5 - iter 42/64 - loss 0.80878493 - time (sec): 5.51 - samples/sec: 243.99 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:50,781 epoch 5 - iter 48/64 - loss 0.81581626 - time (sec): 6.18 - samples/sec: 248.69 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:51,408 epoch 5 - iter 54/64 - loss 0.82706751 - time (sec): 6.80 - samples/sec: 254.01 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:52,304 epoch 5 - iter 60/64 - loss 0.83341895 - time (sec): 7.70 - samples/sec: 249.39 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:52,577 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:47:52,577 EPOCH 5 done: loss 0.8351 - lr: 0.100000
#> 2024-09-23 11:47:53,679 DEV : loss 0.7678796648979187 - f1-score (micro avg) 0.5556
#> 2024-09-23 11:47:54,030 - 1 epochs without improvement
#> 2024-09-23 11:47:54,031 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:47:55,003 epoch 6 - iter 6/64 - loss 0.80075386 - time (sec): 0.97 - samples/sec: 197.50 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:55,702 epoch 6 - iter 12/64 - loss 0.82380461 - time (sec): 1.67 - samples/sec: 229.83 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:56,414 epoch 6 - iter 18/64 - loss 0.81097129 - time (sec): 2.38 - samples/sec: 241.71 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:57,336 epoch 6 - iter 24/64 - loss 0.82260606 - time (sec): 3.31 - samples/sec: 232.36 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:57,985 epoch 6 - iter 30/64 - loss 0.81337555 - time (sec): 3.95 - samples/sec: 242.82 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:58,672 epoch 6 - iter 36/64 - loss 0.80567515 - time (sec): 4.64 - samples/sec: 248.20 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:47:59,386 epoch 6 - iter 42/64 - loss 0.80949111 - time (sec): 5.35 - samples/sec: 251.00 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:48:00,048 epoch 6 - iter 48/64 - loss 0.80318874 - time (sec): 6.02 - samples/sec: 255.29 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:48:00,907 epoch 6 - iter 54/64 - loss 0.79650182 - time (sec): 6.88 - samples/sec: 251.30 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:48:01,586 epoch 6 - iter 60/64 - loss 0.79675878 - time (sec): 7.56 - samples/sec: 254.12 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:48:01,871 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:48:01,871 EPOCH 6 done: loss 0.7947 - lr: 0.100000
#> 2024-09-23 11:48:02,984 DEV : loss 0.7552272081375122 - f1-score (micro avg) 0.56
#> 2024-09-23 11:48:03,332 - 2 epochs without improvement
#> 2024-09-23 11:48:03,333 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:48:04,281 epoch 7 - iter 6/64 - loss 0.74463288 - time (sec): 0.95 - samples/sec: 202.51 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:48:04,954 epoch 7 - iter 12/64 - loss 0.78036737 - time (sec): 1.62 - samples/sec: 236.80 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:48:05,599 epoch 7 - iter 18/64 - loss 0.78813493 - time (sec): 2.27 - samples/sec: 254.11 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:48:06,259 epoch 7 - iter 24/64 - loss 0.77977300 - time (sec): 2.93 - samples/sec: 262.41 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:48:07,171 epoch 7 - iter 30/64 - loss 0.79734032 - time (sec): 3.84 - samples/sec: 250.12 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:48:07,903 epoch 7 - iter 36/64 - loss 0.81214167 - time (sec): 4.57 - samples/sec: 252.09 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:48:08,602 epoch 7 - iter 42/64 - loss 0.80828580 - time (sec): 5.27 - samples/sec: 255.08 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:48:09,308 epoch 7 - iter 48/64 - loss 0.79509349 - time (sec): 5.98 - samples/sec: 257.04 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:48:10,207 epoch 7 - iter 54/64 - loss 0.79817696 - time (sec): 6.87 - samples/sec: 251.39 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:48:10,862 epoch 7 - iter 60/64 - loss 0.79470734 - time (sec): 7.53 - samples/sec: 255.01 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:48:11,126 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:48:11,126 EPOCH 7 done: loss 0.7914 - lr: 0.100000
#> 2024-09-23 11:48:12,205 DEV : loss 1.179593563079834 - f1-score (micro avg) 0.4533
#> 2024-09-23 11:48:12,556 - 3 epochs without improvement
#> 2024-09-23 11:48:12,556 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:48:13,520 epoch 8 - iter 6/64 - loss 0.78675573 - time (sec): 0.96 - samples/sec: 199.39 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:48:14,222 epoch 8 - iter 12/64 - loss 0.76353798 - time (sec): 1.67 - samples/sec: 230.54 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:48:14,910 epoch 8 - iter 18/64 - loss 0.77087792 - time (sec): 2.35 - samples/sec: 244.76 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:48:15,615 epoch 8 - iter 24/64 - loss 0.78264684 - time (sec): 3.06 - samples/sec: 251.08 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:48:16,273 epoch 8 - iter 30/64 - loss 0.78193761 - time (sec): 3.72 - samples/sec: 258.29 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:48:17,183 epoch 8 - iter 36/64 - loss 0.79271775 - time (sec): 4.63 - samples/sec: 249.01 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:48:17,964 epoch 8 - iter 42/64 - loss 0.80765662 - time (sec): 5.41 - samples/sec: 248.53 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:48:18,639 epoch 8 - iter 48/64 - loss 0.81049288 - time (sec): 6.08 - samples/sec: 252.53 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:48:19,385 epoch 8 - iter 54/64 - loss 0.81939115 - time (sec): 6.83 - samples/sec: 253.07 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:48:20,256 epoch 8 - iter 60/64 - loss 0.82118929 - time (sec): 7.70 - samples/sec: 249.37 - lr: 0.100000 - momentum: 0.000000
#> 2024-09-23 11:48:20,534 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:48:20,534 EPOCH 8 done: loss 0.8182 - lr: 0.100000
#> 2024-09-23 11:48:21,444 DEV : loss 0.706107497215271 - f1-score (micro avg) 0.5467
#> 2024-09-23 11:48:22,039 - 4 epochs without improvement (above 'patience')-> annealing learning_rate to [0.05]
#> 2024-09-23 11:48:22,040 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:48:22,727 epoch 9 - iter 6/64 - loss 0.66559014 - time (sec): 0.69 - samples/sec: 279.70 - lr: 0.050000 - momentum: 0.000000
#> 2024-09-23 11:48:23,626 epoch 9 - iter 12/64 - loss 0.66215663 - time (sec): 1.59 - samples/sec: 242.16 - lr: 0.050000 - momentum: 0.000000
#> 2024-09-23 11:48:24,341 epoch 9 - iter 18/64 - loss 0.64717902 - time (sec): 2.30 - samples/sec: 250.38 - lr: 0.050000 - momentum: 0.000000
#> 2024-09-23 11:48:25,023 epoch 9 - iter 24/64 - loss 0.65857949 - time (sec): 2.98 - samples/sec: 257.43 - lr: 0.050000 - momentum: 0.000000
#> 2024-09-23 11:48:25,708 epoch 9 - iter 30/64 - loss 0.65033504 - time (sec): 3.67 - samples/sec: 261.71 - lr: 0.050000 - momentum: 0.000000
#> 2024-09-23 11:48:26,407 epoch 9 - iter 36/64 - loss 0.64554224 - time (sec): 4.37 - samples/sec: 263.82 - lr: 0.050000 - momentum: 0.000000
#> 2024-09-23 11:48:27,309 epoch 9 - iter 42/64 - loss 0.64265556 - time (sec): 5.27 - samples/sec: 255.08 - lr: 0.050000 - momentum: 0.000000
#> 2024-09-23 11:48:28,006 epoch 9 - iter 48/64 - loss 0.64192964 - time (sec): 5.97 - samples/sec: 257.45 - lr: 0.050000 - momentum: 0.000000
#> 2024-09-23 11:48:28,684 epoch 9 - iter 54/64 - loss 0.64669667 - time (sec): 6.64 - samples/sec: 260.10 - lr: 0.050000 - momentum: 0.000000
#> 2024-09-23 11:48:29,362 epoch 9 - iter 60/64 - loss 0.64877095 - time (sec): 7.32 - samples/sec: 262.24 - lr: 0.050000 - momentum: 0.000000
#> 2024-09-23 11:48:29,843 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:48:29,844 EPOCH 9 done: loss 0.6509 - lr: 0.050000
#> 2024-09-23 11:48:30,958 DEV : loss 0.5682870745658875 - f1-score (micro avg) 0.7244
#> 2024-09-23 11:48:31,312 - 0 epochs without improvement
#> 2024-09-23 11:48:31,312 saving best model
#> 2024-09-23 11:48:31,596 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:48:32,402 epoch 10 - iter 6/64 - loss 0.58569980 - time (sec): 0.81 - samples/sec: 238.21 - lr: 0.050000 - momentum: 0.000000
#> 2024-09-23 11:48:33,130 epoch 10 - iter 12/64 - loss 0.62290008 - time (sec): 1.53 - samples/sec: 250.33 - lr: 0.050000 - momentum: 0.000000
#> 2024-09-23 11:48:34,008 epoch 10 - iter 18/64 - loss 0.63693839 - time (sec): 2.41 - samples/sec: 238.84 - lr: 0.050000 - momentum: 0.000000
#> 2024-09-23 11:48:34,744 epoch 10 - iter 24/64 - loss 0.63064030 - time (sec): 3.15 - samples/sec: 243.98 - lr: 0.050000 - momentum: 0.000000
#> 2024-09-23 11:48:35,374 epoch 10 - iter 30/64 - loss 0.63178459 - time (sec): 3.78 - samples/sec: 254.10 - lr: 0.050000 - momentum: 0.000000
#> 2024-09-23 11:48:36,026 epoch 10 - iter 36/64 - loss 0.63311307 - time (sec): 4.43 - samples/sec: 260.03 - lr: 0.050000 - momentum: 0.000000
#> 2024-09-23 11:48:36,955 epoch 10 - iter 42/64 - loss 0.63235253 - time (sec): 5.36 - samples/sec: 250.81 - lr: 0.050000 - momentum: 0.000000
#> 2024-09-23 11:48:37,682 epoch 10 - iter 48/64 - loss 0.62940139 - time (sec): 6.09 - samples/sec: 252.36 - lr: 0.050000 - momentum: 0.000000
#> 2024-09-23 11:48:38,408 epoch 10 - iter 54/64 - loss 0.62466771 - time (sec): 6.81 - samples/sec: 253.68 - lr: 0.050000 - momentum: 0.000000
#> 2024-09-23 11:48:39,092 epoch 10 - iter 60/64 - loss 0.62534599 - time (sec): 7.50 - samples/sec: 256.15 - lr: 0.050000 - momentum: 0.000000
#> 2024-09-23 11:48:39,554 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:48:39,554 EPOCH 10 done: loss 0.6239 - lr: 0.050000
#> 2024-09-23 11:48:40,468 DEV : loss 0.6515768766403198 - f1-score (micro avg) 0.6178
#> 2024-09-23 11:48:41,072 - 1 epochs without improvement
#> 2024-09-23 11:48:41,340 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:48:41,340 Loading model from best epoch ...
#> 2024-09-23 11:48:42,638
#> Results:
#> - F-score (micro) 0.752
#> - F-score (macro) 0.7517
#> - Accuracy 0.752
#>
#> By class:
#> precision recall f1-score support
#>
#> POSITIVE 0.7597 0.7597 0.7597 129
#> NEGATIVE 0.7438 0.7438 0.7438 121
#>
#> accuracy 0.7520 250
#> macro avg 0.7517 0.7517 0.7517 250
#> weighted avg 0.7520 0.7520 0.7520 250
#>
#> 2024-09-23 11:48:42,638 ----------------------------------------------------------------------------------------------------
#> $test_score
#> [1] 0.752
Loading and Using the Classifiers
After training the text classification model, the resulting classifier will already be stored in memory as part of the classifier variable. It is possible, however, that your Python session exited after training. If so, you’ll need to load the model into memory with the following:
TextClassifier <- flair_models()$TextClassifier
classifier <- TextClassifier$load('classifier/best-model.pt')
We import the Sentence object. Now, we can generate predictions on some example text inputs.
Sentence <- flair_data()$Sentence
sentence <- Sentence("great")
classifier$predict(sentence)
print(sentence$labels)
#> [[1]]
#> 'Sentence[1]: "great"'/'POSITIVE' (1.0)
sentence <- Sentence("sad")
classifier$predict(sentence)
print(sentence$labels)
#> [[1]]
#> 'Sentence[1]: "sad"'/'NEGATIVE' (0.7621)
Training RNNs
Here, we train a sentiment analysis model to categorize text. In this case, we also include a pipeline that implements the use of Recurrent Neural Networks (RNN). This makes them particularly effective for tasks involving sequential data. This section also show you how to implement one of most powerful features in flaiR, stacked embeddings. You can stack multiple embeddings with different layers and let the classifier learn from different types of features. In Flair NLP, and with the flaiR package, it’s very easy to accomplish this task.
Import Necessary Modules
library(flaiR)
WordEmbeddings <- flair_embeddings()$WordEmbeddings
FlairEmbeddings <- flair_embeddings()$FlairEmbeddings
DocumentRNNEmbeddings <- flair_embeddings()$DocumentRNNEmbeddings
TextClassifier <- flair_models()$TextClassifier
ModelTrainer <- flair_trainers()$ModelTrainer
Get the IMDB Corpus
The IMDB movie review dataset is used here, which is a commonly
utilized dataset for sentiment analysis. $downsample(0.1)
method means only 10% of the dataset is used, allowing for a faster
demonstration.
# load the IMDB file and downsize it to 0.1
IMDB <- flair_datasets()$IMDB
corpus <- IMDB()$downsample(0.1)
#> 2024-09-23 11:48:43,042 Reading data from /Users/yenchiehliao/.flair/datasets/imdb_v4-rebalanced
#> 2024-09-23 11:48:43,042 Train: /Users/yenchiehliao/.flair/datasets/imdb_v4-rebalanced/train.txt
#> 2024-09-23 11:48:43,042 Dev: None
#> 2024-09-23 11:48:43,042 Test: None
#> 2024-09-23 11:48:43,555 No test split found. Using 10% (i.e. 5000 samples) of the train split as test data
#> 2024-09-23 11:48:43,566 No dev split found. Using 10% (i.e. 4500 samples) of the train split as dev data
#> 2024-09-23 11:48:43,566 Initialized corpus /Users/yenchiehliao/.flair/datasets/imdb_v4-rebalanced (label type name is 'sentiment')
# create the label dictionary
lbl_type <- 'sentiment'
label_dict <- corpus$make_label_dictionary(label_type=lbl_type)
#> 2024-09-23 11:48:43,578 Computing label dictionary. Progress:
#> 2024-09-23 11:48:49,992 Dictionary created for label 'sentiment' with 2 values: POSITIVE (seen 2056 times), NEGATIVE (seen 1994 times)
Stacked Embeddings
This is one of Flair’s most powerful features: it allows for the integration of embeddings to enable the model to learn from more sparse features. Three types of embeddings are utilized here: GloVe embeddings, and two types of Flair embeddings (forward and backward). Word embeddings are used to convert words into vectors.
# make a list of word embeddings
word_embeddings <- list(WordEmbeddings('glove'),
FlairEmbeddings('news-forward-fast'),
FlairEmbeddings('news-backward-fast'))
# initialize the document embeddings
document_embeddings <- DocumentRNNEmbeddings(word_embeddings,
hidden_size = 512L,
reproject_words = TRUE,
reproject_words_dimension = 256L)
# create a Text Classifier with the embeddings and label dictionary
classifier <- TextClassifier(document_embeddings,
label_dictionary=label_dict, label_type='class')
# initialize the text classifier trainer with our corpus
trainer <- ModelTrainer(classifier, corpus)
Start the Training
For the sake of this example, setting max_epochs to 5. You might want to increase this for better performance.
It is worth noting that the learning rate is a parameter that
determines the step size at each iteration while moving towards a
minimum of the loss function. A smaller learning rate could slow down
the learning process, but it could lead to more precise convergence.
mini_batch_size
determines the number of samples that will
be used to compute the gradient at each step. The ‘L’ in 32L is used in
R to denote that the number is an integer.
patience
(aka early stop) is a hyper-parameter used in
conjunction with early stopping to avoid overfitting. It determines the
number of epochs the training process will tolerate without improvements
before stopping the training. Setting max_epochs to 5 means the
algorithm will make five passes through the dataset.
# note: the 'L' in 32L is used in R to denote that the number is an integer.
trainer$train('models/sentiment',
learning_rate=0.1,
mini_batch_size=32L,
patience=5L,
max_epochs=5L)
To Apply the Trained Model for Prediction
sentence <- "This movie was really exciting!"
classifier$predict(sentence)
print(sentence.labels)
Finetune Transformers
We use data from The Temporal Focus of Campaign Communication
(2020 JOP) as an example. Let’s assume we receive the data for
training from different times. First, suppose you have a dataset of 1000
entries called cc_muller_old
. On another day, with the help
of nice friends, you receive another set of data, adding 2000 entries in
a dataset called cc_muller_new
. Both subsets are from
data(cc_muller)
. We will show how to fine-tune a
transformer model with cc_muller_old
, and then continue
with another round of fine-tuning using cc_muller_new
.
Fine-tuning a Transformers Model
Step 1 Load Necessary Modules from Flair
Load necessary classes from flair
package.
# Sentence is a class for holding a text sentence
Sentence <- flair_data()$Sentence
# Corpus is a class for text corpora
Corpus <- flair_data()$Corpus
# TransformerDocumentEmbeddings is a class for loading transformer
TransformerDocumentEmbeddings <- flair_embeddings()$TransformerDocumentEmbeddings
# TextClassifier is a class for text classification
TextClassifier <- flair_models()$TextClassifier
# ModelTrainer is a class for training and evaluating models
ModelTrainer <- flair_trainers()$ModelTrainer
We use purrr to help us split sentences using Sentence from
flair_data()
, then use map2 to add labels, and finally use
Corpus
to segment the data.
library(purrr)
data(cc_muller)
cc_muller_old <- cc_muller[1:1000,]
old_text <- map(cc_muller_old$text, Sentence)
old_labels <- as.character(cc_muller_old$class)
old_text <- map2(old_text, old_labels, ~ {
.x$add_label("classification", .y)
.x
})
set.seed(2046)
sample <- sample(c(TRUE, FALSE), length(old_text), replace=TRUE, prob=c(0.8, 0.2))
old_train <- old_text[sample]
old_test <- old_text[!sample]
test_id <- sample(c(TRUE, FALSE), length(old_test), replace=TRUE, prob=c(0.5, 0.5))
old_test <- old_test[test_id]
old_dev <- old_test[!test_id]
If you do not provide a development set (dev set) while using Flair, it will automatically split the training data into training and development datasets. The test set is used for training the model and evaluating its final performance, whereas the development set is used for adjusting model parameters and preventing overfitting, or in other words, for early stopping of the model.
old_corpus <- Corpus(train = old_train, test = old_test)
#> 2024-09-23 11:48:52,427 No dev split found. Using 10% (i.e. 80 samples) of the train split as dev data
Step 3 Load distilbert
Transformer
document_embeddings <- TransformerDocumentEmbeddings('distilbert-base-uncased', fine_tune=TRUE)
First, the $make_label_dictionary
function is used to
automatically create a label dictionary for the classification task. The
label dictionary is a mapping from label to index, which is used to map
the labels to a tensor of label indices. Besides classification tasks,
flaiR also supports other label types for training custom model, such as
ner
, pos
and sentiment
. From the
cc_muller dataset: Future (seen 423 times), Present (seen 262 times),
Past (seen 131 times)
old_label_dict <- old_corpus$make_label_dictionary(label_type="classification")
#> 2024-09-23 11:48:53,542 Computing label dictionary. Progress:
#> 2024-09-23 11:48:53,547 Dictionary created for label 'classification' with 3 values: Future (seen 380 times), Present (seen 232 times), Past (seen 111 times)
TextClassifier
is used to create a text classifier. The
classifier takes the document embeddings (importing from
'distilbert-base-uncased'
from Hugging Face) and the label
dictionary as input. The label type is also specified as
classification.
old_classifier <- TextClassifier(document_embeddings,
label_dictionary = old_label_dict,
label_type='classification')
Step 4 Start Training
ModelTrainer
is used to train the model.
old_trainer <- ModelTrainer(model = old_classifier, corpus = old_corpus)
old_trainer$train("vignettes/inst/muller-campaign-communication",
learning_rate=0.02,
mini_batch_size=8L,
anneal_with_restarts = TRUE,
save_final_model=TRUE,
max_epochs=1L)
#> 2024-09-23 11:48:53,673 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:48:53,673 Model: "TextClassifier(
#> (embeddings): TransformerDocumentEmbeddings(
#> (model): DistilBertModel(
#> (embeddings): Embeddings(
#> (word_embeddings): Embedding(30523, 768, padding_idx=0)
#> (position_embeddings): Embedding(512, 768)
#> (LayerNorm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)
#> (dropout): Dropout(p=0.1, inplace=False)
#> )
#> (transformer): Transformer(
#> (layer): ModuleList(
#> (0-5): 6 x TransformerBlock(
#> (attention): MultiHeadSelfAttention(
#> (dropout): Dropout(p=0.1, inplace=False)
#> (q_lin): Linear(in_features=768, out_features=768, bias=True)
#> (k_lin): Linear(in_features=768, out_features=768, bias=True)
#> (v_lin): Linear(in_features=768, out_features=768, bias=True)
#> (out_lin): Linear(in_features=768, out_features=768, bias=True)
#> )
#> (sa_layer_norm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)
#> (ffn): FFN(
#> (dropout): Dropout(p=0.1, inplace=False)
#> (lin1): Linear(in_features=768, out_features=3072, bias=True)
#> (lin2): Linear(in_features=3072, out_features=768, bias=True)
#> (activation): GELUActivation()
#> )
#> (output_layer_norm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)
#> )
#> )
#> )
#> )
#> )
#> (decoder): Linear(in_features=768, out_features=3, bias=True)
#> (dropout): Dropout(p=0.0, inplace=False)
#> (locked_dropout): LockedDropout(p=0.0)
#> (word_dropout): WordDropout(p=0.0)
#> (loss_function): CrossEntropyLoss()
#> (weights): None
#> (weight_tensor) None
#> )"
#> 2024-09-23 11:48:53,673 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:48:53,673 Corpus: 723 train + 80 dev + 85 test sentences
#> 2024-09-23 11:48:53,673 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:48:53,673 Train: 723 sentences
#> 2024-09-23 11:48:53,673 (train_with_dev=False, train_with_test=False)
#> 2024-09-23 11:48:53,674 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:48:53,674 Training Params:
#> 2024-09-23 11:48:53,674 - learning_rate: "0.02"
#> 2024-09-23 11:48:53,674 - mini_batch_size: "8"
#> 2024-09-23 11:48:53,674 - max_epochs: "1"
#> 2024-09-23 11:48:53,674 - shuffle: "True"
#> 2024-09-23 11:48:53,674 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:48:53,674 Plugins:
#> 2024-09-23 11:48:53,674 - AnnealOnPlateau | patience: '3', anneal_factor: '0.5', min_learning_rate: '0.0001'
#> 2024-09-23 11:48:53,674 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:48:53,674 Final evaluation on model from best epoch (best-model.pt)
#> 2024-09-23 11:48:53,674 - metric: "('micro avg', 'f1-score')"
#> 2024-09-23 11:48:53,674 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:48:53,674 Computation:
#> 2024-09-23 11:48:53,674 - compute on device: cpu
#> 2024-09-23 11:48:53,674 - embedding storage: cpu
#> 2024-09-23 11:48:53,674 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:48:53,674 Model training base path: "vignettes/inst/muller-campaign-communication"
#> 2024-09-23 11:48:53,674 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:48:53,674 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:48:56,314 epoch 1 - iter 9/91 - loss 0.89193513 - time (sec): 2.64 - samples/sec: 27.27 - lr: 0.020000 - momentum: 0.000000
#> 2024-09-23 11:48:58,531 epoch 1 - iter 18/91 - loss 0.91054643 - time (sec): 4.86 - samples/sec: 29.65 - lr: 0.020000 - momentum: 0.000000
#> 2024-09-23 11:49:01,114 epoch 1 - iter 27/91 - loss 0.85574824 - time (sec): 7.44 - samples/sec: 29.04 - lr: 0.020000 - momentum: 0.000000
#> 2024-09-23 11:49:03,511 epoch 1 - iter 36/91 - loss 0.75750717 - time (sec): 9.84 - samples/sec: 29.28 - lr: 0.020000 - momentum: 0.000000
#> 2024-09-23 11:49:05,905 epoch 1 - iter 45/91 - loss 0.72997100 - time (sec): 12.23 - samples/sec: 29.43 - lr: 0.020000 - momentum: 0.000000
#> 2024-09-23 11:49:08,354 epoch 1 - iter 54/91 - loss 0.69230357 - time (sec): 14.68 - samples/sec: 29.43 - lr: 0.020000 - momentum: 0.000000
#> 2024-09-23 11:49:10,767 epoch 1 - iter 63/91 - loss 0.67165820 - time (sec): 17.09 - samples/sec: 29.49 - lr: 0.020000 - momentum: 0.000000
#> 2024-09-23 11:49:13,292 epoch 1 - iter 72/91 - loss 0.69162410 - time (sec): 19.62 - samples/sec: 29.36 - lr: 0.020000 - momentum: 0.000000
#> 2024-09-23 11:49:15,776 epoch 1 - iter 81/91 - loss 0.68180142 - time (sec): 22.10 - samples/sec: 29.32 - lr: 0.020000 - momentum: 0.000000
#> 2024-09-23 11:49:18,047 epoch 1 - iter 90/91 - loss 0.67393552 - time (sec): 24.37 - samples/sec: 29.54 - lr: 0.020000 - momentum: 0.000000
#> 2024-09-23 11:49:18,196 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:49:18,196 EPOCH 1 done: loss 0.6731 - lr: 0.020000
#> 2024-09-23 11:49:19,127 DEV : loss 0.6989173889160156 - f1-score (micro avg) 0.75
#> 2024-09-23 11:49:19,128 - 0 epochs without improvement
#> 2024-09-23 11:49:19,128 saving best model
#> 2024-09-23 11:49:19,747 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:49:19,748 Loading model from best epoch ...
#> 2024-09-23 11:49:21,806
#> Results:
#> - F-score (micro) 0.7294
#> - F-score (macro) 0.7689
#> - Accuracy 0.7294
#>
#> By class:
#> precision recall f1-score support
#>
#> Present 0.5400 1.0000 0.7013 27
#> Future 1.0000 0.5116 0.6769 43
#> Past 1.0000 0.8667 0.9286 15
#>
#> accuracy 0.7294 85
#> macro avg 0.8467 0.7928 0.7689 85
#> weighted avg 0.8539 0.7294 0.7291 85
#>
#> 2024-09-23 11:49:21,806 ----------------------------------------------------------------------------------------------------
#> $test_score
#> [1] 0.7294118
Continue Fine-tuning with New Dataset
Now, we can continue to fine tune the already fine tuned model with
an additional 2000 pieces of data. First, let’s say we have another 2000
entries called cc_muller_new
. We can fine-tune the previous
model with these 2000 entries. The steps are the same as before. For
this case, we don’t need to split the dataset again. We can use the
entire 2000 entries as the training set and use the
old_test
set to evaluate how well our refined model
performs.
Step 1 Load the
muller-campaign-communication
Model
Load the model (old_model
) you have already fine tuned
from previous stage and let’s fine tune it with the new data,
new_corpus
.
old_model <- TextClassifier$load("vignettes/inst/muller-campaign-communication/best-model.pt")
Step 2 Convert the New Data to Sentence and Corpus
library(purrr)
cc_muller_new <- cc_muller[1001:3000,]
new_text <- map(cc_muller_new$text, Sentence)
new_labels <- as.character(cc_muller_new$class)
new_text <- map2(new_text, new_labels, ~ {
.x$add_label("classification", .y)
.x
})
new_corpus <- Corpus(train=new_text, test=old_test)
#> 2024-09-23 11:49:23,287 No dev split found. Using 10% (i.e. 200 samples) of the train split as dev data
Step 3 Create a New Model Trainer with the Old Model and New Corpus
new_trainer <- ModelTrainer(old_model, new_corpus)
Step 4 Train the New Model
new_trainer$train("vignettes/inst/new-muller-campaign-communication",
learning_rate=0.002,
mini_batch_size=8L,
max_epochs=1L)
#> 2024-09-23 11:49:23,367 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:49:23,367 Model: "TextClassifier(
#> (embeddings): TransformerDocumentEmbeddings(
#> (model): DistilBertModel(
#> (embeddings): Embeddings(
#> (word_embeddings): Embedding(30523, 768, padding_idx=0)
#> (position_embeddings): Embedding(512, 768)
#> (LayerNorm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)
#> (dropout): Dropout(p=0.1, inplace=False)
#> )
#> (transformer): Transformer(
#> (layer): ModuleList(
#> (0-5): 6 x TransformerBlock(
#> (attention): MultiHeadSelfAttention(
#> (dropout): Dropout(p=0.1, inplace=False)
#> (q_lin): Linear(in_features=768, out_features=768, bias=True)
#> (k_lin): Linear(in_features=768, out_features=768, bias=True)
#> (v_lin): Linear(in_features=768, out_features=768, bias=True)
#> (out_lin): Linear(in_features=768, out_features=768, bias=True)
#> )
#> (sa_layer_norm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)
#> (ffn): FFN(
#> (dropout): Dropout(p=0.1, inplace=False)
#> (lin1): Linear(in_features=768, out_features=3072, bias=True)
#> (lin2): Linear(in_features=3072, out_features=768, bias=True)
#> (activation): GELUActivation()
#> )
#> (output_layer_norm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)
#> )
#> )
#> )
#> )
#> )
#> (decoder): Linear(in_features=768, out_features=3, bias=True)
#> (dropout): Dropout(p=0.0, inplace=False)
#> (locked_dropout): LockedDropout(p=0.0)
#> (word_dropout): WordDropout(p=0.0)
#> (loss_function): CrossEntropyLoss()
#> (weights): None
#> (weight_tensor) None
#> )"
#> 2024-09-23 11:49:23,367 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:49:23,367 Corpus: 1800 train + 200 dev + 85 test sentences
#> 2024-09-23 11:49:23,367 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:49:23,367 Train: 1800 sentences
#> 2024-09-23 11:49:23,367 (train_with_dev=False, train_with_test=False)
#> 2024-09-23 11:49:23,367 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:49:23,367 Training Params:
#> 2024-09-23 11:49:23,367 - learning_rate: "0.002"
#> 2024-09-23 11:49:23,367 - mini_batch_size: "8"
#> 2024-09-23 11:49:23,367 - max_epochs: "1"
#> 2024-09-23 11:49:23,367 - shuffle: "True"
#> 2024-09-23 11:49:23,367 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:49:23,367 Plugins:
#> 2024-09-23 11:49:23,367 - AnnealOnPlateau | patience: '3', anneal_factor: '0.5', min_learning_rate: '0.0001'
#> 2024-09-23 11:49:23,367 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:49:23,368 Final evaluation on model from best epoch (best-model.pt)
#> 2024-09-23 11:49:23,368 - metric: "('micro avg', 'f1-score')"
#> 2024-09-23 11:49:23,368 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:49:23,368 Computation:
#> 2024-09-23 11:49:23,368 - compute on device: cpu
#> 2024-09-23 11:49:23,368 - embedding storage: cpu
#> 2024-09-23 11:49:23,368 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:49:23,368 Model training base path: "vignettes/inst/new-muller-campaign-communication"
#> 2024-09-23 11:49:23,368 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:49:23,368 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:49:29,710 epoch 1 - iter 22/225 - loss 0.51795530 - time (sec): 6.34 - samples/sec: 27.75 - lr: 0.002000 - momentum: 0.000000
#> 2024-09-23 11:49:35,515 epoch 1 - iter 44/225 - loss 0.42166825 - time (sec): 12.15 - samples/sec: 28.98 - lr: 0.002000 - momentum: 0.000000
#> 2024-09-23 11:49:42,018 epoch 1 - iter 66/225 - loss 0.41576015 - time (sec): 18.65 - samples/sec: 28.31 - lr: 0.002000 - momentum: 0.000000
#> 2024-09-23 11:49:48,875 epoch 1 - iter 88/225 - loss 0.38241305 - time (sec): 25.51 - samples/sec: 27.60 - lr: 0.002000 - momentum: 0.000000
#> 2024-09-23 11:49:55,636 epoch 1 - iter 110/225 - loss 0.37842205 - time (sec): 32.27 - samples/sec: 27.27 - lr: 0.002000 - momentum: 0.000000
#> 2024-09-23 11:50:01,632 epoch 1 - iter 132/225 - loss 0.38865886 - time (sec): 38.26 - samples/sec: 27.60 - lr: 0.002000 - momentum: 0.000000
#> 2024-09-23 11:50:07,730 epoch 1 - iter 154/225 - loss 0.39412351 - time (sec): 44.36 - samples/sec: 27.77 - lr: 0.002000 - momentum: 0.000000
#> 2024-09-23 11:50:14,073 epoch 1 - iter 176/225 - loss 0.39203615 - time (sec): 50.70 - samples/sec: 27.77 - lr: 0.002000 - momentum: 0.000000
#> 2024-09-23 11:50:19,765 epoch 1 - iter 198/225 - loss 0.37966077 - time (sec): 56.40 - samples/sec: 28.09 - lr: 0.002000 - momentum: 0.000000
#> 2024-09-23 11:50:26,061 epoch 1 - iter 220/225 - loss 0.38061988 - time (sec): 62.69 - samples/sec: 28.07 - lr: 0.002000 - momentum: 0.000000
#> 2024-09-23 11:50:27,556 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:50:27,557 EPOCH 1 done: loss 0.3877 - lr: 0.002000
#> 2024-09-23 11:50:30,116 DEV : loss 0.3974565863609314 - f1-score (micro avg) 0.855
#> 2024-09-23 11:50:30,118 - 0 epochs without improvement
#> 2024-09-23 11:50:30,119 saving best model
#> 2024-09-23 11:50:30,728 ----------------------------------------------------------------------------------------------------
#> 2024-09-23 11:50:30,728 Loading model from best epoch ...
#> 2024-09-23 11:50:32,840
#> Results:
#> - F-score (micro) 0.8471
#> - F-score (macro) 0.8583
#> - Accuracy 0.8471
#>
#> By class:
#> precision recall f1-score support
#>
#> Future 0.8605 0.8605 0.8605 43
#> Present 0.7586 0.8148 0.7857 27
#> Past 1.0000 0.8667 0.9286 15
#>
#> accuracy 0.8471 85
#> macro avg 0.8730 0.8473 0.8583 85
#> weighted avg 0.8527 0.8471 0.8487 85
#>
#> 2024-09-23 11:50:32,840 ----------------------------------------------------------------------------------------------------
#> $test_score
#> [1] 0.8470588
More R tutorial and documentation see here.