Spaces:
Runtime error
Runtime error
| import streamlit as st | |
| import numpy as np | |
| # TODO: move to 'utils' | |
| mystyle = ''' | |
| <style> | |
| p { | |
| text-align: justify; | |
| } | |
| </style> | |
| ''' | |
| st.markdown(mystyle, unsafe_allow_html=True) | |
| def divider(): | |
| _, c, _ = st.columns(3) | |
| c.divider() | |
| def get_embeddings(text): | |
| return np.array(openai.Embedding.create(input=text, model=EMBEDDING_MODEL)["data"][0]["embedding"]) | |
| st.title("Transformers: Tokenisers and Embeddings") | |
| preface_image, preface_text, = st.columns(2) | |
| # preface_image.image("https://static.streamlit.io/examples/dice.jpg") | |
| # preface_image.image("""https://assets.digitalocean.com/articles/alligator/boo.svg""") | |
| preface_text.write("""\ | |
| *Transformers represent a revolutionary class of machine learning architectures that have sparked | |
| immense interest. While numerous insightful tutorials are available, the evolution of transformer architectures over | |
| the last few years has led to significant simplifications. These advancements have made it increasingly | |
| straightforward to understand their inner workings. In this series of articles, I aim to provide a direct, clear explanation of | |
| how and why modern transformers function, unburdened by the historical complexities associated with their inception.* | |
| """) | |
| divider() | |
| st.write("""\ | |
| In order to understand the recent success in AI we need to understand the Transformer architecture. Its | |
| rise in the field of Natural Language Processing (NLP) is largely attributed to a combination of several key | |
| advancements: | |
| - Tokenisers and Embeddings | |
| - Attention and Self-Attention | |
| - Encoder-Decoder architecture | |
| Understanding these foundational concepts is crucial to comprehending the overall structure and function of the | |
| Transformer model. They are the building blocks from which the rest of the model is constructed, and their roles | |
| within the architecture are essential to the model's ability to process and generate language. In my view, | |
| a comprehensive and simple explanation may give a reader a significant advantage in using LLMs. Feynman once said: | |
| "*I think I can safely say that nobody understands quantum mechanics.*". Because he couldn't explain it to a freshman. | |
| Given the importance and complexity of these concepts, I have chosen to dedicate the first article in this series | |
| solely to Tokenisation and embeddings. The decision to separate the topics into individual articles is driven by a | |
| desire to provide a thorough and in-depth understanding of each component of the Transformer model. | |
| Note: *HuggingFace provides an exceptional [tutorial on Transformer models](https://huggingface.co/docs/transformers/index). | |
| That tutorial is particularly beneficial for readers willing to dive into advanced topics.* | |
| """) | |
| with st.expander("Copernicus Museum in Warsaw"): | |
| st.write("""\ | |
| Have you ever visited the Copernicus Museum in Warsaw? It's an engaging interactive hub that allows | |
| you to familiarise yourself with various scientific topics. The experience is both entertaining and educational, | |
| providing the opportunity to explore different concepts firsthand. **They even feature a small neural network that | |
| illustrates the neuron activation process during the recognition of handwritten digits!** | |
| I encourage you not to hesitate in modifying parameters or experimenting with different models in the provided | |
| examples. This hands-on exploration can significantly enhance your learning experience. So, let's begin our journey | |
| through this virtual, interactive museum of AI. Enjoy the exploration! | |
| """) | |
| st.image("https://i.pinimg.com/originals/04/11/2c/04112c791a859d07a01001ac4f436e59.jpg") | |
| divider() | |
| st.header("Tokenisers and Tokenisation") | |
| st.write("""\ | |
| Tokenisation is the initial step in the data preprocessing pipeline for natural language processing (NLP) | |
| models. It involves breaking down a piece of text—whether a sentence, paragraph, or document—into smaller units, | |
| known as "tokens". In English and many other languages, a token often corresponds to a word, but it can also be a | |
| subword, character, or n-gram. The choice of token size depends on various factors, including the task at hand and | |
| the language of the text. | |
| """) | |
| from transformers import AutoTokenizer | |
| sentence = st.text_input("Consider the sentence: (you can change it):", value="Tokenising text is a fundamental step for NLP models.") | |
| sentence_split = sentence.split() | |
| tokenizer = AutoTokenizer.from_pretrained('bert-base-uncased') | |
| sentence_tokenise_bert = tokenizer.tokenize(sentence) | |
| sentence_encode_bert = tokenizer.encode(sentence) | |
| sentence_encode_bert = list(zip(sentence_tokenise_bert, sentence_encode_bert)) | |
| st.write(f"""\ | |
| A basic word-level tokenisation, which splits a text by spaces, would produce next tokens: | |
| """) | |
| st.code(f""" | |
| {sentence_split} | |
| """) | |
| st.write(f"""\ | |
| However, we notice that the punctuation may attached to the words. It is disadvantageous, how the tokenization dealt with the word "Don't". | |
| "Don't" stands for "do not", so it would be better tokenized as ["Do", "n't"]. (Hint: try another sentence: "I musn't tell lies. Don't do this.") This is where things start getting complicated, | |
| and part of the reason each model has its own tokenizer type. Depending on the rules we apply for tokenizing a text, | |
| a different tokenized output is generated for the same text. | |
| A more sophisticated algorithm, with several optimizations, might generate a different set of tokens: | |
| """) | |
| st.code(f""" | |
| {sentence_tokenise_bert} | |
| """) | |
| with st.expander("click here to look at the Python code:"): | |
| st.code(f"""\ | |
| from transformers import AutoTokenizer | |
| sentence = "{sentence}" | |
| sentence_split = sentence.split() | |
| tokenizer = AutoTokenizer.from_pretrained('bert-base-uncased') | |
| sentence_tokenise_bert = tokenizer.tokenize(sentence) | |
| sentence_encode_bert = tokenizer.encode(sentence) | |
| sentence_encode_bert = list(zip(sentence_tokenise_bert, sentence_encode_bert)) | |
| """, language='python') | |
| st.write(""" | |
| As machine learning models, including Transformers, work with numbers rather than words, each vocabulary | |
| entry is assigned a corresponding numerical value. Here is a potential key-value, vocabulary-based representation of | |
| the input (so called 'token ids'): | |
| """ | |
| ) | |
| st.code(f""" | |
| {sentence_encode_bert} | |
| """) | |
| st.write(""" | |
| What distinguishes subword Tokenisation is its reliance on statistical rules and algorithms, learned from | |
| the pretraining corpus. The resulting Tokeniser creates a vocabulary, which usually represents the most frequently | |
| used words and subwords. For example, Byte Pair Encoding (BPE) first encodes the most frequent words as single | |
| tokens, while less frequent words are represented by multiple tokens, each representing a word part. | |
| There are numerous different Tokenisers available, including spaCy, Moses, Byte-Pair Encoding (BPE), | |
| Byte-level BPE, WordPiece, Unigram, and SentencePiece. It's crucial to choose a specific Tokeniser and stick with it. | |
| Changing the Tokeniser is akin to altering the model's language on the fly—imagine studying physics in English and | |
| then taking the exam in French or Spanish. You might get lucky, but it's a considerable risk. | |
| """) | |
| training_dataset = """\ | |
| Beautiful is better than ugly. | |
| Explicit is better than implicit. | |
| Simple is better than complex. | |
| Complex is better than complicated. | |
| Flat is better than nested. | |
| Sparse is better than dense. | |
| Readability counts. | |
| """ | |
| tokeniser_name = st.selectbox(label="Choose your tokeniser", options=["BPE", 'Unigram', 'WordPiece']) | |
| if tokeniser_name == 'BPE': | |
| st.subheader("Byte-Pair Encoding (BPE)") | |
| st.write("""\ | |
| Byte-Pair Encoding (BPE) was introduced in [Neural Machine Translation of Rare Words with Subword | |
| Units (Sennrich et al., 2015)](https://arxiv.org/abs/1508.07909). BPE relies on a pre-tokenizer that splits the | |
| training data into words. Pre-tokenization can be as simple as space tokenization, e.g. GPT-2, Roberta. More | |
| advanced pre-tokenization include rule-based tokenization, e.g. XLM, FlauBERT which uses Moses for most | |
| languages, or GPT which uses Spacy and ftfy, to count the frequency of each word in the training corpus. | |
| After pre-tokenization, a set of unique words has been created and the frequency with which each word occurred in the | |
| training data has been determined. Next, BPE creates a base vocabulary consisting of all symbols that occur in the | |
| set of unique words and learns merge rules to form a new symbol from two symbols of the base vocabulary. It does so | |
| until the vocabulary has attained the desired vocabulary size. Note that the desired vocabulary size is a | |
| hyperparameter to define before training the tokenizer. | |
| As an example, let’s assume that after pre-tokenization, the following set of words including their frequency has | |
| been determined: | |
| """) | |
| st.code(""" ("hug", 10), ("pug", 5), ("pun", 12), ("bun", 4), ("hugs", 5) """) | |
| st.write("""\ | |
| Consequently, the base vocabulary is ["b", "g", "h", "n", "p", "s", "u"]. Splitting all words into symbols of the base vocabulary, we obtain: | |
| """) | |
| st.code(""" ("h" "u" "g", 10), ("p" "u" "g", 5), ("p" "u" "n", 12), ("b" "u" "n", 4), ("h" "u" "g" "s", 5) """) | |
| st.write("""\ | |
| BPE then counts the frequency of each possible symbol pair and picks the symbol pair that occurs | |
| most frequently. In the example above "h" followed by "u" is present 10 + 5 = 15 times (10 times in the 10 | |
| occurrences of "hug", 5 times in the 5 occurrences of "hugs"). However, the most frequent symbol pair is "u" | |
| followed by "g", occurring 10 + 5 + 5 = 20 times in total. Thus, the first merge rule the tokenizer learns is to | |
| group all "u" symbols followed by a "g" symbol together. Next, "ug" is added to the vocabulary. The set of words | |
| then becomes | |
| """) | |
| st.code(""" ("h" "ug", 10), ("p" "ug", 5), ("p" "u" "n", 12), ("b" "u" "n", 4), ("h" "ug" "s", 5) """) | |
| st.write("""\ | |
| BPE then identifies the next most common symbol pair. It’s "u" followed by "n", which occurs 16 | |
| times. "u", "n" is merged to "un" and added to the vocabulary. The next most frequent symbol pair is "h" followed | |
| by "ug", occurring 15 times. Again the pair is merged and "hug" can be added to the vocabulary. | |
| At this stage, the vocabulary is ["b", "g", "h", "n", "p", "s", "u", "ug", "un", "hug"] and our set of unique words is represented as | |
| """) | |
| st.code(""" ("hug", 10), ("p" "ug", 5), ("p" "un", 12), ("b" "un", 4), ("hug" "s", 5) """) | |
| st.write("""\ | |
| Assuming, that the Byte-Pair Encoding training would stop at this point, the learned merge rules | |
| would then be applied to new words (as long as those new words do not include symbols that were not in the base | |
| vocabulary). For instance, the word "bug" would be tokenized to ["b", "ug"] but "mug" would be tokenized as [ | |
| "[unk]", "ug"] since the symbol "m" is not in the base vocabulary. In general, single letters such as "m" are not | |
| replaced by the "[unk]" symbol because the training data usually includes at least one occurrence of each letter, | |
| but it is likely to happen for very special characters like emojis. | |
| As mentioned earlier, the vocabulary size, i.e. the base vocabulary size + the number of merges, is a hyperparameter | |
| to choose. For instance GPT has a vocabulary size of 40,478 since they have 478 base characters and chose to stop | |
| training after 40,000 merges. | |
| """) | |
| st.subheader(":green[Try Yourself:]") | |
| st.write(f"""\ | |
| *Using text area field below try to find or create a comprehensive vocabulary (training dataset) for Tokenisation, which can enhance the | |
| efficiency of the process. This approach helps to eliminate unknown tokens, thereby making the token sequence | |
| more understandable and containing less tokens (ids)* | |
| """) | |
| training_dataset = st.text_area("*Training Dataset - Vocabulary:*", value=training_dataset, height=200) | |
| training_dataset = training_dataset.split('\n') | |
| vocabulary_size = st.number_input("Vocabulary Size:", value=100000) | |
| sentence = st.text_input(label="*Text to tokenise:*", | |
| value="[CLS] Tokenising text is a fundamental step for NLP models. [SEP] [PAD] [PAD] [PAD]") | |
| from tokenizers import Tokenizer, decoders, models, normalizers, pre_tokenizers, trainers | |
| tokenizer = Tokenizer(models.BPE(unk_token="[UNK]")) | |
| tokenizer.pre_tokenizer = pre_tokenizers.Whitespace() | |
| tokenizer.decoder = decoders.ByteLevel() | |
| trainer = trainers.BpeTrainer(special_tokens=["[UNK]", "[CLS]", "[SEP]", "[PAD]", "[MASK]"], vocab_size=vocabulary_size) | |
| tokenizer.train_from_iterator(training_dataset, trainer=trainer) | |
| output = tokenizer.encode(sentence) | |
| st.write("*Tokens:*") | |
| st.code(f"""{output.tokens}""") | |
| st.code(f"""\ | |
| ids: {output.ids} | |
| attention_mast: {output.attention_mask} | |
| """) | |
| st.write(""" *well done if you get ids like these: [1, 57, 49, 28, 10, 58, 55, 52, 31, 54, 5, 2, 3, 3, 3]!*""") | |
| with st.expander("Python code:"): | |
| st.code(f""" | |
| from tokenizers import Tokenizer, decoders, models, normalizers, pre_tokenizers, trainers | |
| tokenizer = Tokenizer(models.BPE(unk_token="[UNK]")) | |
| tokenizer.pre_tokenizer = pre_tokenizers.Whitespace() | |
| tokenizer.decoder = decoders.ByteLevel() | |
| trainer = trainers.BpeTrainer( | |
| special_tokens=["[UNK]", "[CLS]", "[SEP]", "[PAD]", "[MASK]"], | |
| vocab_size={vocabulary_size}) | |
| training_dataset = {training_dataset} | |
| tokenizer.train_from_iterator(training_dataset, trainer=trainer) | |
| output = tokenizer.encode("{sentence}") | |
| """, language='python') | |
| elif tokeniser_name == 'Unigram': | |
| st.subheader("""Unigram""") | |
| st.write("""\ | |
| Unigram is a subword tokenization algorithm introduced in [Subword Regularization: Improving Neural | |
| Network Translation Models with Multiple Subword Candidates (Kudo, 2018)](https://arxiv.org/pdf/1804.10959.pdf). | |
| In contrast to BPE or WordPiece, Unigram initializes its base vocabulary to a large number of symbols and | |
| progressively trims down each symbol to obtain a smaller vocabulary. The base vocabulary could for instance | |
| correspond to all pre-tokenized words and the most common substrings. Unigram is not used directly for any of the | |
| models in the transformers, but it’s used in conjunction with SentencePiece. | |
| At each training step, the Unigram algorithm defines a loss (often defined as the log-likelihood) over the training | |
| data given the current vocabulary and a unigram language model. Then, for each symbol in the vocabulary, | |
| the algorithm computes how much the overall loss would increase if the symbol was to be removed from the vocabulary. | |
| Unigram then removes p (with p usually being 10% or 20%) percent of the symbols whose loss increase is the lowest, | |
| i.e. those symbols that least affect the overall loss over the training data. This process is repeated until the | |
| vocabulary has reached the desired size. The Unigram algorithm always keeps the base characters so that any word can | |
| be tokenized. | |
| Because Unigram is not based on merge rules (in contrast to BPE and WordPiece), the algorithm has several ways of | |
| tokenizing new text after training. As an example, if a trained Unigram tokenizer exhibits the vocabulary: | |
| """) | |
| st.code(""" ["b", "g", "h", "n", "p", "s", "u", "ug", "un", "hug"] """) | |
| st.write("""\ | |
| "hugs" could be tokenized both as ["hug", "s"], ["h", "ug", "s"] or ["h", "u", "g", "s"]. So which | |
| one to choose? Unigram saves the probability of each token in the training corpus on top of saving the vocabulary | |
| so that the probability of each possible tokenization can be computed after training. The algorithm simply picks | |
| the most likely tokenization in practice, but also offers the possibility to sample a possible tokenization | |
| according to their probabilities. | |
| """) | |
| st.subheader(":green[Try Yourself:]") | |
| st.write(f"""\ | |
| *Using text area field below try to find or create a comprehensive vocabulary (training dataset) for Tokenisation, which can enhance the | |
| efficiency of the process. This approach helps to eliminate unknown tokens, thereby making the token sequence | |
| more understandable and containing less tokens (ids)* | |
| """) | |
| training_dataset = st.text_area("*Training Dataset - Vocabulary(change it and looks at resulted tokens):*", value=training_dataset, height=200) | |
| training_dataset = training_dataset.split('\n') | |
| vocabulary_size = st.number_input("Vocabulary Size:", value=100000) | |
| sentence = st.text_input(label="*Text to tokenise:*", | |
| value="[CLS] Tokenising text is a fundamental step for NLP models. [SEP] [PAD] [PAD] [PAD]") | |
| from tokenizers import Tokenizer, decoders, models, normalizers, pre_tokenizers, trainers | |
| tokenizer = Tokenizer(models.Unigram()) | |
| tokenizer.pre_tokenizer = pre_tokenizers.Whitespace() | |
| tokenizer.decoder = decoders.ByteLevel() | |
| trainer = trainers.UnigramTrainer( | |
| vocab_size=vocabulary_size, | |
| unk_token="[UNK]", | |
| # initial_alphabet=pre_tokenizers.ByteLevel.alphabet(), | |
| special_tokens=["[UNK]", "[CLS]", "[SEP]", "[PAD]", "[MASK]"], | |
| ) | |
| tokenizer.train_from_iterator(training_dataset, trainer=trainer) | |
| output = tokenizer.encode(sentence) | |
| # TODO: make it more visible, container with a differect color or something | |
| st.write("*Tokens:*") | |
| st.code(f"""{output.tokens}""") | |
| st.code(f"""\ | |
| ids: {output.ids} | |
| attention_mast: {output.attention_mask} | |
| """) | |
| st.write(""" *well done if you get ids like these: [1, 57, 49, 28, 10, 58, 55, 52, 31, 54, 5, 2, 3, 3, 3]!*""") | |
| with st.expander("Python code:"): | |
| st.code(f"""\ | |
| from tokenizers import Tokenizer, decoders, models, normalizers, pre_tokenizers, trainers | |
| tokenizer = Tokenizer(models.Unigram()) | |
| tokenizer.pre_tokenizer = pre_tokenizers.Whitespace() | |
| tokenizer.decoder = decoders.ByteLevel() | |
| trainer = trainers.UnigramTrainer( | |
| vocab_size={vocabulary_size}, | |
| special_tokens=["[UNK]", "[CLS]", "[SEP]", "[PAD]", "[MASK]"], | |
| ) | |
| training_dataset = {training_dataset} | |
| tokenizer.train_from_iterator(training_dataset, trainer=trainer) | |
| output = tokenizer.encode("{sentence}") | |
| """, language='python') | |
| elif tokeniser_name == 'WordPiece': | |
| st.subheader("""WordPiece""") | |
| st.write("""\ | |
| WordPiece is the subword tokenization algorithm used for BERT, DistilBERT, and Electra. The | |
| algorithm was outlined in [Japanese and Korean Voice Search (Schuster et al., | |
| 2012)](https://static.googleusercontent.com/media/research.google.com/ja//pubs/archive/37842.pdf) and is very | |
| similar to BPE. WordPiece first initializes the vocabulary to include every character present in the training | |
| data and progressively learns a given number of merge rules. In contrast to BPE, WordPiece does not choose the | |
| most frequent symbol pair, but the one that maximizes the likelihood of the training data once added to the | |
| vocabulary. | |
| So what does this mean exactly? Referring to the example from BPE tokeniser, maximizing the likelihood of the training data is | |
| equivalent to finding the symbol pair, whose probability divided by the probabilities of its first symbol followed by | |
| its second symbol is the greatest among all symbol pairs. E.g. "u", followed by "g" would have only been merged if | |
| the probability of "ug" divided by "u", "g" would have been greater than for any other symbol pair. Intuitively, | |
| WordPiece is slightly different to BPE in that it evaluates what it loses by merging two symbols to ensure it’s worth | |
| it. | |
| """) | |
| st.subheader(":green[Try Yourself:]") | |
| st.write(f"""\ | |
| *Using text area field below try to find or create a comprehensive vocabulary (training dataset) for Tokenisation, which can enhance the | |
| efficiency of the process. This approach helps to eliminate unknown tokens, thereby making the token sequence | |
| more understandable and containing less tokens (ids)* | |
| """) | |
| training_dataset = st.text_area("*Training Dataset - Vocabulary(change it and looks at resulted tokens):*", | |
| value=training_dataset, height=200) | |
| training_dataset = training_dataset.split('\n') | |
| vocabulary_size = st.number_input("Vocabulary Size:", value=100000) | |
| sentence = st.text_input(label="*Text to tokenise:*", | |
| value="[CLS] Tokenising text is a fundamental step for NLP models. [SEP] [PAD] [PAD] [PAD]") | |
| from tokenizers import Tokenizer, decoders, models, pre_tokenizers, trainers | |
| tokenizer = Tokenizer(models.WordPiece(unk_token="[UNK]")) | |
| tokenizer.pre_tokenizer = pre_tokenizers.Whitespace() | |
| tokenizer.decoder = decoders.ByteLevel() | |
| trainer = trainers.WordPieceTrainer( | |
| vocab_size=vocabulary_size, | |
| special_tokens=["[UNK]", "[CLS]", "[SEP]", "[PAD]", "[MASK]"], | |
| ) | |
| tokenizer.train_from_iterator(training_dataset, trainer=trainer) | |
| output = tokenizer.encode(sentence) | |
| # TODO: make it more visible, container with a differect color or something | |
| st.write("*Tokens:*") | |
| st.code(f"""{output.tokens}""") | |
| st.code(f"""\ | |
| ids: {output.ids} | |
| attention_mast: {output.attention_mask} | |
| """) | |
| st.write(""" *well done if you get ids like these: [1, 76, 72, 50, 10, 77, 71, 68, 66, 78, 5, 2, 3, 3, 3]!*""") | |
| with st.expander("Python code:"): | |
| st.code(f"""\ | |
| from tokenizers import Tokenizer, decoders, models, pre_tokenizers, trainers | |
| tokenizer = Tokenizer(models.WordPiece(unk_token="[UNK]")) | |
| trainer = trainers.WordPieceTrainer( | |
| vocab_size={vocabulary_size}, | |
| special_tokens=["[UNK]", "[CLS]", "[SEP]", "[PAD]", "[MASK]"], | |
| ) | |
| training_dataset = {training_dataset} | |
| tokenizer.train_from_iterator(training_dataset, trainer=trainer) | |
| output = tokenizer.encode("{sentence}") | |
| """, language='python') | |
| with st.expander("Special tokens meaning:"): | |
| st.write("""\ | |
| \\#\\# prefix: It means that the preceding string is not whitespace, any token with this prefix should be | |
| merged with the previous token when you convert the tokens back to a string. | |
| [UNK]: Stands for "unknown". This token is used to represent any word that is not in the model's vocabulary. Since | |
| most models have a fixed-size vocabulary, it's not possible to have a unique token for every possible word. The [UNK] | |
| token is used as a catch-all for any words the model hasn't seen before. E.g. in our example we 'decided' that Large | |
| Language (LL) abbreviation is not part of the model's vocabulary. | |
| [CLS]: Stands for "classification". In models like BERT, this token is added at the beginning of every input | |
| sequence. The representation (embedding) of this token is used as the aggregate sequence representation for | |
| classification tasks. In other words, the model is trained to encode the meaning of the entire sequence into this token. | |
| [SEP]: Stands for "separator". This token is used to separate different sequences when the model needs to take more | |
| than one input sequence. For example, in question-answering tasks, the model takes two inputs: a question and a | |
| passage that contains the answer. The two inputs are separated by a [SEP] token. | |
| [MASK]: This token is specific to models like BERT, which are trained with a masked language modelling objective. | |
| During training, some percentage of the input tokens are replaced with the [MASK] token, and the model's goal is to | |
| predict the original value of the masked tokens. | |
| [PAD]: Stands for "padding". This token is used to fill in the extra spaces when batching sequences of different | |
| lengths together. Since models require input sequences to be the same length, shorter sequences are extended with [ | |
| PAD] tokens. In our example, we extended the length of the input sequence to 16 tokens. | |
| """) | |
| with st.expander("References:"): | |
| st.write("""\ | |
| - https://huggingface.co/docs/transformers/tokenizer_summary | |
| - https://huggingface.co/docs/tokenizers/training_from_memory | |
| """) | |
| divider() | |
| st.header("Embeddings") | |
| st.write("""\ | |
| Following tokenization, each token is transformed into a vector of numeric characteristics, a process | |
| known as 'embedding.' In this context, 'embedding' refers to the mapping of the discrete, categorical space of words | |
| or tokens into a continuous, numeric space, which the model can manipulate more effectively. | |
| Each dimension in this high-dimensional space can encapsulate a different facet of the token's meaning. For instance, | |
| one dimension might capture the tense of a token if it's a verb, while another dimension might capture the degree of | |
| positivity or negativity if the token is an adjective expressing sentiment. For instance: | |
| """) | |
| st.code("""\ | |
| "I" -> [noun, person] | |
| "love" -> [verb, feeling] | |
| "machine" -> [noun, automation] | |
| "learn" -> [verb, knowledge] | |
| "##ing" -> [gerund, continues] | |
| """) | |
| st.write("""\ | |
| The actual embeddings in a typical NLP model would be in a much higher-dimensional space (often several hundred dimensions), but the idea is the same. | |
| Embeddings are dynamically learned from the data, with the model adjusting these embeddings during | |
| training to minimize the discrepancy between the predicted and actual outputs for a set of training examples. | |
| Consequently, tokens with similar meanings often end up with similar embeddings. | |
| In the context of Transformers, these embeddings are the inputs that the model uses. Once again, we represent all the | |
| characteristics using numbers, not words. | |
| """) | |
| st.write("""\ | |
| Let's explore embeddings in more detail. We can take an experimental approach by encoding two specific | |
| words and examining the corresponding embedding vectors they generate. To make our exploration more accessible, | |
| we'll visualise a portion of these vectors, thereby unveiling the underlying structure of embeddings. Pay attention | |
| to common patterns and peaks, try to find two words that yield differing embeddings. | |
| """) | |
| col1, col2, col3 = st.columns(3) | |
| token_king = col1.text_input("Choose a word:", value="king") | |
| token_queen = col2.text_input("Choose a word:", value="queen") | |
| token_dots = col3.number_input("Number of dots:", value=50, min_value=0, max_value=1536) | |
| from torch import nn | |
| from transformers import AutoConfig | |
| from transformers import AutoTokenizer | |
| import pandas as pd | |
| import openai | |
| import plotly.express as px | |
| model_ckpt = 'bert-base-uncased' | |
| tokenizer = AutoTokenizer.from_pretrained(model_ckpt) | |
| king_id = tokenizer(token_king, return_tensors="pt", add_special_tokens=False) | |
| queen_id = tokenizer(token_queen, return_tensors="pt", add_special_tokens=False) | |
| config = AutoConfig.from_pretrained(model_ckpt) | |
| token_emb = nn.Embedding(config.vocab_size, config.hidden_size) | |
| king_embeddings = token_emb(king_id.input_ids) | |
| queen_embeddings = token_emb(queen_id.input_ids) | |
| king_emb_np = king_embeddings.reshape(-1).detach().numpy() | |
| queen_emb_np = queen_embeddings.reshape(-1).detach().numpy() | |
| openai.api_key = st.secrets["OPENAI_API_KEY"] | |
| EMBEDDING_MODEL = 'text-embedding-ada-002' | |
| EMBEDDING_CTX_LENGTH = 8191 | |
| EMBEDDING_ENCODING = 'cl100k_base' | |
| king = get_embeddings(token_king) | |
| queen = get_embeddings(token_queen) | |
| df = pd.DataFrame({f'"{token_king}" embeddings': king_emb_np, f'"{token_queen}" embeddings': queen_emb_np}) | |
| fig = px.line(df[:token_dots], title=f"Google's 'bert-base-uncased' model embeddings, embedding vector size: {len(queen_emb_np)}") | |
| fig.update_layout(legend=dict(orientation="h")) | |
| st.plotly_chart(fig, use_container_width=True) | |
| with st.expander("Python Code:"): | |
| st.code(f"""\ | |
| from torch import nn | |
| from transformers import AutoConfig | |
| model_ckpt = 'bert-base-uncased' | |
| tokenizer = AutoTokenizer.from_pretrained(model_ckpt) | |
| king_id = tokenizer("{token_king}", return_tensors="pt", add_special_tokens=False) | |
| queen_id = tokenizer("{token_queen}", return_tensors="pt", add_special_tokens=False) | |
| config = AutoConfig.from_pretrained(model_ckpt) | |
| token_emb = nn.Embedding(config.vocab_size, config.hidden_size) | |
| king_embeddings = token_emb(king_id.input_ids) | |
| queen_embeddings = token_emb(queen_id.input_ids) | |
| """) | |
| df = pd.DataFrame({f'"{token_king}" embeddings': king, f'"{token_queen}" embeddings': queen}) | |
| fig = px.line(df[:token_dots], title=f"OpenAI's 'text-embedding-ada-002' model embeddings, embedding vector size: {len(queen)}") | |
| fig.update_layout(legend=dict(orientation="h")) | |
| st.plotly_chart(fig, use_container_width=True) | |
| with st.expander("Python Code:"): | |
| st.code(f"""\ | |
| import openai | |
| import numpy as np | |
| EMBEDDING_MODEL = 'text-embedding-ada-002' | |
| king_embeddings = np.array(openai.Embedding.create(input="{token_king}", model=EMBEDDING_MODEL)["data"][0]["embedding"]) | |
| queen_embeddings = np.array(openai.Embedding.create(input="{token_queen}", model=EMBEDDING_MODEL)["data"][0]["embedding"]) | |
| """) | |
| st.write("""\ | |
| The similarity can be represented as a similarity score. Identical words naturally have the highest | |
| score (black colours), while unrelated terms have lower scores (white colours). To compute this score, | |
| we construct a matrix infused with our embedding vectors. Each row in this matrix corresponds to a unique word in the | |
| sentence, while each column aligns with another word. The value at the intersection of row i and column j represents | |
| the score between word i and word j. For a clearer understanding, let's visualise this matrix using a heatmap. Each | |
| cell in the grid corresponds to a pair of words, and the colour of the cell indicates the similarity (correlation) | |
| score between those two words. The intensity of the colour directly corresponds to the magnitude of the score - the | |
| darker the hue, the higher the score. | |
| """) | |
| st.write("""Here is a heatmap of the score matrix for the sentence:""") | |
| sentence = st.text_input(label="*words to explore embeddings*", value="a the king queen space sit eat from on") | |
| sentence = sentence.split() | |
| input = {word: get_embeddings(word) for word in sentence} | |
| scores_matrix = np.zeros((len(sentence), len(sentence))) | |
| for i, word_i in enumerate(sentence): | |
| for j, word_j in enumerate(sentence): | |
| scores_matrix[i, j] = np.dot(input[word_i], input[word_j]) | |
| fig = px.imshow(scores_matrix, x=sentence, y=sentence, color_continuous_scale="hot_r") | |
| fig.update_layout(coloraxis_showscale=False) | |
| fig.update_layout(width=6000) | |
| st.plotly_chart(fig, use_container_width=True) | |
| st.subheader("Vector Databases") | |
| st.write("""\ | |
| In a vector database, each item (e.g., a document) is represented as a point in a multidimensional | |
| space. Each point is a vector that represents the features of the item. The goal is to place similar items close to | |
| each other and dissimilar items farther apart. In the case of documents, the features could be derived from the words | |
| in the document, and the similarity might be based on the overlapping words or concepts between the documents. | |
| The retrieval of documents based on search terms involves two main steps: | |
| - **Vectorization of the search query**: The search query is converted into a vector using the same process used to vectorize the documents in the database. | |
| - **Vector similarity search**: The vector database then identifies the vectors that are closest to the query vector. | |
| This is typically done using a distance metric like Euclidean distance or cosine similarity. The documents | |
| corresponding to these vectors are returned as the search results. | |
| As you correctly assumed, we leverage embedding algorithms to vectorise documents. Let's generate a 3D | |
| visualization of the document vectors and a search query. For simplicity, let's assume we have a vector database | |
| of documents that has been reduced to 3 dimensions, and we'll also have a 3D vector for a search query. | |
| """) | |
| with st.expander("The Euclidean distance between two points in 3D space is calculated as:"): | |
| st.latex("""\\text{Distance}(A(x_1, y_1, z_1), B(x_2, y_2, z_2)) = \sqrt{(x_2 - x_1)^2 + (y_2 - y_1)^2 + (z_2 - z_1)^2}""") | |
| st.write("""\ | |
| The document that corresponds to the vector with the smallest distance to the query vector is | |
| considered the most relevant document. The 3D plot above now shows lines from the query vector (in red) to each | |
| document vector (in blue). Each line represents the Euclidean distance from the query vector to a document vector. | |
| """) | |
| embeddings = st.text_input("vector space:", value="king queen prince princess counselor minister teacher") | |
| embeddings = embeddings.split() | |
| embeddings_query = st.text_input(label="search term", value='woman') | |
| import numpy as np | |
| import plotly.express as px | |
| import plotly.graph_objects as go | |
| from sklearn.manifold import TSNE | |
| embeddings = {word: get_embeddings(word) for word in embeddings} | |
| embeddings[embeddings_query] = get_embeddings(embeddings_query) | |
| tsne = TSNE(n_components=3, perplexity=3, random_state=0) | |
| embedding_matrix = np.array(list(embeddings.values())) | |
| reduced_embeddings = tsne.fit_transform(embedding_matrix) | |
| df = pd.DataFrame(reduced_embeddings, columns=["X", "Y", "Z"]) | |
| df["Word"] = list(embeddings.keys()) | |
| fig = px.scatter_3d(df, x="X", y="Y", z="Z", text="Word", title="Vector Space", width=800, height=800) | |
| docs = reduced_embeddings[:-1] | |
| query = reduced_embeddings[-1] | |
| distances = np.linalg.norm(docs - query, axis=1) | |
| closest_doc_index = np.argmin(distances) | |
| closest_doc = docs[closest_doc_index] | |
| for doc in docs: | |
| fig.add_trace(go.Scatter3d(x=[query[0], doc[0]], y=[query[1], doc[1]], z=[query[2], doc[2]], mode='lines', line=dict(color='purple', width=2, dash='dash'))) | |
| fig.add_trace(go.Scatter3d(x=[query[0], closest_doc[0]], y=[query[1], closest_doc[1]], z=[query[2], closest_doc[2]], name='closest', mode='lines', line=dict(color='purple', width=2))) | |
| fig.update_layout(legend=dict(orientation="h")) | |
| st.plotly_chart(fig, use_container_width=True) | |
| st.write("""\ | |
| This visualization represents the core concept of a vector database search. The database converts the | |
| search query into a vector, then finds the document vectors that are closest to the query vector. Those documents are | |
| considered the most relevant to the search query. | |
| It's important to note that in a real-world application, the vectors would likely exist in much higher dimensional | |
| space. However, the same principles apply: the search algorithm finds the document vectors that are nearest to the | |
| query vector based on some distance metric. | |
| """) | |
| st.subheader(":green[Try Yourself]") | |
| st.write("""\ | |
| *There is a vector database containing two words (documents): 'king' and 'queen'. Your task is to pinpoint search | |
| terms that would yield either of these words. To facilitate this, use the previously presented similarity matrix to | |
| seek out words that give a higher correlation with the word in question. For instance, you might want to explore | |
| terms such as 'king', 'queen', 'dog', 'prince', 'man', 'minister', 'boy'.* | |
| """) | |
| embeddings_query = st.text_input(label="search term") | |
| from langchain.embeddings.openai import OpenAIEmbeddings | |
| from langchain.vectorstores import FAISS | |
| from langchain.schema.document import Document | |
| def create_vector_database(): | |
| return FAISS.from_documents([Document(page_content="king"), Document(page_content="queen")], OpenAIEmbeddings(model=EMBEDDING_MODEL)) | |
| db = create_vector_database() | |
| def search_vector_database(term): | |
| embedding_vector = OpenAIEmbeddings(model=EMBEDDING_MODEL).embed_query(term) | |
| docs = db.similarity_search_by_vector(embedding_vector) | |
| return docs | |
| if embeddings_query is not None and embeddings_query != '': | |
| docs = search_vector_database(embeddings_query) | |
| st.warning(docs[0].page_content) | |
| with st.expander("Python Code:"): | |
| st.code(f"""\ | |
| from langchain.embeddings.openai import OpenAIEmbeddings | |
| from langchain.vectorstores import FAISS | |
| from langchain.schema.document import Document | |
| db = FAISS.from_documents([Document(page_content="king"), Document(page_content="queen")], OpenAIEmbeddings(model=EMBEDDING_MODEL)) | |
| embedding_vector = OpenAIEmbeddings(model=EMBEDDING_MODEL).embed_query("{embeddings_query}") | |
| docs = db.similarity_search_by_vector(embedding_vector) | |
| """) | |
| divider() | |
| st.subheader("Conclusion") | |
| st.write("""\ | |
| As embedding algorithms are trained on a vast corpus of data, they inherently encapsulate a rich | |
| tapestry of information about our language and even the world at large. Therefore, they can be used for: | |
| - Search (where results are ranked by relevance to a query string) | |
| - Clustering (where text strings are grouped by similarity) | |
| - Recommendations (where items with related text strings are recommended) | |
| - Anomaly detection (where outliers with little relatedness are identified) | |
| - Diversity measurement (where similarity distributions are analyzed) | |
| - Classification (where text strings are classified by their most similar label) | |
| """) | |
| with st.expander("References:"): | |
| st.write("""\ | |
| - https://huggingface.co/blog/getting-started-with-embeddings | |
| - https://huggingface.co/blog/1b-sentence-embeddings | |
| - https://platform.openai.com/docs/guides/embeddings/use-cases | |
| """) | |
| # ********************************************* | |
| divider() | |
| st.header("Dimensionality Reduction (optional)") | |
| st.write("""\ | |
| As was mentioned above, embedding vectors are learned in such a way that words with similar meanings | |
| are located close to each other in the space. However, this is an abstract concept that might be difficult to | |
| explore, understand and visualise in a 2D space because word embeddings typically have hundreds of dimensions. To | |
| solve this, we can use techniques like Principal Component Analysis (PCA) or t-SNE to reduce the dimensionality of | |
| the vectors and plot them. | |
| """) | |
| st.write("""But first, let's talk about the meaning of dimensionality reduction using simplified use-case:""") | |
| dimensionality_name = st.selectbox(label="Choose your example", options=["Simplified", "PCA", 't-SNE']) | |
| if dimensionality_name == 'Simplified': | |
| _, col2, _ = st.columns(3) | |
| col2.image("assets/img.png") | |
| st.write("""\ | |
| **Step 1: The context**\n | |
| We have a 3D object (your hand) and a light source that's casting a 2D shadow of your hand onto a | |
| wall. The shadow is a simpler, lower-dimensional representation of your hand. | |
| **Step 2: Identifying the dimensions**\n | |
| In this case, the dimensions are the different aspects of your hand that can be | |
| observed: the length of your fingers, the width of your palm, the height (or depth) of your hand, the scars, | |
| the colour of the skin, etc. However, we have a problem: we can't easily visualise or understand all these dimensions | |
| at once. Just as it's hard to imagine a 6-dimensional space. | |
| **Step 3: Deciding on important dimensions**\n | |
| Let's say you want to compare the number of fingers of different hands. In | |
| this case, you don't need to know about the depth of the hand, the width of the palm, or other details like freckles, | |
| scars, or skin colour. You just need a shadow that clearly shows the fingers. So, you decide to focus on the length | |
| of the fingers, which can be easily shown in the shadow. | |
| **Step 4: Reducing dimensions**\n | |
| This is where you actually perform dimensionality reduction. You orient your hand in such | |
| a way (giving the wall a high-five) that the shadow clearly shows the fingers. You've effectively reduced the | |
| dimensions from 3D to 2D. Your hand is still a 3D object, but its shadow — the simplified representation you're using | |
| for your comparison — is 2D. | |
| **Step 5: Interpretation**\n | |
| This hand and shadow example shows how dimensionality reduction simplifies a complex object ( | |
| the 3D hand) into a lower-dimensional representation (the 2D shadow) that retains the most important information (the | |
| number of fingers) while discarding the less important details (like the depth of the hand, skin colour, etc.). It's | |
| a process of prioritisation and simplification that makes it easier for us to understand and analyse the data (or the | |
| hands, in this case). | |
| """) | |
| elif dimensionality_name == 'PCA': | |
| st.write("""\ | |
| **Step 1: Understanding PCA**\n | |
| PCA is a popular method for dimensionality reduction. It identifies the | |
| axes in the feature space along which the original data varies the most. These axes are known as the principal | |
| components, and they are orthogonal (perpendicular) to each other. | |
| **Step 2: Projecting the Data**\n | |
| Imagine that instead of just casting a shadow on the wall, you can cast your hand's | |
| shadow onto a number of walls arranged at different angles around your hand. Each shadow is a different projection of | |
| your hand. In PCA, these different walls represent different principal components, and the shadow on each wall is a | |
| projection of your hand onto that principal component. | |
| **Step 3: Choosing the Best Projection**\n | |
| Now, consider the shadow that most accurately portrays the number of fingers on | |
| your hand. This shadow corresponds to the principal component that captures the most variance in the data. In PCA, | |
| this would be the first principal component. | |
| **Step 4: Secondary Features**\n | |
| Next, consider the shadow that, while not as accurate as the first, still gives a | |
| reasonable representation of your hand, such as showing the width of your palm. This shadow represents the second | |
| principal component, which captures the second highest amount of variance in the data. | |
| **Step 5: Reduction of Dimensions**\n | |
| In the process of reducing dimensions, we select the top few principal components ( | |
| shadows) that capture the most variance. The other dimensions (shadows) are discarded. So, instead of having to | |
| consider the complex 3D structure of your hand, you can simply look at one or two shadows that give you the most | |
| information about the hand. | |
| **Step 6: Transformation**\n | |
| Finally, we transform the original data into the reduced dimensional space defined by the | |
| selected principal components. This is analogous to replacing each hand with the selected shadows for further analysis. | |
| By using PCA, we can reduce the complexity of the data (from a 3D hand to a 2D or even 1D shadow), while still | |
| retaining the most important information (like the number of fingers or the width of the palm). This makes the data | |
| easier to visualize, understand, and work with. | |
| """) | |
| embedding_dim = 1536 | |
| embeddings = st.text_input("words to explore:", | |
| value="king queen man woman prince princess counselor minister teacher") | |
| embeddings = embeddings.split() | |
| embeddings = {word: get_embeddings(word) for word in embeddings} | |
| from sklearn.decomposition import PCA | |
| pca = PCA(n_components=2) | |
| embedding_matrix = np.array(list(embeddings.values())) | |
| reduced_embeddings = pca.fit_transform(embedding_matrix) | |
| df = pd.DataFrame(reduced_embeddings, columns=["X", "Y"]) | |
| df["Word"] = list(embeddings.keys()) | |
| fig = px.scatter(df, x="X", y="Y", text="Word", title="Word Embeddings", width=800, height=800) | |
| st.plotly_chart(fig, use_container_width=True) | |
| st.code(f"""\ | |
| from sklearn.decomposition import PCA | |
| import numpy as np | |
| pca = PCA(n_components=2) | |
| embedding_matrix = np.array(list(embeddings.values())) | |
| reduced_embeddings = pca.fit_transform(embedding_matrix) | |
| """, language='python') | |
| elif dimensionality_name == 't-SNE': | |
| st.write("""\ | |
| **Step 1: Understanding t-SNE**\n | |
| t-SNE is a technique for dimensionality reduction that is particularly | |
| well-suited for the visualization of high-dimensional datasets. Unlike PCA, which is a linear technique, | |
| t-SNE is a non-linear technique, making it better at capturing complex polynomial relationships between variables. | |
| **Step 2: Measuring Similarities**\n | |
| Imagine that instead of just one hand, you have many hands casting shadows. Each hand | |
| is different - some hands might have longer fingers, some might have a wider palm, and so on. Each hand has its own | |
| "neighborhood" of similar hands. In t-SNE, these neighborhoods are represented mathematically by a probability | |
| distribution. Hands that are very similar to each other have a high probability of being "neighbors", while hands | |
| that are very different have a low probability. | |
| **Step 3: Creating a Map**\n | |
| t-SNE creates a map (or a projection) where hands that were close in the high-dimensional | |
| space (similar hands) are still close in the low-dimensional space (in their shadows), and hands that were far apart | |
| in the high-dimensional space (different hands) are still far apart in the low-dimensional space. This map is created | |
| in such a way that it minimizes the difference between the distances in the high-dimensional space and the distances | |
| in the low-dimensional space. | |
| **Step 4: Reducing Dimensions**\n | |
| The process of reducing dimensions in t-SNE involves optimizing the locations of each | |
| hand's shadow in the low-dimensional space such that the overall configuration of shadows best represents the | |
| similarities between the hands in the high-dimensional space. | |
| **Step 5: Interpretation**\n | |
| The result of t-SNE is a map where similar hands are located close together and dissimilar | |
| hands are located far apart. This makes it easier to visualize clusters or groups of similar hands. | |
| t-SNE, therefore, helps us to project high-dimensional data into a lower-dimensional space in a way that preserves | |
| the structure of the data as much as possible, making it easier to visualize and understand the relationships in the | |
| data. | |
| """) | |
| embedding_dim = 1536 | |
| embeddings = st.text_input("words to explore:", | |
| value="king queen man woman prince princess counselor minister teacher") | |
| embeddings = embeddings.split() | |
| embeddings = {word: get_embeddings(word) for word in embeddings} | |
| from sklearn.manifold import TSNE | |
| tsne = TSNE(n_components=2, perplexity=2, random_state=0) | |
| embedding_matrix = np.array(list(embeddings.values())) | |
| reduced_embeddings = tsne.fit_transform(embedding_matrix) | |
| df = pd.DataFrame(reduced_embeddings, columns=["X", "Y"]) | |
| df["Word"] = list(embeddings.keys()) | |
| fig = px.scatter(df, x="X", y="Y", text="Word", title="Word Embeddings", width=800, height=800) | |
| st.plotly_chart(fig, use_container_width=True) | |
| st.code(f"""\ | |
| from sklearn.manifold import TSNE | |
| import numpy as np | |
| tsne = TSNE(n_components=2, perplexity=2, random_state=0) | |
| embedding_matrix = np.array(list(embeddings.values())) | |
| reduced_embeddings = tsne.fit_transform(embedding_matrix) | |
| """, language='python') | |
| with st.expander("References:"): | |
| st.write("""\ | |
| - https://hex.tech/blog/dimensionality-reduction/ | |
| - https://github.com/openai/openai-cookbook/blob/main/examples/Visualizing_embeddings_in_2D.ipynb | |
| """) |