I am building an app around GPT-3, and I would like to know how much tokens every request I make uses. Is this possible and how ?
Counting Tokens with Actual Tokenizer
To do this in python, first install the transformers package to enable the GPT-2 Tokenizer, which is the same tokenizer used for [GPT-3]:
pip install transformers
Then, to tokenize the string "Hello world", you have a choice of using GPT2TokenizerFast or GPT2Tokenizer.
from transformers import GPT2TokenizerFast\
tokenizer = GPT2TokenizerFast.from_pretrained("gpt2")\
number_of_tokens = len(tokenizer("Hello world")['input_ids'])
or
from transformers import GPT2Tokenizer\
tokenizer = GPT2Tokenizer.from_pretrained("gpt2")\
number_of_tokens = len(tokenizer("Hello world")['input_ids'])
In either case, tokenizer() produces a python list of token representing the string, which can the be counted with len(). The documentation doesn't mention any differences in behavior between the two methods. I tested both methods on both text and code and they gave the same numbers. The from_pretrained methods are unpleasantly slow: 28s for GPT2Tokenizer, and 56s for GPT2TokenizerFast. The load time dominates the experience, so I suggest NOT the "fast" method. (Note: the first time you run either of the from_pretrained methods, a 3MB model will be downloaded and installed, which takes a couple minutes.)
Approximating Token Counts
The tokenizers are slow and heavy, but approximations can be to go back and forth between them, using nothing but the number of characters or tokens. I developed the following approximations by observing the behavior of the GPT-2 tokenizer. They hold well for English text and python code. The 3rd and 4th functions are perhaps the most useful since they let us quickly fit a text in the GPT-3's token limit.
import math
def nchars_to_ntokens_approx(nchars):
#returns an estimate of #tokens corresponding to #characters nchars
return max(0,int((nchars - init_offset)*math.exp(-1)))
def ntokens_to_nchars_approx(ntokens):
#returns an estimate of #characters corresponding to #tokens ntokens
return max(0,int(ntokens*math.exp(1) ) + 2 )
def nchars_leq_ntokens_approx(maxTokens):
#returns a number of characters very likely to correspond <= maxTokens
sqrt_margin = 0.5
lin_margin = 1.010175047 #= e - 1.001 - sqrt(1 - sqrt_margin) #ensures return 1 when maxTokens=1
return max( 0, int(maxTokens*math.exp(1) - lin_margin - math.sqrt(max(0,maxTokens - sqrt_margin) ) ))
def truncate_text_to_maxTokens_approx(text, maxTokens):
#returns a truncation of text to make it (likely) fit within a token limit
#So the output string is very likely to have <= maxTokens, no guarantees though.
char_index = min( len(text), nchars_leq_ntokens_approx(maxTokens) )
return text[:char_index]
OPEN-AI charges GPT-3 usage through tokens, this counts both the prompt and the answer. For OPEN-AI 750 words would have an equivalent of around 1000 tokens or a token to word ratio of 1.4 . Pricing of the token depends of the plan you are on.
I do not know of more accurate ways of estimating cost. Perhaps using GPT-2 tokenizer from Hugging face can help. I know the tokens from the GPT-2 tokenizer are accepted when passed to GPT-3 in the logit bias array, so there is a degree of equivalence between GPT-2 tokens and GPT-3 tokens.
However GPT-2 and GPT-3 models are different and GPT-3 famously has more parameters than GPT-3 so GPT-2 estimations are probably lower token wise. I am sure you can write a simple program that estimates the price by comparing prompts and token usage, but that might take some time.
Here is an example from openai-cookbook that worked perfectly for me:
import tiktoken
def num_tokens_from_string(string: str, encoding_name: str) -> int:
"""Returns the number of tokens in a text string."""
encoding = tiktoken.get_encoding(encoding_name)
num_tokens = len(encoding.encode(string))
return num_tokens
num_tokens_from_string("tiktoken is great!", "gpt2")
>6
Code to count how much tokens a GPT-3 request used:
def count_tokens(input: str):
tokenizer = GPT2TokenizerFast.from_pretrained("gpt2")
res = tokenizer(input)['input_ids']
return len(res)
print(count_tokens("Hello world"))
Related
I'm using customized text with 'Prompt' and 'Completion' to train new model.
Here's the tutorial I used to create customized model from my data:
beta.openai.com/docs/guides/fine-tuning/advanced-usage
However even after training the model and sending prompt text to the model, I'm still getting generic results which are not always suitable for me.
How I can make sure completion results for my prompts will be only from the text I used for the model and not from the generic OpenAI models?
Can I use some flags to eliminate results from generic models?
Wrong goal: OpenAI API should answer from the fine-tuning dataset if the prompt is similar to the one from the fine-tuning dataset
It's the completely wrong logic. Forget about fine-tuning. As stated on the official OpenAI website:
Fine-tuning lets you get more out of the models available through the
API by providing:
Higher quality results than prompt design
Ability to train on more examples than can fit in a prompt
Token savings due to shorter prompts
Lower latency requests
Fine-tuning is not about answering with a specific answer from the fine-tuning dataset.
Fine-tuning helps the model gain more knowledge, but it has nothing to do with how the model answers. Why? The answer we get from the fine-tuned model is based on all knowledge (i.e., fine-tuned model knowledge = default knowledge + fine-tuning knowledge).
Although GPT-3 models have a lot of general knowledge, sometimes we want the model to answer with a specific answer (i.e., "fact").
Correct goal: Answer with a "fact" when asked about a "fact", otherwise answer with the OpenAI API
Note: For better (visual) understanding, the following code was ran and tested in Jupyter.
STEP 1: Create a .csv file with "facts"
To keep things simple, let's add two companies (i.e., ABC and XYZ) with a content. The content in our case will be a 1-sentence description of the company.
companies.csv
Run print_dataframe.ipynb to print the dataframe.
print_dataframe.ipynb
import pandas as pd
df = pd.read_csv('companies.csv')
df
We should get the following output:
STEP 2: Calculate an embedding vector for every "fact"
An embedding is a vector of numbers that helps us understand how semantically similar or different the texts are. The closer two embeddings are to each other, the more similar are their contents (source).
Let's test the Embeddings endpoint first. Run get_embedding.ipynb with an input This is a test.
Note: In the case of Embeddings endpoint, the parameter prompt is called input.
get_embedding.ipynb
import openai
openai.api_key = '<OPENAI_API_KEY>'
def get_embedding(model: str, text: str) -> list[float]:
result = openai.Embedding.create(
model = model,
input = text
)
return result['data'][0]['embedding']
print(get_embedding('text-embedding-ada-002', 'This is a test'))
We should get the following output:
What we see in the screenshot above is This is a test as an embedding vector. More precisely, we get a 1536-dimensional embedding vector (i.e., there are 1536 numbers inside). You are probably familiar with a 3-dimensional space (i.e., X, Y, Z). Well, this is a 1536-dimensional space which is very hard to imagine.
There are two things we need to understand at this point:
Why do we need to transform text into an embedding vector (i.e., numbers)? Because later on, we can compare embedding vectors and figure out how similar the two texts are. We can't compare texts as such.
Why are there exactly 1536 numbers inside the embedding vector? Because the text-embedding-ada-002 model has an output dimension of 1536. It's pre-defined.
Now we can create an embedding vector for each "fact". Run get_all_embeddings.ipynb.
get_all_embeddings.ipynb
import openai
from openai.embeddings_utils import get_embedding
import pandas as pd
openai.api_key = '<OPENAI_API_KEY>'
df = pd.read_csv('companies.csv')
df['embedding'] = df['content'].apply(lambda x: get_embedding(x, engine = 'text-embedding-ada-002'))
df.to_csv('companies_embeddings.csv')
The code above will take the first company (i.e., x), get its 'content' (i.e., "fact") and apply the function get_embedding using the text-embedding-ada-002 model. It will save the embedding vector of the first company in a new column named 'embedding'. Then it will take the second company, the third company, the fourth company, etc. At the end, the code will automatically generate a new .csv file named companies_embeddings.csv.
Saving embedding vectors locally (i.e., in a .csv file) means we don't have to call the OpenAI API every time we need them. We calculate an embedding vector for a given "fact" once and that's it.
Run print_dataframe_embeddings.ipynb to print the dataframe with the new column named 'embedding'.
print_dataframe_embeddings.ipynb
import pandas as pd
import numpy as np
df = pd.read_csv('companies_embeddings.csv')
df['embedding'] = df['embedding'].apply(eval).apply(np.array)
df
We should get the following output:
STEP 3: Calculate an embedding vector for the input and compare it with embedding vectors from the companies_embeddings.csv using cosine similarity
We need to calculate an embedding vector for the input so that we can compare the input with a given "fact" and see how similar these two texts are. Actually, we compare the embedding vector of the input with the embedding vector of the "fact". Then we compare the input with the second "fact", the third "fact", the fourth "fact", etc. Run get_cosine_similarity.ipynb.
get_cosine_similarity.ipynb
import openai
from openai.embeddings_utils import cosine_similarity
import pandas as pd
openai.api_key = '<OPENAI_API_KEY>'
my_model = 'text-embedding-ada-002'
my_input = '<INSERT_INPUT>'
def get_embedding(model: str, text: str) -> list[float]:
result = openai.Embedding.create(
model = my_model,
input = my_input
)
return result['data'][0]['embedding']
input_embedding_vector = get_embedding(my_model, my_input)
df = pd.read_csv('companies_embeddings.csv')
df['embedding'] = df['embedding'].apply(eval).apply(np.array)
df['similarity'] = df['embedding'].apply(lambda x: cosine_similarity(x, input_embedding_vector))
df
The code above will take the input and compare it with the first fact. It will save the calculated similarity of the two in a new column named 'similarity'. Then it will take the second fact, the third fact, the fourth fact, etc.
If my_input = 'Tell me something about company ABC':
If my_input = 'Tell me something about company XYZ':
If my_input = 'Tell me something about company Apple':
We can see that when we give Tell me something about company ABC as an input, it's the most similar to the first "fact". When we give Tell me something about company XYZ as an input, it's the most similar to the second "fact". Whereas, if we give Tell me something about company Apple as an input, it's the least similar to any of these two "facts".
STEP 4: Answer with the most similar "fact" if similarity is above our threshold, otherwise answer with the OpenAI API
Let's set our similarity threshold to >= 0.9. The code below should answer with the most similar "fact" if similarity is >= 0.9, otherwise answer with the OpenAI API. Run get_answer.ipynb.
get_answer.ipynb
# Imports
import openai
from openai.embeddings_utils import cosine_similarity
import pandas as pd
import numpy as np
# Insert your API key
openai.api_key = '<OPENAI_API_KEY>'
# Insert OpenAI text embedding model and input
my_model = 'text-embedding-ada-002'
my_input = '<INSERT_INPUT>'
# Calculate embedding vector for the input using OpenAI Embeddings endpoint
def get_embedding(model: str, text: str) -> list[float]:
result = openai.Embedding.create(
model = my_model,
input = my_input
)
return result['data'][0]['embedding']
# Save embedding vector of the input
input_embedding_vector = get_embedding(my_model, my_input)
# Calculate similarity between the input and "facts" from companies_embeddings.csv file which we created before
df = pd.read_csv('companies_embeddings.csv')
df['embedding'] = df['embedding'].apply(eval).apply(np.array)
df['similarity'] = df['embedding'].apply(lambda x: cosine_similarity(x, input_embedding_vector))
# Find the highest similarity value in the dataframe column 'similarity'
highest_similarity = df['similarity'].max()
# If the highest similarity value is equal or higher than 0.9 then print the 'content' with the highest similarity
if highest_similarity >= 0.9:
fact_with_highest_similarity = df.loc[df['similarity'] == highest_similarity, 'content']
print(fact_with_highest_similarity)
# Else pass input to the OpenAI Completions endpoint
else:
response = openai.Completion.create(
model = 'text-davinci-003',
prompt = my_input,
max_tokens = 30,
temperature = 0
)
content = response['choices'][0]['text'].replace('\n', '')
print(content)
If my_input = 'Tell me something about company ABC' and the threshold is >= 0.9 we should get the following answer from the companies_embeddings.csv:
If my_input = 'Tell me something about company XYZ' and the threshold is >= 0.9 we should get the following answer from the companies_embeddings.csv:
If my_input = 'Tell me something about company Apple' and the threshold is >= 0.9 we should get the following answer from the OpenAI API:
Goal: set min_length and max_length in Hugging Face Transformers generator query.
I've passed 50, 200 as these parameters. Yet, the length of my outputs are much higher...
There's no runtime failure.
from transformers import pipeline, set_seed
generator = pipeline('text-generation', model='gpt2')
set_seed(42)
def query(payload, multiple, min_char_len, max_char_len):
print(min_char_len, max_char_len)
list_dict = generator(payload, min_length=min_char_len, max_length=max_char_len, num_return_sequences=multiple)
test = [d['generated_text'].split(payload)[1].strip() for d in list_dict]
for t in test: print(len(t))
return test
query('example', 1, 50, 200)
Output:
50 200
Setting `pad_token_id` to `eos_token_id`:50256 for open-end generation.
1015
Explanation:
As explained by Narsil on Hugging Face 🤗 Transformers Git Issue response
Models, don't ingest the text one character at a time, but one token
at a time. There are different algorithms to achieve this but
basically "My name is Nicolas" gets transformers into ["my", " name",
" is", " nic", "olas"] for instance, and each of those tokens have a
number.
So when you are generating tokens, they can contain themselves 1 or
more characters (usually several and almost any common word for
instance). That's why you are seeing 1015 instead of your expected 200
(the tokens here have an average of 5 chars)
Solution:
As I resolved...
Rename min_char_len, max_char_len to min_tokens, max_tokens and
simply reduce their values by a ~1/4 or 1/5.
I am still a beginner with neural networks and NLP.
In this code I'm training cleaned text (some tweets) with skip-gram.
But I do not know if I do it correctly.
Can anyone inform me about the correctness of this skip-gram text training?
Any help is appreciated.
This my code :
from nltk import word_tokenize
from gensim.models.phrases import Phrases, Phraser
sent = [row.split() for row in X['clean_text']]
phrases = Phrases(sent, max_vocab_size = 50, progress_per=10000)
bigram = Phraser(phrases)
sentences = bigram[sent]
from gensim.models import Word2Vec
w2v_model = Word2Vec(window=5,
size = 300,
sg=1)
w2v_model.build_vocab(sentences)
w2v_model.train(sentences, total_examples=w2v_model.corpus_count, epochs=25)
del sentences #to reduce memory usage
def get_mat(model, corpus, size):
vecs = np.zeros((len(corpus), size))
n = 0
for i in corpus.index:
vecs[i] = np.zeros(size).reshape((1, size))
for word in str(corpus.iloc[i,0]).split():
try:
vecs[i] += model[word]
#n += 1
except KeyError:
continue
return vecs
X_sg = get_vectors(w2v_model, X, 300)
del X
X_sg=pd.DataFrame(X_sg)
X_sg.head()
from sklearn import preprocessing
scale = preprocessing.normalize
X_sg=scale(X_sg)
for i in range(len(X_sg)):
X_sg[i]+=1 #I did this because some weights where negative! So could not
#apply LSTM on them later
You haven't mentioned if you've received any errors, or unsatisfactory results, so it's hard to know what kind of help you might need.
Your specific lines of code involving the Word2Vec model are roughly correct: plausibly-useful parameters (if you have a dataset large enough to train 300-dimensional vectors), and the proper steps. So the real proof would be whether your results are acceptable.
Regarding your attempted use of Phrases bigram-creation beforehand:
You should get things generally working and with promising results before adding this extra pre-processing complexity.
The parameter max_vocab_size=50 is seriously misguided and may make the phrases-step pointless. The max_vocab_size is a hard cap on how many words/bigrams are tallied by the class, as a way to cap its memory-usage. (Whenever the number of known words/bigrams hits this cap, many lower-frequency words/bigrams are pruned – in practice, a majority of all words/bigrams each pruning, giving up a lot of accuracy in return for capped memory usage.) The max_vocab_size default in gensim is 40,000,000 – but the default in the Google word2phrase.c source on which gensim's method is based was 500,000,000. By using just 50, it's not really going to learn anything useful about just whatever 50 words/bigrams survive the many prunings.
Regarding your get_mat() function & later DataFrame code, i have no idea what you're trying to do with it, so can't offer any opinion on it.
I'm trying to do data enhancement with a FAQ dataset. I change words, specifically nouns, by most similar words with Wordnet checking the similarity with Spacy. I use multiple for loop to go through my dataset.
import spacy
import nltk
from nltk.corpus import wordnet as wn
import pandas as pd
nlp = spacy.load('en_core_web_md')
nltk.download('wordnet')
questions = pd.read_csv("FAQ.csv")
list_questions = []
for question in questions.values:
list_questions.append(nlp(question[0]))
for question in list_questions:
for token in question:
treshold = 0.5
if token.pos_ == 'NOUN':
wordnet_syn = wn.synsets(str(token), pos=wn.NOUN)
for syn in wordnet_syn:
for lemma in syn.lemmas():
similar_word = nlp(lemma.name())
if similar_word.similarity(token) != 1. and similar_word.similarity(token) > treshold:
good_word = similar_word
treshold = token.similarity(similar_word)
However, the following warning is printed several times and I don't understand why :
UserWarning: [W008] Evaluating Doc.similarity based on empty vectors.
It is my similar_word.similarity(token) which creates the problem but I don't understand why.
The form of my list_questions is :
list_questions = [Do you have a paper or other written explanation to introduce your model's details?, Where is the BERT code come from?, How large is a sentence vector?]
I need to check token but also the similar_word in the loop, for example, I still get the error here :
tokens = nlp(u'dog cat unknownword')
similar_word = nlp(u'rabbit')
if(similar_word):
for token in tokens:
if (token):
print(token.text, similar_word.similarity(token))
You get that error message when similar_word is not a valid spacy document. E.g. this is a minimal reproducible example:
import spacy
nlp = spacy.load('en_core_web_md') # make sure to use larger model!
tokens = nlp(u'dog cat')
#similar_word = nlp(u'rabbit')
similar_word = nlp(u'')
for token in tokens:
print(token.text, similar_word.similarity(token))
If you change the '' to be 'rabbit' it works fine. (Cats are apparently just a fraction more similar to rabbits than dogs are!)
(UPDATE: As you point out, unknown words also trigger the warning; they will be valid spacy objects, but not have any word vector.)
So, one fix would be to check similar_word is valid, including having a valid word vector, before calling similarity():
import spacy
nlp = spacy.load('en_core_web_md') # make sure to use larger model!
tokens = nlp(u'dog cat')
similar_word = nlp(u'')
if(similar_word and similar_word.vector_norm):
for token in tokens:
if(token and token.vector_norm):
print(token.text, similar_word.similarity(token))
Alternative Approach:
You could suppress the particular warning. It is W008. I believe setting an environmental variable SPACY_WARNING_IGNORE=W008 before running your script would do it. (Not tested.)
(See source code)
By the way, similarity() might cause some CPU load, so is worth storing in a variable, instead of calculating it three times as you currently do. (Some people might argue that is premature optimization, but I think it might also make the code more readable.)
I have suppress the W008 warning by setting environmental variable by using this code in run file.
import os
app = Flask(__name__)
app.config['SPACY_WARNING_IGNORE'] = "W008"
os.environ["SPACY_WARNING_IGNORE"] = "W008"
if __name__ == "__main__":
app.run(host='0.0.0.0', port=5000)
I am trying to follow the official Doc2Vec Gensim tutorial mentioned here - https://github.com/RaRe-Technologies/gensim/blob/develop/docs/notebooks/doc2vec-lee.ipynb
I modified the code in line 10 to determine best matching document for the given query and everytime I run, I get a completely different resultset. My new code iin line 10 of the notebook is:
inferred_vector = model.infer_vector(['only', 'you', 'can', 'prevent', 'forest', 'fires'])
sims = model.docvecs.most_similar([inferred_vector], topn=len(model.docvecs))
rank = [docid for docid, sim in sims]
print(rank)
Everytime I run the piece of code, I get different set of documents that are matching with this query: "only you can prevent forest fires". The difference is stark and just does not seem to match.
Is Doc2Vec not a suitable match for querying and information extraction? Or are there bugs?
Look into the code, in infer_vector you are using parts of the algorithm that is non-deterministic. Initialization of word vector is deterministic - see the code of seeded_vector, but when we look further, i.e., random sampling of words, negative sampling (updating only sample of word vector per iteration) could cause non-deterministic output (thanks #gojomo).
def seeded_vector(self, seed_string):
"""Create one 'random' vector (but deterministic by seed_string)"""
# Note: built-in hash() may vary by Python version or even (in Py3.x) per launch
once = random.RandomState(self.hashfxn(seed_string) & 0xffffffff)
return (once.rand(self.vector_size) - 0.5) / self.vector_size
Set negative=0 to avoid randomization:
import numpy as np
from gensim.models.doc2vec import Doc2Vec, TaggedDocument
documents = [list('asdf'), list('asfasf')]
documents = [TaggedDocument(doc, [i]) for i, doc in enumerate(documents)]
model = Doc2Vec(documents, vector_size=20, window=5, min_count=1, negative=0, workers=6, epochs=10)
a = list('test sample')
b = list('testtesttest')
for s in (a, b):
v1 = model.infer_vector(s)
for i in range(100):
v2 = model.infer_vector(s)
assert np.all(v1 == v2), "Failed on %s" % (''.join(s))